khive-runtime 0.8.0

Composable Service API: entity/note CRUD, graph traversal, hybrid search, curation.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
//! khived daemon server — persistent warm runtime over a Unix socket.
//!
//! The daemon binds `~/.khive/khived.sock`, accepts length-prefixed request
//! frames, dispatches them through a [`DaemonDispatch`] implementor, and serves
//! results back. It is transport-agnostic: the MCP crate provides the dispatch
//! impl, but any future client (CLI, HTTP gateway) can reuse this server.
//!
//! The client side (forwarding, auto-spawn) lives in the transport crate
//! (e.g. `khive-mcp`), not here.

use std::sync::Arc;

#[cfg(unix)]
use std::io::Write as _;
#[cfg(unix)]
use std::os::unix::fs::PermissionsExt;
#[cfg(unix)]
use std::os::unix::io::AsRawFd;
use std::path::PathBuf;

#[cfg(unix)]
use async_trait::async_trait;
#[cfg(unix)]
use libc;
use serde::{Deserialize, Serialize};
#[cfg(unix)]
use tokio::io::{AsyncReadExt, AsyncWriteExt};
#[cfg(unix)]
use tokio::net::{UnixListener, UnixStream};

#[cfg(unix)]
use khive_db::{run_checkpoint_task, CheckpointConfig, CheckpointLifecycleOwner, ConnectionPool};

#[cfg(unix)]
use crate::pack::RequestIdentity;

/// Maximum frame size accepted in either direction.
pub const MAX_FRAME_BYTES: usize = 8 * 1024 * 1024;

/// Wire protocol version for the daemon IPC framing.
///
/// Increment this constant whenever the request or response frame shape
/// changes in a backward-incompatible way. The client sends its version
/// in every request; the daemon rejects mismatches with an explicit error
/// that names both sides so the operator knows exactly what to do
/// (`make local` rebuilds the client binary).
/// See `docs/api/daemon.md#protocol_version` for the version-by-version history.
pub const PROTOCOL_VERSION: u32 = 4;

const DEFAULT_DRAIN_TIMEOUT_SECS: u64 = 10;

// ── paths ─────────────────────────────────────────────────────────────────────

/// Base `.khive` directory used to anchor every advisory lock/socket/pid
/// path below. Pure path computation — portable on every target, even
/// though most of its callers (socket/pid paths) are unix-only.
///
/// Resolution order: `HOME`, then `USERPROFILE` (the conventional Windows
/// home variable), then a platform-specific last resort. On non-unix targets
/// the last resort is the OS temp directory (per-user on Windows), so the
/// lock path can never be working-directory-relative there: two processes
/// opening the same database from different working directories must resolve
/// the same lock file. On unix the last resort stays the historical `"."` —
/// a shared world-writable anchor such as `/tmp` would be worse, since a
/// local attacker could pre-claim the directory and the socket/lock files
/// under it before the daemon's first run.
fn khive_dir() -> PathBuf {
    khive_root_from(
        std::env::var("HOME").ok(),
        std::env::var("USERPROFILE").ok(),
    )
}

/// Env-free core of [`khive_dir`], split out so the fallback chain is
/// testable without mutating process-global environment variables.
fn khive_root_from(home: Option<String>, userprofile: Option<String>) -> PathBuf {
    home.filter(|v| !v.trim().is_empty())
        .or_else(|| userprofile.filter(|v| !v.trim().is_empty()))
        .map(PathBuf::from)
        .unwrap_or_else(last_resort_root)
        .join(".khive")
}

/// See [`khive_dir`] for why the two arms differ.
#[cfg(unix)]
fn last_resort_root() -> PathBuf {
    PathBuf::from(".")
}

#[cfg(not(unix))]
fn last_resort_root() -> PathBuf {
    std::env::temp_dir()
}

/// Unix socket path the daemon binds and clients connect to.
///
/// Overridable via the `KHIVE_SOCKET` env var (for tests and ops).
#[cfg(unix)]
pub fn socket_path() -> PathBuf {
    if let Ok(p) = std::env::var("KHIVE_SOCKET") {
        if !p.is_empty() {
            return PathBuf::from(p);
        }
    }
    khive_dir().join("khived.sock")
}

/// PID file path written by the daemon.
///
/// Overridable via the `KHIVE_PID` env var.
#[cfg(unix)]
pub fn pid_path() -> PathBuf {
    if let Ok(p) = std::env::var("KHIVE_PID") {
        if !p.is_empty() {
            return PathBuf::from(p);
        }
    }
    khive_dir().join("khived.pid")
}

/// Advisory lock file used to serialize stale-daemon recovery across concurrent
/// clients (flock/`File::lock` on the file; released when the lock file
/// handle is dropped). Path computation is portable; `kkernel exec`'s
/// non-unix local-construction guard (`kkernel::exec::acquire_local_construction_guard`)
/// shares this exact path with the unix daemon-boot guard so the two stay
/// mutually exclusive.
///
/// Overridable via the `KHIVE_LOCK` env var (for tests).
pub fn lock_path() -> PathBuf {
    if let Ok(p) = std::env::var("KHIVE_LOCK") {
        if !p.is_empty() {
            return PathBuf::from(p);
        }
    }
    khive_dir().join("khived.recovery.lock")
}

/// Advisory lock file used to serialize RECOVERY (kill+respawn) attempts
/// across concurrent clients only — the daemon's own boot sequence never
/// acquires this file ([`lock_path`] / [`acquire_daemon_boot_guard`] is the
/// boot-side lock). A recoverer holding this lock across dead-confirmation
/// → kill → spawn (khive-mcp's `kill_and_respawn`) therefore can never
/// deadlock against a peer daemon's boot, unlike holding the shared boot
/// lock for that whole span would.
///
/// Overridable via the `KHIVE_RECOVERER_LOCK` env var (for tests).
#[cfg(unix)]
pub fn recoverer_lock_path() -> PathBuf {
    if let Ok(p) = std::env::var("KHIVE_RECOVERER_LOCK") {
        if !p.is_empty() {
            return PathBuf::from(p);
        }
    }
    khive_dir().join("khived.recoverer.lock")
}

#[cfg(unix)]
fn open_lock_file(path: &std::path::Path) -> std::io::Result<std::fs::File> {
    if let Some(parent) = path.parent() {
        let _ = std::fs::create_dir_all(parent);
    }
    std::fs::OpenOptions::new()
        .create(true)
        .truncate(false)
        .write(true)
        .open(path)
}

#[cfg(unix)]
fn acquire_flock_blocking(path: &std::path::Path, label: &str) -> Option<std::fs::File> {
    let file = match open_lock_file(path) {
        Ok(f) => f,
        Err(e) => {
            tracing::warn!(error = %e, path = ?path, "cannot open {label} lock file");
            return None;
        }
    };
    // SAFETY: flock is a POSIX advisory lock with no memory side-effects.
    let rc = unsafe { libc::flock(file.as_raw_fd(), libc::LOCK_EX) };
    if rc != 0 {
        tracing::warn!("flock LOCK_EX failed on {label} lock");
        return None;
    }
    Some(file)
}

/// Acquire an exclusive advisory flock on the recovery/startup lock file.
///
/// The returned `File` holds the lock for its lifetime; dropping it releases
/// it.  Used by both the client (serializing kill+spawn) and the daemon server
/// (serializing cleanup+bind+pid-write) so the two critical sections are
/// mutually exclusive across processes.
#[cfg(unix)]
pub fn acquire_recovery_lock() -> Option<std::fs::File> {
    acquire_flock_blocking(&lock_path(), "recovery")
}

/// Attempt to acquire an exclusive advisory flock on `path`, retrying with a
/// non-blocking `flock(LOCK_NB)` until `deadline` elapses. Bounded alternative
/// to `acquire_recovery_lock`/`acquire_daemon_boot_guard`'s unbounded blocking
/// flock — see `docs/api/daemon.md#try_acquire_flock_until` for why a caller
/// merely detecting lock freedom needs a deadline instead.
///
/// - `Ok(Some(file))` — the lock was free within the deadline.
/// - `Ok(None)` — `deadline` elapsed while the lock stayed held; an explicit
///   "could not confirm" outcome, distinct from a hard I/O error.
/// - `Err(_)` — the lock file could not be opened, or `flock` failed for a
///   reason other than contention.
///
/// Blocking (paces retries with `std::thread::sleep`) — async callers must
/// run this via `spawn_blocking`.
#[cfg(unix)]
fn try_acquire_flock_until(
    path: &std::path::Path,
    deadline: std::time::Instant,
) -> std::io::Result<Option<std::fs::File>> {
    let file = open_lock_file(path)?;
    let poll_interval = std::time::Duration::from_millis(10);
    loop {
        // SAFETY: flock is a POSIX advisory lock with no memory side-effects.
        let rc = unsafe { libc::flock(file.as_raw_fd(), libc::LOCK_EX | libc::LOCK_NB) };
        if rc == 0 {
            return Ok(Some(file));
        }
        let err = std::io::Error::last_os_error();
        if err.raw_os_error() != Some(libc::EWOULDBLOCK) {
            return Err(err);
        }
        let now = std::time::Instant::now();
        if now >= deadline {
            return Ok(None);
        }
        std::thread::sleep(poll_interval.min(deadline - now));
    }
}

/// Bounded, deadline-aware variant of [`acquire_daemon_boot_guard`]: attempts
/// the SAME boot/recovery lock ([`lock_path`]) but gives up at `deadline`
/// instead of blocking forever. For callers that need to detect "is a boot in
/// progress right now" without risking an unbounded wait behind a wedged
/// holder: e.g. khive-mcp's `confirm_genuinely_dead` re-probing rounds,
/// where `DEAD_CONFIRM_ROUNDS` must bound elapsed time, not just probe count.
#[cfg(unix)]
pub fn try_acquire_daemon_boot_guard_until(
    deadline: std::time::Instant,
) -> std::io::Result<Option<DaemonBootGuard>> {
    try_acquire_flock_until(&lock_path(), deadline)
}

/// Bounded, deadline-aware acquisition of the recoverer-only lock
/// ([`recoverer_lock_path`]). See [`try_acquire_daemon_boot_guard_until`] for
/// the shared rationale — a second recoverer waiting for a peer's dead
/// confirmation/kill/spawn critical section must give up and report
/// "uncertain" rather than block forever if that peer is itself wedged.
#[cfg(unix)]
pub fn try_acquire_recoverer_lock_until(
    deadline: std::time::Instant,
) -> std::io::Result<Option<std::fs::File>> {
    try_acquire_flock_until(&recoverer_lock_path(), deadline)
}

/// Guard returned by [`acquire_daemon_boot_guard`], held across cold-boot
/// schema initialization (migrations + pack schema plans / FTS DDL) through
/// daemon bind + pid-write.
#[cfg(unix)]
pub type DaemonBootGuard = std::fs::File;

/// Acquire the recovery/boot lock, treating failure as fatal.
///
/// Unlike [`acquire_recovery_lock`] (best-effort, `None` on failure: used by
/// shutdown cleanup, where skipping unlink is safer than blocking forever),
/// daemon-mode boot must hold this lock across migrations/FTS DDL through
/// bind+pid-write. Silently continuing with no lock reopens the cold-boot FTS
/// race this guard exists to close, so callers that are about to run
/// daemon-mode boot (or wait for one to quiesce) must fail loudly instead of
/// proceeding unguarded.
#[cfg(unix)]
pub fn acquire_daemon_boot_guard() -> anyhow::Result<DaemonBootGuard> {
    acquire_recovery_lock()
        .ok_or_else(|| anyhow::anyhow!("failed to acquire daemon boot/recovery lock"))
}

/// Identity of a bound Unix socket path, used to tell "the socket I bound" apart
/// from "a same-path socket some other daemon bound after mine was removed".
///
/// A socket path can be recreated by a different process between the time
/// this daemon captures its identity and the time it later checks it, so `dev`
/// and `ino` (not the path) are what must match for cleanup to be safe.
#[cfg(unix)]
#[derive(Clone, Copy, PartialEq, Eq)]
struct SocketIdentity {
    dev: u64,
    ino: u64,
}

#[cfg(unix)]
fn socket_identity(path: &std::path::Path) -> Option<SocketIdentity> {
    use std::os::unix::fs::MetadataExt;
    let meta = std::fs::metadata(path).ok()?;
    Some(SocketIdentity {
        dev: meta.dev(),
        ino: meta.ino(),
    })
}

// ── connection principal ──────────────────────────────────────────────────────

/// The uid on the other end of an accepted connection, read from the kernel.
///
/// This is the only identity on this socket the caller cannot choose. Every
/// identity field on the request frame — `namespace`, `actor_id`,
/// `visible_namespaces`, `config_id` — is supplied by the connecting process,
/// so none of them can answer "who is this". A check reading self-asserted
/// fields is not a weak gate, it is not a gate: anyone who wants to pass it
/// asserts the passing values.
///
/// `getpeereid(2)` on macOS/BSD, `SO_PEERCRED` on Linux. Both report the peer's
/// credentials as recorded by the kernel at connect time.
#[cfg(unix)]
pub(crate) fn peer_uid(stream: &UnixStream) -> std::io::Result<u32> {
    use std::os::fd::AsRawFd;
    let fd = stream.as_raw_fd();

    #[cfg(any(target_os = "macos", target_os = "ios", target_vendor = "apple"))]
    {
        let mut uid: libc::uid_t = 0;
        let mut gid: libc::gid_t = 0;
        // SAFETY: `fd` is a live connected socket owned by `stream` for the
        // duration of this call; both out-params are valid initialized locals.
        let rc = unsafe { libc::getpeereid(fd, &mut uid, &mut gid) };
        if rc != 0 {
            return Err(std::io::Error::last_os_error());
        }
        Ok(uid as u32)
    }

    #[cfg(target_os = "linux")]
    {
        let mut cred = libc::ucred {
            pid: 0,
            uid: 0,
            gid: 0,
        };
        let mut len = std::mem::size_of::<libc::ucred>() as libc::socklen_t;
        // SAFETY: `fd` is a live connected socket owned by `stream`; `cred` is
        // an initialized local of exactly `len` bytes, which is what
        // SO_PEERCRED writes.
        let rc = unsafe {
            libc::getsockopt(
                fd,
                libc::SOL_SOCKET,
                libc::SO_PEERCRED,
                (&mut cred as *mut libc::ucred).cast::<libc::c_void>(),
                &mut len,
            )
        };
        if rc != 0 {
            return Err(std::io::Error::last_os_error());
        }
        Ok(cred.uid)
    }

    #[cfg(not(any(
        target_os = "linux",
        target_os = "macos",
        target_os = "ios",
        target_vendor = "apple"
    )))]
    {
        let _ = fd;
        Err(std::io::Error::new(
            std::io::ErrorKind::Unsupported,
            "peer-credential capture is not implemented for this platform",
        ))
    }
}

/// Whether a connection from `uid` may be served by this daemon.
///
/// ADR-096 accepted per-request identity threading **for the single-principal
/// owner-only socket only**, and rested that on three things: the `0600`
/// socket, all connections being the same uid, and the database being already
/// same-uid-accessible. The socket mode is asserted at bind. This asserts the
/// second, which previously had no representation in the code at all — nothing
/// read peer identity, so nothing could notice when it stopped being true.
///
/// **Principal is not attribution.** Many `actor_id`s over one socket is
/// exactly what ADR-096 shipped and what every seat on a normal host does;
/// refusing a second distinct actor would break the accepted design. The
/// principal is the uid, and this refuses only a genuinely foreign one.
///
/// **There is deliberately no configuration escape hatch.** A flag permitting
/// other uids would not weaken this assertion, it would delete it, in the way
/// hardest to notice later: the check still exists, its tests still pass, and
/// the deployment that matters has it off. A deployment that genuinely needs
/// multiple uids needs a code change and a gated ADR — which is precisely the
/// decision that should be impossible to make by accident.
#[cfg(unix)]
pub(crate) fn uid_is_permitted(peer: u32, daemon_euid: u32) -> bool {
    peer == daemon_euid
}

// ── wire types ────────────────────────────────────────────────────────────────

/// Request frame sent from a client to the daemon.
#[derive(Serialize, Deserialize, Default)]
pub struct DaemonRequestFrame {
    pub ops: String,
    pub presentation: Option<String>,
    pub presentation_per_op: Option<Vec<Option<String>>>,
    /// The client's resolved storage/gate default namespace for this request.
    ///
    /// As of protocol version 3 (ADR-096) the daemon serves the request under
    /// this namespace instead of rejecting on mismatch: a per-request
    /// identity input, not a same-process-identity assertion.
    pub namespace: String,
    /// The client's resolved write-stamp / gate actor identity (ADR-057),
    /// carried on the frame so the warm daemon stamps writes with the
    /// *caller's* actor instead of its own baked `actor_id` (ADR-096). `None`
    /// mints `ActorRef::anonymous()`, matching an unconfigured actor.
    #[serde(default)]
    pub actor_id: Option<String>,
    /// Opaque process provenance resolved in the originating client process.
    /// It is carried per request because a shared warm daemon's environment
    /// does not identify the worker that submitted the operation. Protocol v4
    /// makes this field part of dispatch semantics: a v3 daemon must reject the
    /// request rather than execute it while silently discarding provenance.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub process_ref: Option<String>,
    /// The client's resolved extra read-visibility namespaces (ADR-007 Rule
    /// 3b), carried on the frame so the warm daemon widens read scope to
    /// match the caller's own configuration rather than its own baked
    /// `visible_namespaces` (ADR-096). Empty means no extra visibility beyond
    /// `namespace` itself.
    #[serde(default)]
    pub visible_namespaces: Vec<String>,
    /// Fingerprint of the client's engine-coherence config: packs, db target,
    /// embedders, backend routing, and construction-baked outbound policy.
    /// Identity fields are carried separately in this frame. The daemon rejects
    /// a request whose `config_id` differs from its own so a restricted client
    /// (e.g. `--pack kg`, `--db :memory:`) never dispatches through the broader
    /// default daemon. See ADR-027 / ADR-049 / ADR-096.
    #[serde(default)]
    pub config_id: String,
    /// IPC protocol version sent by the client. Pre-versioning clients omit
    /// this field (deserializes to 0). The daemon compares against
    /// [`PROTOCOL_VERSION`] and rejects mismatches with an explicit error.
    #[serde(default)]
    pub protocol_version: u32,
    /// When `true`, the daemon returns an identity frame (ok=true, result=None)
    /// immediately after identity validation — without calling the dispatcher.
    /// Used by the client's under-lock recovery probe to confirm a daemon is
    /// alive and identity-matching without dispatching any mutating verb.
    /// Pre-probe clients omit this field (deserializes to false → normal dispatch).
    #[serde(default)]
    pub probe_only: bool,
    /// When `true`, the daemon returns a point-in-time [`MetricsSnapshot`] of
    /// its server-side gauges (a read-only measurement surface for the
    /// load/perf harness) instead of dispatching any op. Handled before the
    /// `config_id` equality reject: a gauge read is process-global and
    /// namespace/config-agnostic, not a namespaced record operation.
    /// READ-ONLY — this field is the only input the frame accepts for a
    /// metrics request; there is no reset or mutation reachable over the
    /// wire. Pre-metrics clients omit this field (deserializes to `false` →
    /// normal dispatch, unaffected).
    #[serde(default)]
    pub metrics_only: bool,
    /// Output format for this request (ADR-078). Forwarded to the daemon's
    /// serialization seam. `None` means use the daemon's resolved default.
    #[serde(default)]
    pub format: Option<String>,
    /// Per-operation output format overrides (ADR-078).
    #[serde(default)]
    pub format_per_op: Option<Vec<Option<String>>>,
    /// Whether this request originated from the agent-facing MCP `request`
    /// tool (the wire surface). When `true`, the daemon rejects
    /// `Visibility::Subhandler` verbs: agents must not invoke internal
    /// subhandlers. When `false` (the default, and the only value any
    /// operator path sends), subhandlers are allowed: `kkernel exec` and
    /// other in-process callers are trusted operator surfaces.
    ///
    /// This is the origin discriminator, not a daemon-vs-local one: operator
    /// requests flow through the daemon by default too, so the gate cannot
    /// key on transport.
    #[serde(default)]
    pub from_wire: bool,
    /// Caller-supplied correlation id (khive#948): a `u64` from the caller's
    /// own process-local monotonic counter, echoed back unchanged on
    /// [`DaemonResponseFrame::request_id`] and stamped into the dispatch's
    /// audit event (`resource.request_id`) so a benchmark harness can join
    /// its own pre-send sample to the server-side audit row for the same
    /// request. Purely additive — `#[serde(default)]` matches
    /// `metrics_only`/`format`/`format_per_op` precedent, no
    /// `PROTOCOL_VERSION` bump. `None` means the caller supplied no id.
    #[serde(default)]
    pub request_id: Option<u64>,
}

/// Response frame sent from the daemon back to a client.
#[derive(Serialize, Deserialize, Debug)]
pub struct DaemonResponseFrame {
    pub ok: bool,
    pub result: Option<String>,
    pub error: Option<String>,
    pub namespace_mismatch: bool,
    /// Set when the request's `config_id` does not match the daemon's. Like
    /// `namespace_mismatch`, this signals the client to fall back to local
    /// dispatch rather than execute under a different runtime/config.
    #[serde(default)]
    pub config_mismatch: bool,
    /// The `config_id` the daemon dispatched under, echoed back so the client
    /// can positively confirm the result came from a matching runtime. A
    /// pre-`config_id` daemon omits this field (deserializes to `None`), which
    /// the client treats as a mismatch and falls back to local dispatch — this
    /// closes the upgrade window where a new restricted client could otherwise
    /// trust a still-warm legacy daemon's broader registry.
    #[serde(default)]
    pub served_config_id: Option<String>,
    /// Set when the client's `protocol_version` does not match the daemon's
    /// [`PROTOCOL_VERSION`]. The client must treat this as a hard error and
    /// surface the human-readable `error` field rather than falling back to
    /// local dispatch (which would hide the version skew).
    #[serde(default)]
    pub version_mismatch: bool,
    /// The daemon's [`PROTOCOL_VERSION`], echoed in error responses so the
    /// client can include both sides in the diagnostic message. Pre-versioning
    /// daemons omit this field (deserializes to 0).
    #[serde(default)]
    pub daemon_protocol_version: u32,
    /// Populated when the request set `metrics_only: true`: a point-in-time
    /// snapshot of the daemon's server-side gauges. `None` on every other
    /// response, and on any response from a daemon that predates this field
    /// (client-side back-compat via `#[serde(default)]`, matching
    /// `served_config_id`'s upgrade-window handling above).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub metrics: Option<MetricsSnapshot>,
    /// Echo of the request's `request_id` (khive#948), present whenever the
    /// frame that produced this response carried one — including on every
    /// error/denied arm, not only success, so a client can join a failure
    /// the same way it joins a success. `#[serde(default)]` so an older
    /// daemon's response (predating this field) deserializes to `None`
    /// rather than a parse error.
    #[serde(default)]
    pub request_id: Option<u64>,
}

/// Point-in-time snapshot of the daemon's server-side gauges — the
/// load/perf harness read-surface (measurement substrate, not a product feature).
///
/// Every field here is a **server-side** gauge reachable from `handle_conn`
/// without any mutation: [`khive_storage::tx_registry`] (ADR-091 Plank 0,
/// process-global singleton), the main pool's backend-keyed routine WAL
/// sample and process TRUNCATE counters (`khive_db::checkpoint`), and the
/// ADR-067 Component A write queue depth/latest writer-stage sample. There is
/// no reset reachable through this type or through [`DaemonRequestFrame`] —
/// gauges out, nothing in.
#[derive(Serialize, Deserialize, Debug, Clone, Default, PartialEq)]
pub struct MetricsSnapshot {
    /// Last-observed WAL page count from the periodic checkpoint tick.
    /// `None` when the checkpoint task has never ticked in this process
    /// (for example, an in-memory dispatcher with no pool, or a daemon that
    /// just started and hasn't hit its first tick yet).
    pub wal_pages: Option<u64>,
    /// Logical frames present in the WAL at the main backend's most recent
    /// periodic PASSIVE checkpoint. Kept separate from physical allocation.
    #[serde(default)]
    pub wal_log_frames: Option<u64>,
    /// Frames backfilled by that same periodic PASSIVE pass.
    #[serde(default)]
    pub wal_checkpointed_frames: Option<u64>,
    /// Logical frames still pending after that pass (`log - checkpointed`).
    #[serde(default)]
    pub wal_pending_frames: Option<u64>,
    /// Physical `-wal` sidecar bytes captured at the same routine tick.
    #[serde(default)]
    pub wal_physical_bytes: Option<u64>,
    /// Wall-clock timestamp of the routine WAL sample, for staleness checks.
    #[serde(default)]
    pub wal_observed_at_unix_ms: Option<u64>,
    /// Total WAL TRUNCATE escalation attempts (ADR-091 Plank 2) made in this
    /// process's lifetime, regardless of whether they succeeded in reclaiming
    /// pages.
    pub wal_truncate_attempts: u64,
    /// Current consecutive-failure count for TRUNCATE attempts that failed to
    /// bring the WAL back below `warn_pages`; resets to 0 the next time an
    /// attempt clears it.
    pub wal_truncate_consecutive_failures: u64,
    /// Total checkpoint ticks skipped because the dedicated checkpoint
    /// connection was unavailable (ADR-091 checkpoint-pressure telemetry),
    /// across this process's lifetime. `#[serde(default)]` so an older client
    /// decoding a newer daemon's snapshot (or vice versa) does not fail.
    #[serde(default)]
    pub wal_checkpoint_skipped_ticks: u64,
    /// Current consecutive-skip run length; 0 once the next tick is observed.
    #[serde(default)]
    pub wal_checkpoint_consecutive_skips: u64,
    /// WAL page count last known at the time of the most recent skip, if any
    /// skip has occurred yet in this process.
    #[serde(default)]
    pub wal_checkpoint_last_skip_wal_pages: Option<u64>,
    /// Age, in microseconds, of the oldest currently-open transaction
    /// registry entry (ADR-091 Plank 0). `None` when no transaction is
    /// currently open.
    pub oldest_pinned_tx_micros: Option<u64>,
    /// Diagnostic label of the oldest currently-open transaction registry
    /// entry, if any and if it was registered with one.
    pub oldest_pinned_tx_label: Option<String>,
    /// Number of currently open transaction registry entries.
    pub open_tx_count: usize,
    /// Current write-queue backlog depth (ADR-067 Component A): requests
    /// enqueued but not yet accepted by the `WriterTask` drain loop. `None`
    /// unless the write queue is enabled (`KHIVE_WRITE_QUEUE=1`) and a
    /// file-backed pool is available.
    pub write_queue_depth: Option<usize>,
    /// The write queue's configured bounded capacity
    /// (`PoolConfig::write_queue_capacity`), gated the same as
    /// `write_queue_depth`.
    pub write_queue_capacity: Option<usize>,
    /// Latest completed writer-task span: bounded-channel admission/backlog.
    #[serde(default)]
    pub write_last_queue_wait_micros: Option<u64>,
    /// Latest completed writer-task span: `BEGIN IMMEDIATE` acquisition.
    #[serde(default)]
    pub write_last_transaction_acquire_micros: Option<u64>,
    /// Latest completed writer-task span: application transaction body.
    #[serde(default)]
    pub write_last_body_micros: Option<u64>,
    /// Latest completed writer-task span: SQLite COMMIT/fsync phase.
    #[serde(default)]
    pub write_last_commit_micros: Option<u64>,
    /// Whole latest writer-task request span, retained for compatibility and
    /// comparison with the decomposed stages.
    #[serde(default)]
    pub write_last_total_micros: Option<u64>,
    /// Wall-clock timestamp of the writer-stage sample.
    #[serde(default)]
    pub write_last_observed_at_unix_ms: Option<u64>,
}

// ── framing ───────────────────────────────────────────────────────────────────

/// Read one length-prefixed frame (4-byte BE u32 length + JSON bytes).
#[cfg(unix)]
pub async fn read_frame<R>(stream: &mut R) -> std::io::Result<Vec<u8>>
where
    R: tokio::io::AsyncRead + Unpin,
{
    let mut len_buf = [0u8; 4];
    stream.read_exact(&mut len_buf).await?;
    let len = u32::from_be_bytes(len_buf) as usize;
    if len > MAX_FRAME_BYTES {
        return Err(std::io::Error::new(
            std::io::ErrorKind::InvalidData,
            format!("daemon frame of {len} bytes exceeds {MAX_FRAME_BYTES} cap"),
        ));
    }
    let mut buf = vec![0u8; len];
    stream.read_exact(&mut buf).await?;
    Ok(buf)
}

/// Write one length-prefixed frame.
#[cfg(unix)]
pub async fn write_frame<W>(stream: &mut W, payload: &[u8]) -> std::io::Result<()>
where
    W: tokio::io::AsyncWrite + Unpin,
{
    if payload.len() > MAX_FRAME_BYTES {
        return Err(std::io::Error::new(
            std::io::ErrorKind::InvalidData,
            format!(
                "daemon frame of {} bytes exceeds {MAX_FRAME_BYTES} cap",
                payload.len()
            ),
        ));
    }
    let len = (payload.len() as u32).to_be_bytes();
    stream.write_all(&len).await?;
    stream.write_all(payload).await?;
    stream.flush().await?;
    Ok(())
}

// ── dispatch trait ────────────────────────────────────────────────────────────

/// Transport-agnostic dispatch interface for the daemon server.
///
/// The MCP crate implements this by dispatching through the shared request body
/// while honoring [`DaemonRequestFrame::from_wire`] (so subhandler visibility is
/// gated by request origin, not by transport); any future transport can do the
/// same.
#[cfg(unix)]
#[async_trait]
pub trait DaemonDispatch: Clone + Send + Sync + 'static {
    /// Dispatch a verb-DSL request string and return the rendered result.
    ///
    /// `from_wire` carries the origin discriminator from
    /// [`DaemonRequestFrame::from_wire`]: when `true`, the implementor enforces
    /// verb visibility (rejects `Visibility::Subhandler` verbs); when `false`,
    /// the request is from a trusted operator surface and subhandlers pass.
    ///
    /// `identity` is the per-request identity context threaded from the frame
    /// (ADR-096): `Some(..)` when serving a request forwarded over the
    /// daemon socket (built from `frame.namespace` / `frame.actor_id` /
    /// `frame.visible_namespaces` by the connection handler), `None` for any
    /// other dispatch path. Implementors should mint the storage/gate token from
    /// `identity` when present and fall back to their own construction-baked
    /// identity when absent, so pure local (non-daemon) dispatch is unchanged.
    #[allow(clippy::too_many_arguments)]
    async fn dispatch(
        &self,
        ops: String,
        presentation: Option<String>,
        presentation_per_op: Option<Vec<Option<String>>>,
        format: Option<String>,
        format_per_op: Option<Vec<Option<String>>>,
        from_wire: bool,
        identity: Option<RequestIdentity>,
    ) -> Result<String, String>;

    /// Warm every pack's in-memory state (ANN indexes, etc.).
    async fn warm_all(&self);

    /// The namespace this dispatcher was configured for.
    fn namespace(&self) -> &str;

    /// Fingerprint of this dispatcher's resolved runtime config (packs, db
    /// target, embedders). Used to reject forwarded requests from clients whose
    /// config differs, so a restricted client cannot dispatch through a broader
    /// daemon.
    fn config_id(&self) -> &str;

    /// Return the pool to use for background WAL checkpointing, if available.
    ///
    /// Implementors backed by a file-based SQLite database should return
    /// `Some(pool_arc)`. In-memory or test dispatchers that have no pool
    /// return `None` and the checkpoint task is not spawned.
    ///
    /// The default implementation returns `None`.
    fn pool_for_checkpoint(&self) -> Option<Arc<ConnectionPool>> {
        None
    }

    /// File-backed backend pools beyond [`Self::pool_for_checkpoint`]'s pool
    /// (ADR-091 Amendment 3): one checkpoint task is spawned per entry here,
    /// in addition to the one spawned for the primary pool, so a
    /// multi-backend deployment gets PASSIVE/TRUNCATE checkpointing and
    /// sidecar enumeration on every file-backed backend it wired, not only
    /// the main one.
    ///
    /// The default implementation returns an empty `Vec` — an implementor
    /// with only one backend (or none) needs no override.
    fn secondary_pools_for_checkpoint(&self) -> Vec<Arc<ConnectionPool>> {
        Vec::new()
    }

    /// Return the audit `EventStore` the checkpoint task should append
    /// ADR-094 lifecycle events (`CheckpointOutcomeRecorded`) to, if any.
    ///
    /// Mirrors [`Self::pool_for_checkpoint`]'s default-`None` shape: an
    /// implementor with no configured event store (or no pool at all) simply
    /// gets a checkpoint task that never appends events — the checkpoint
    /// task itself remains fully functional either way.
    ///
    /// The default implementation returns `None`.
    fn event_store_for_checkpoint(&self) -> Option<Arc<dyn khive_storage::EventStore>> {
        None
    }
}

#[cfg(unix)]
struct CheckpointTaskSpec {
    pool: Arc<ConnectionPool>,
    lifecycle_owner: Option<CheckpointLifecycleOwner>,
    is_main: bool,
}

/// Build checkpoint-task fan-out and designate one lifecycle owner.
///
/// The main checkpoint task owns lifecycle emission when it exists. If the
/// main backend is in-memory and therefore has no checkpoint task, the first
/// file-backed secondary owns emission instead. All remaining tasks are
/// explicit non-owners.
#[cfg(unix)]
fn checkpoint_task_specs(
    main_pool: Option<Arc<ConnectionPool>>,
    secondary_pools: Vec<Arc<ConnectionPool>>,
    event_store: Option<Arc<dyn khive_storage::EventStore>>,
    namespace: String,
) -> Vec<CheckpointTaskSpec> {
    let mut tasks = Vec::with_capacity(usize::from(main_pool.is_some()) + secondary_pools.len());
    if let Some(pool) = main_pool {
        tasks.push(CheckpointTaskSpec {
            pool,
            lifecycle_owner: None,
            is_main: true,
        });
    }
    tasks.extend(secondary_pools.into_iter().map(|pool| CheckpointTaskSpec {
        pool,
        lifecycle_owner: None,
        is_main: false,
    }));

    if let (Some(task), Some(event_store)) = (tasks.first_mut(), event_store) {
        task.lifecycle_owner = Some(CheckpointLifecycleOwner::new(event_store, namespace));
    }
    tasks
}

// ── tracked background tasks ─────────────────────────────────────────────────
//
// Pack handlers (e.g. memory.recall's ADR-081 serve-ledger append) fire
// fire-and-forget `tokio::spawn`ed work off the response path so the caller
// never waits on a cross-pack dispatch or a SQL write. Left untracked, that
// work is invisible to `drain()`: a SIGTERM landing between the response
// returning and the spawned task completing can abort it mid-flight with no
// log and no row. `track_background_task` gives such spawns a process-wide
// presence that `drain()` waits on, exactly like the `active` counter does
// for in-flight connections: the caller still only pays for the spawn +
// counter increment, never the task's own work.
static BACKGROUND_TASKS: std::sync::OnceLock<Arc<std::sync::atomic::AtomicUsize>> =
    std::sync::OnceLock::new();

fn background_tasks() -> &'static Arc<std::sync::atomic::AtomicUsize> {
    BACKGROUND_TASKS.get_or_init(|| Arc::new(std::sync::atomic::AtomicUsize::new(0)))
}

/// Decrements the shared background-task counter from `Drop`, so the count
/// comes back down whether the tracked future returns normally, panics, or
/// is cancelled — a plain post-`await` `fetch_sub` only covers the return
/// path and leaks the count forever on a panic, since unwinding skips every
/// statement after the panic point.
struct BackgroundTaskGuard {
    counter: Arc<std::sync::atomic::AtomicUsize>,
}

impl Drop for BackgroundTaskGuard {
    fn drop(&mut self) {
        self.counter
            .fetch_sub(1, std::sync::atomic::Ordering::SeqCst);
    }
}

/// Spawn a task that daemon shutdown's `drain()` waits for and return its join
/// handle. Retaining the handle lets boot coordinators form an explicit barrier;
/// dropping it deliberately detaches the task while the background counter still
/// keeps daemon drain aware of its lifetime. The decrement happens via
/// `BackgroundTaskGuard`'s `Drop`, including panic and cancellation paths.
pub fn spawn_tracked_task<F, T>(fut: F) -> tokio::task::JoinHandle<T>
where
    F: std::future::Future<Output = T> + Send + 'static,
    T: Send + 'static,
{
    background_tasks().fetch_add(1, std::sync::atomic::Ordering::SeqCst);
    let guard = BackgroundTaskGuard {
        counter: background_tasks().clone(),
    };
    tokio::spawn(async move {
        let _guard = guard;
        fut.await
    })
}

/// Spawn a fire-and-forget task through [`spawn_tracked_task`].
///
/// Callers that need a boot or shutdown barrier should retain and await the
/// returned handle from [`spawn_tracked_task`] instead of detaching it here.
pub fn track_background_task<F>(fut: F)
where
    F: std::future::Future<Output = ()> + Send + 'static,
{
    drop(spawn_tracked_task(fut));
}

/// Current count of in-flight tasks started via [`track_background_task`].
/// Exposed for tests; `drain()` reads the shared counter directly.
pub fn background_task_count() -> usize {
    background_tasks().load(std::sync::atomic::Ordering::SeqCst)
}

/// Process-wide daemon shutdown signal (ADR-119).
///
/// Cancelled exactly once, when the daemon's unified shutdown future resolves
/// — before `drain()` begins waiting on tracked tasks — so long-running
/// daemon components supervised outside this module observe shutdown through
/// the same path the daemon itself does, rather than inventing their own.
/// Clones share the underlying token; child tokens derived from it are
/// cancelled transitively.
///
/// In non-daemon processes the token simply never fires.
pub fn daemon_shutdown_token() -> tokio_util::sync::CancellationToken {
    static TOKEN: std::sync::OnceLock<tokio_util::sync::CancellationToken> =
        std::sync::OnceLock::new();
    TOKEN
        .get_or_init(tokio_util::sync::CancellationToken::new)
        .clone()
}

// ── active background phase names (ADR-103) ──────────────────────────────────
//
// A lightweight, best-effort process-wide gauge of which named background
// phases (e.g. `ann_warm`) are in flight right now, read by `comm.health`'s
// resource self-report so a caller can see "what is the daemon doing" at a
// glance without correlating timestamps across the event log itself. Counted
// per name rather than boolean, since more than one occurrence of the same
// named phase can legitimately overlap (e.g. two embedding models warming
// concurrently) — the name only drops out of the reported set once every
// concurrent occurrence has ended.
static ACTIVE_PHASES: std::sync::OnceLock<
    std::sync::Mutex<std::collections::HashMap<String, usize>>,
> = std::sync::OnceLock::new();

fn active_phases() -> &'static std::sync::Mutex<std::collections::HashMap<String, usize>> {
    ACTIVE_PHASES.get_or_init(|| std::sync::Mutex::new(std::collections::HashMap::new()))
}

/// RAII guard for one occurrence of a named background phase. Increments the
/// phase's count on creation (see [`register_active_phase`]); decrements on
/// `Drop`, so the count comes back down whether the guarded work returns
/// normally, panics, or is cancelled — the same rationale as
/// `BackgroundTaskGuard` above.
pub struct PhaseGuard {
    name: String,
}

impl Drop for PhaseGuard {
    fn drop(&mut self) {
        let mut map = active_phases()
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        if let Some(count) = map.get_mut(&self.name) {
            *count -= 1;
            if *count == 0 {
                map.remove(&self.name);
            }
        }
    }
}

/// Register one occurrence of a named background phase as currently active.
/// Returns a guard: drop it (or let it fall out of scope) when the phase
/// ends. Best-effort process-wide gauge only, read by `comm.health` — never
/// load-bearing for correctness.
pub fn register_active_phase(name: &str) -> PhaseGuard {
    let mut map = active_phases()
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner);
    *map.entry(name.to_string()).or_insert(0) += 1;
    PhaseGuard {
        name: name.to_string(),
    }
}

/// Currently active background-phase names, sorted for deterministic output.
/// Empty when no tracked phase is in flight.
pub fn active_phase_names() -> Vec<String> {
    let map = active_phases()
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner);
    let mut names: Vec<String> = map.keys().cloned().collect();
    names.sort();
    names
}

// ── server ────────────────────────────────────────────────────────────────────

/// Build a point-in-time [`MetricsSnapshot`] of this process's server-side
/// gauges. Called only from `handle_conn`'s `metrics_only` arm — a
/// process-global, read-only assembly with no side effects of its own.
/// See `docs/api/daemon.md#build_metrics_snapshot` for where each gauge is sourced
/// from and why.
#[cfg(unix)]
fn build_metrics_snapshot<D: DaemonDispatch>(dispatcher: &D) -> MetricsSnapshot {
    let open_tx_count = khive_storage::tx_registry::snapshot().len();
    // ADR-091 Amendment 3: deliberately the process-wide aggregate, not a
    // backend-attributed view — this gauge reports "oldest pinned tx in this
    // process" across every wired backend, not an attribution claim about
    // which database it belongs to. Attribution consumers (the session sweep,
    // the per-backend checkpoint tasks) use the scoped `oldest_for` views in
    // khive-db instead.
    let (oldest_pinned_tx_micros, oldest_pinned_tx_label) =
        match khive_storage::tx_registry::oldest() {
            Some((_id, age, label)) => (Some(age.as_micros() as u64), label),
            None => (None, None),
        };

    let checkpoint_pool = dispatcher.pool_for_checkpoint();
    let routine_wal = checkpoint_pool
        .as_deref()
        .and_then(khive_db::checkpoint::routine_wal_observation);
    let writer_stages = checkpoint_pool
        .as_deref()
        .and_then(khive_db::writer_task::last_writer_stage_observation);
    let (write_queue_depth, write_queue_capacity) = checkpoint_pool
        .as_ref()
        .and_then(|pool| pool.writer_task_handle().ok().flatten())
        .map(|handle| (Some(handle.queue_depth()), Some(handle.capacity())))
        .unwrap_or((None, None));

    MetricsSnapshot {
        wal_pages: routine_wal.as_ref().map(|sample| sample.log_frames),
        wal_log_frames: routine_wal.as_ref().map(|sample| sample.log_frames),
        wal_checkpointed_frames: routine_wal
            .as_ref()
            .map(|sample| sample.checkpointed_frames),
        wal_pending_frames: routine_wal.as_ref().map(|sample| sample.pending_frames),
        wal_physical_bytes: routine_wal
            .as_ref()
            .and_then(|sample| sample.physical_wal_bytes),
        wal_observed_at_unix_ms: routine_wal
            .as_ref()
            .map(|sample| sample.observed_at_unix_ms),
        wal_truncate_attempts: khive_db::checkpoint::truncate_attempts(),
        wal_truncate_consecutive_failures: khive_db::checkpoint::truncate_consecutive_failures(),
        wal_checkpoint_skipped_ticks: khive_db::checkpoint::checkpoint_skipped_ticks(),
        wal_checkpoint_consecutive_skips: khive_db::checkpoint::checkpoint_consecutive_skips(),
        wal_checkpoint_last_skip_wal_pages: khive_db::checkpoint::checkpoint_last_skip_wal_pages(),
        oldest_pinned_tx_micros,
        oldest_pinned_tx_label,
        open_tx_count,
        write_queue_depth,
        write_queue_capacity,
        write_last_queue_wait_micros: writer_stages
            .as_ref()
            .map(|sample| sample.queue_wait_micros),
        write_last_transaction_acquire_micros: writer_stages
            .as_ref()
            .map(|sample| sample.transaction_acquire_micros),
        write_last_body_micros: writer_stages.as_ref().map(|sample| sample.body_micros),
        write_last_commit_micros: writer_stages.as_ref().map(|sample| sample.commit_micros),
        write_last_total_micros: writer_stages.as_ref().map(|sample| sample.total_micros),
        write_last_observed_at_unix_ms: writer_stages
            .as_ref()
            .map(|sample| sample.observed_at_unix_ms),
    }
}

#[cfg(unix)]
async fn wait_for_peer_disconnect(read: &mut tokio::net::unix::OwnedReadHalf) {
    let mut byte = [0u8; 1];
    // One request is admitted per connection. EOF, a read error, or any
    // subsequent byte all make this connection no longer a valid response
    // peer, so the first completed read is the entire observation.
    let _ = read.read(&mut byte).await;
}

#[cfg(all(unix, test))]
async fn handle_conn<D: DaemonDispatch>(stream: UnixStream, dispatcher: D) {
    handle_conn_with_shutdown(stream, dispatcher, None).await;
}

#[cfg(unix)]
async fn handle_conn_with_shutdown<D: DaemonDispatch>(
    mut stream: UnixStream,
    dispatcher: D,
    shutdown: Option<tokio::sync::watch::Receiver<bool>>,
) {
    let (local_shutdown_tx, local_shutdown_rx) = tokio::sync::watch::channel(false);
    let shutdown = shutdown.unwrap_or(local_shutdown_rx);
    // Keeps the fallback receiver open in direct/test calls. Production owns
    // a sender at the daemon-run scope and passes its receiver above.
    let _local_shutdown_tx = local_shutdown_tx;
    let raw = match read_frame(&mut stream).await {
        Ok(r) => r,
        Err(e) => {
            tracing::debug!(error = %e, "failed to read daemon request frame");
            return;
        }
    };
    let frame: DaemonRequestFrame = match serde_json::from_slice(&raw) {
        Ok(f) => f,
        Err(e) => {
            tracing::debug!(error = %e, "failed to decode daemon request frame");
            return;
        }
    };
    let (mut peer_read, mut peer_write) = stream.into_split();

    let served_config_id = Some(dispatcher.config_id().to_string());
    let resp = if frame.protocol_version != PROTOCOL_VERSION {
        let msg = format!(
            "daemon protocol mismatch: client={} daemon={} — \
             rebuild/update the client binary (make local)",
            frame.protocol_version, PROTOCOL_VERSION,
        );
        tracing::warn!(
            client_version = frame.protocol_version,
            daemon_version = PROTOCOL_VERSION,
            "daemon protocol version mismatch"
        );
        DaemonResponseFrame {
            ok: false,
            result: None,
            error: Some(msg),
            namespace_mismatch: false,
            config_mismatch: false,
            served_config_id,
            version_mismatch: true,
            daemon_protocol_version: PROTOCOL_VERSION,
            metrics: None,
            request_id: frame.request_id,
        }
    } else if frame.metrics_only {
        // Process-global gauge read: namespace/config-agnostic, so this is
        // handled BEFORE the `config_id` equality reject below (unlike every
        // other arm) — a metrics probe must work regardless of which
        // client's config is asking, since it never touches the dispatcher's
        // packs/db/embed registry. READ-ONLY: builds a snapshot and returns
        // immediately, never reaching the ops-dispatch arm.
        DaemonResponseFrame {
            ok: true,
            result: None,
            error: None,
            namespace_mismatch: false,
            config_mismatch: false,
            served_config_id,
            version_mismatch: false,
            daemon_protocol_version: PROTOCOL_VERSION,
            metrics: Some(build_metrics_snapshot(&dispatcher)),
            request_id: frame.request_id,
        }
    // There is no `frame.namespace != dispatcher.namespace()` reject here.
    // The daemon accepts and serves the request under the frame's own
    // identity (namespace / actor / visible_namespaces, built into a
    // `RequestIdentity` below) over its one shared warm registry, rather
    // than rejecting a differently-attributed same-uid connection to a cold
    // local-dispatch fallback. `config_id`: which governs packs/db/embed
    // coherence for the shared warm engine: remains a hard reject; it is
    // not an identity field and softening it would let a restricted client
    // dispatch through an incompatible broader daemon.
    } else if frame.config_id != dispatcher.config_id() {
        DaemonResponseFrame {
            ok: false,
            result: None,
            error: None,
            namespace_mismatch: false,
            config_mismatch: true,
            served_config_id,
            version_mismatch: false,
            daemon_protocol_version: PROTOCOL_VERSION,
            metrics: None,
            request_id: frame.request_id,
        }
    } else if frame.probe_only {
        // Probe-only request: identity checks passed; return immediately without
        // dispatching any verb. The client uses this to confirm the daemon is
        // alive and identity-matching without triggering any mutation.
        DaemonResponseFrame {
            ok: true,
            result: None,
            error: None,
            namespace_mismatch: false,
            config_mismatch: false,
            served_config_id,
            version_mismatch: false,
            daemon_protocol_version: PROTOCOL_VERSION,
            metrics: None,
            request_id: frame.request_id,
        }
    } else {
        // Build the per-request identity context from the frame so the
        // implementor mints the storage/gate token from the CALLER's
        // identity, not the dispatcher's own construction-baked scalars.
        // This is always `Some` here: every frame that reaches this arm
        // carries a `namespace` (required on the wire) plus whatever
        // `actor_id`/`visible_namespaces` the client resolved (defaulting to
        // `None`/`vec![]` for an older, field-absent payload, which is
        // exactly the prior anonymous/no-extra-visibility behavior).
        let identity = RequestIdentity {
            namespace: frame.namespace.clone(),
            actor_id: frame.actor_id.clone(),
            visible_namespaces: frame.visible_namespaces.clone(),
            process_ref: frame.process_ref.clone(),
            request_id: frame.request_id,
        };
        let (read_cancel_tx, read_cancel_rx) = tokio::sync::watch::channel(false);
        let dispatch = khive_storage::scope_request_read_cancellation(
            shutdown,
            khive_storage::scope_request_read_cancellation(
                read_cancel_rx,
                khive_storage::scope_request_read_deadline(
                    khive_storage::request_read_timeout_from_env(),
                    dispatcher.dispatch(
                        frame.ops,
                        frame.presentation,
                        frame.presentation_per_op,
                        frame.format,
                        frame.format_per_op,
                        frame.from_wire,
                        Some(identity),
                    ),
                ),
            ),
        );
        tokio::pin!(dispatch);
        let dispatch_result = tokio::select! {
            result = &mut dispatch => result,
            _ = wait_for_peer_disconnect(&mut peer_read) => {
                let _ = read_cancel_tx.send(true);
                dispatch.await
            }
        };
        match dispatch_result {
            Ok(result) => DaemonResponseFrame {
                ok: true,
                result: Some(result),
                error: None,
                namespace_mismatch: false,
                config_mismatch: false,
                served_config_id,
                version_mismatch: false,
                daemon_protocol_version: PROTOCOL_VERSION,
                metrics: None,
                request_id: frame.request_id,
            },
            Err(e) => DaemonResponseFrame {
                ok: false,
                result: None,
                error: Some(e),
                namespace_mismatch: false,
                config_mismatch: false,
                served_config_id,
                version_mismatch: false,
                daemon_protocol_version: PROTOCOL_VERSION,
                metrics: None,
                request_id: frame.request_id,
            },
        }
    };

    match serde_json::to_vec(&resp) {
        Ok(payload) => {
            if payload.len() > MAX_FRAME_BYTES {
                // The serialized response exceeds the IPC frame cap.  Send a
                // small explicit error frame so the client can distinguish a
                // per-request payload-size failure from a daemon crash.  A
                // client that receives this error frame will NOT trigger
                // stale-daemon kill/respawn (ParseFailure requires a read_frame
                // error, not an ok=false result).
                tracing::warn!(
                    bytes = payload.len(),
                    limit = MAX_FRAME_BYTES,
                    "daemon response exceeds MAX_FRAME_BYTES; sending explicit error frame"
                );
                let err_resp = DaemonResponseFrame {
                    ok: false,
                    result: None,
                    error: Some(format!(
                        "response too large: {} bytes exceeds {} byte IPC cap",
                        payload.len(),
                        MAX_FRAME_BYTES,
                    )),
                    namespace_mismatch: false,
                    config_mismatch: false,
                    served_config_id: resp.served_config_id,
                    version_mismatch: false,
                    daemon_protocol_version: PROTOCOL_VERSION,
                    metrics: None,
                    request_id: resp.request_id,
                };
                if let Ok(err_payload) = serde_json::to_vec(&err_resp) {
                    if let Err(e) = write_frame(&mut peer_write, &err_payload).await {
                        tracing::debug!(error = %e, "failed to write oversized-response error frame");
                    }
                }
            } else if let Err(e) = write_frame(&mut peer_write, &payload).await {
                tracing::debug!(error = %e, "failed to write daemon response frame");
            }
        }
        Err(e) => tracing::warn!(error = %e, "failed to serialize daemon response frame"),
    }
}

/// An accepted connection owns a drain slot from the accept loop until its
/// handler future is dropped. The claim happens before spawning so shutdown
/// cannot observe zero in the gap between `accept()` and the task's first poll.
#[cfg(unix)]
struct ActiveConnectionGuard {
    active: Arc<std::sync::atomic::AtomicUsize>,
}

#[cfg(unix)]
impl ActiveConnectionGuard {
    fn claim(active: Arc<std::sync::atomic::AtomicUsize>) -> Self {
        active.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
        Self { active }
    }
}

#[cfg(unix)]
impl Drop for ActiveConnectionGuard {
    fn drop(&mut self) {
        self.active
            .fetch_sub(1, std::sync::atomic::Ordering::SeqCst);
    }
}

#[cfg(unix)]
fn spawn_connection_task<F>(
    active: Arc<std::sync::atomic::AtomicUsize>,
    future: F,
) -> tokio::task::JoinHandle<()>
where
    F: std::future::Future<Output = ()> + Send + 'static,
{
    let guard = ActiveConnectionGuard::claim(active);
    tokio::spawn(async move {
        let _guard = guard;
        future.await;
    })
}

/// Run the daemon: bind the socket, warm in the background, serve request
/// frames until SIGTERM/SIGINT.
///
/// Fatally acquires its own startup lock, which only protects
/// cleanup→bind→pid-write — `dispatcher` has already run migrations and
/// applied pack schema plans while constructing itself, unguarded. Production
/// boot must go through [`run_daemon_with_boot_guard`] instead, which extends
/// the same lock back over construction. This entry point is for callers
/// (and tests) that build the dispatcher and start serving as one atomic
/// step with no separate boot-guard window to protect.
#[cfg(unix)]
pub async fn run_daemon<D: DaemonDispatch>(dispatcher: D) -> anyhow::Result<()> {
    let boot_guard = Some(acquire_daemon_boot_guard()?);
    run_daemon_with_boot_guard_inner(dispatcher, boot_guard, false).await
}

/// Run a real daemon server for an in-process multi-launch test.
///
/// Separate production daemon candidates have distinct PIDs, so the boot fence
/// recognizes a responsive incumbent and makes later candidates exit. Parallel
/// test launchers share one OS process and therefore one PID; this explicit
/// fault-injection entry point preserves the production fence semantics by
/// allowing a responsive same-PID incumbent to win. Ordinary daemon startup
/// continues to treat a same-PID rendezvous as stale, protecting PID-reuse
/// cleanup behavior.
///
/// A losing candidate still follows the ordinary daemon-exit path and cancels
/// the process-wide component shutdown token. Callers must therefore use a
/// component-free dispatcher; this seam validates socket/PID ownership, not
/// multi-candidate component lifecycle.
#[cfg(all(unix, any(test, feature = "fault-injection")))]
#[doc(hidden)]
pub async fn run_daemon_in_process_test<D: DaemonDispatch>(dispatcher: D) -> anyhow::Result<()> {
    let boot_guard = Some(acquire_daemon_boot_guard()?);
    run_daemon_with_boot_guard_inner(dispatcher, boot_guard, true).await
}

/// Vet the socket's parent directory, re-permissioning it only when it is the
/// directory khive owns by convention.
///
/// `KHIVE_SOCKET` takes an arbitrary path, and the previous unconditional
/// chmod-0700 of its parent was wrong in both directions: pointed at a shared
/// parent like `/tmp` it either failed outright for an ordinary user, or —
/// worse — succeeded when privileged and stripped access for every other
/// process on the machine. A directory we did not create is never modified.
///
/// What the directory must actually prevent is a *takeover of the socket
/// path*: the connection gate is the 0600 socket plus the accept-time
/// peer-uid check, but both defend the daemon's own socket — neither helps
/// once another local user can put *their* listener at the path clients
/// resolve. There are two ways a shared directory allows that, and the
/// sticky bit closes neither: a writer can *pre-bind* the predictable path
/// before this daemon starts (the sticky bit restricts unlinking, not
/// creating), and the directory's *owner* can unlink and rebind even in a
/// 1777 directory (sticky exempts the directory owner). So a caller-chosen
/// directory is served only when it is trusted end to end: owned by this
/// daemon's euid or root, and not writable by group or other at all. The
/// umask-default 0755 stays acceptable; shared sticky directories like
/// `/tmp` do not.
#[cfg(unix)]
pub(crate) fn ensure_socket_dir_is_trusted(parent: &std::path::Path) -> anyhow::Result<()> {
    // SAFETY: `geteuid` is always successful and takes no arguments.
    let daemon_euid = unsafe { libc::geteuid() } as u32;

    if parent == khive_dir() {
        std::fs::set_permissions(parent, std::fs::Permissions::from_mode(0o700)).map_err(|e| {
            anyhow::anyhow!(
                "refusing to start: cannot chmod 0700 {}: {e}. The khive directory must be \
                 owner-only — it is half of the same-uid guarantee this daemon enforces.",
                parent.display()
            )
        })?;
        return ensure_socket_path_is_swap_resistant(parent, daemon_euid);
    }

    // Fail closed on the stat itself: not being able to read the metadata is
    // not the same as the directory passing.
    let meta = std::fs::metadata(parent).map_err(|e| {
        anyhow::anyhow!(
            "refusing to start: cannot stat {}: {e}. The socket directory gates \
             socket-takeover safety, and unreadable metadata is not a passing state.",
            parent.display()
        )
    })?;

    use std::os::unix::fs::MetadataExt;
    let owner = meta.uid();
    if owner != daemon_euid && owner != 0 {
        anyhow::bail!(
            "refusing to start: socket directory {} is owned by uid {owner}, not this \
             daemon's uid ({daemon_euid}) or root. A directory owner can replace the \
             socket regardless of mode bits. Point KHIVE_SOCKET at a directory you own, \
             or unset it for the default.",
            parent.display()
        );
    }

    let mode = meta.permissions().mode();
    if mode & 0o022 != 0 {
        anyhow::bail!(
            "refusing to start: socket directory {} is mode {:04o} — writable by group or \
             other, so another local user could bind their own listener at the socket path \
             (before this daemon starts, the sticky bit does not prevent creating the \
             path). Use a directory only you can write, or unset KHIVE_SOCKET for the \
             default. This daemon is not changing the permissions of a directory it does \
             not own.",
            parent.display(),
            mode & 0o7777
        );
    }

    ensure_socket_path_is_swap_resistant(parent, daemon_euid)
}

/// Walk the socket directory path exactly as the kernel will traverse it at
/// bind time — component by component, following symlinks — and refuse any
/// node another local user could swap after this validation.
///
/// Checking only the canonicalized result is not enough: bind and every
/// client traverse the *original* path, so a symlink component owned by
/// another user can resolve somewhere trusted while it is being validated
/// and be retargeted before the socket is bound. Validating the nodes the
/// traversal actually visits closes that gap: a symlink component is
/// acceptable only when its owner — the only party besides root who can
/// retarget it — is this daemon's euid or root, and every directory
/// component is held to the ancestor rule below.
///
/// The directory rule is deliberately weaker than the immediate parent's.
/// The parent hosts the socket file, where the threat is *creation* of the
/// predictable path — the sticky bit does not restrict creating, so no
/// sticky exception is sound there. An ancestor only threatens via
/// *rename/unlink of an existing entry we own*, which the sticky bit does
/// restrict: in a sticky directory, only the entry's owner, the directory's
/// owner, or root may rename it. So a root-owned `/tmp` (1777) is an
/// acceptable ancestor of a user-owned 0700 socket directory, while a
/// non-sticky group/other-writable ancestor, or one owned by a third uid
/// (who could rename the entry, or chmod the directory first), is refused.
/// Every stat failure fails closed.
#[cfg(unix)]
fn ensure_socket_path_is_swap_resistant(
    parent: &std::path::Path,
    daemon_euid: u32,
) -> anyhow::Result<()> {
    use std::os::unix::fs::MetadataExt;

    let absolute = if parent.is_absolute() {
        parent.to_path_buf()
    } else {
        std::env::current_dir()
            .map_err(|e| {
                anyhow::anyhow!(
                    "refusing to start: cannot resolve the working directory to absolutize \
                     socket directory {}: {e}.",
                    parent.display()
                )
            })?
            .join(parent)
    };

    fn push_components(stack: &mut Vec<std::ffi::OsString>, path: &std::path::Path) {
        let components: Vec<_> = path
            .components()
            .map(|c| c.as_os_str().to_os_string())
            .collect();
        stack.extend(components.into_iter().rev());
    }

    let mut stack: Vec<std::ffi::OsString> = Vec::new();
    push_components(&mut stack, &absolute);
    let mut resolved = std::path::PathBuf::new();
    let mut symlinks_followed = 0u32;

    while let Some(component) = stack.pop() {
        if component == "/" {
            resolved = std::path::PathBuf::from("/");
            continue;
        }
        if component == "." {
            continue;
        }
        if component == ".." {
            resolved.pop();
            continue;
        }
        let candidate = resolved.join(&component);
        let meta = std::fs::symlink_metadata(&candidate).map_err(|e| {
            anyhow::anyhow!(
                "refusing to start: cannot stat socket-path component {}: {e}. An \
                 unreadable component is not a passing one.",
                candidate.display()
            )
        })?;
        let owner = meta.uid();

        if meta.file_type().is_symlink() {
            symlinks_followed += 1;
            if symlinks_followed > 40 {
                anyhow::bail!(
                    "refusing to start: socket path resolves through more than 40 symlinks \
                     at {} — treating this as a loop.",
                    candidate.display()
                );
            }
            if owner != daemon_euid && owner != 0 {
                anyhow::bail!(
                    "refusing to start: socket-path symlink component {} is owned by uid \
                     {owner}, not this daemon's uid ({daemon_euid}) or root — its owner \
                     could retarget it after this check and re-root the socket path. Point \
                     KHIVE_SOCKET somewhere trusted end to end, or unset it for the \
                     default.",
                    candidate.display()
                );
            }
            let target = std::fs::read_link(&candidate).map_err(|e| {
                anyhow::anyhow!(
                    "refusing to start: cannot read socket-path symlink component {}: {e}.",
                    candidate.display()
                )
            })?;
            push_components(&mut stack, &target);
            continue;
        }

        if meta.is_dir() {
            let mode = meta.permissions().mode();
            let sticky = mode & 0o1000 != 0;
            if owner != daemon_euid && owner != 0 {
                anyhow::bail!(
                    "refusing to start: socket-path ancestor {} is owned by uid {owner}, not \
                     this daemon's uid ({daemon_euid}) or root — its owner could rename the \
                     next path component and re-root the socket path. Point KHIVE_SOCKET \
                     somewhere trusted end to end, or unset it for the default.",
                    candidate.display()
                );
            }
            if mode & 0o022 != 0 && !sticky {
                anyhow::bail!(
                    "refusing to start: socket-path ancestor {} is mode {:04o} — writable by \
                     group or other without the sticky bit, so another local user could rename \
                     the next path component and re-root the socket path. Point KHIVE_SOCKET \
                     somewhere trusted end to end, or unset it for the default.",
                    candidate.display(),
                    mode & 0o7777
                );
            }
            resolved = candidate;
            continue;
        }

        anyhow::bail!(
            "refusing to start: socket-path component {} is neither a directory nor a \
             symlink — the socket path cannot traverse it.",
            candidate.display()
        );
    }

    Ok(())
}

/// Run the daemon using a startup lock acquired by the caller *before*
/// building `dispatcher`, so a second process racing to boot (e.g. two
/// `kkernel mcp --daemon` spawns before either has bound its socket) cannot
/// run migrations/FTS DDL concurrently against the same database file.
/// `boot_guard` is only `None` on non-unix targets, where there is no
/// advisory boot lock to hold in the first place; every unix daemon-mode
/// caller passes `Some`.
///
/// The guard is held across cleanup → bind → pid-write, then dropped. The
/// caller must not still be holding a *different* handle to the same lock
/// file when this function is entered — see the "Deadlock note" on the
/// `_startup_lock` binding below for why that would self-deadlock on `flock`.
#[cfg(unix)]
pub async fn run_daemon_with_boot_guard<D: DaemonDispatch>(
    dispatcher: D,
    boot_guard: Option<std::fs::File>,
) -> anyhow::Result<()> {
    run_daemon_with_boot_guard_inner(dispatcher, boot_guard, false).await
}

#[cfg(unix)]
async fn run_daemon_with_boot_guard_inner<D: DaemonDispatch>(
    dispatcher: D,
    boot_guard: Option<std::fs::File>,
    allow_same_process_incumbent: bool,
) -> anyhow::Result<()> {
    // ADR-119: components may have been started by the serve path before this
    // function established (or failed to establish) daemon ownership. Cancel
    // the process-wide shutdown token on EVERY exit — setup failures below,
    // early return when another daemon owns the socket, bind errors, and
    // normal shutdown alike — so supervisors never outlive this process's
    // claim to daemon role. Constructed before any fallible startup work so
    // no error path can precede it. The token is process-lifetime
    // single-shot; a process that stops being (or never becomes) the daemon
    // has no path back except exec.
    struct ComponentTeardown;
    impl Drop for ComponentTeardown {
        fn drop(&mut self) {
            daemon_shutdown_token().cancel();
        }
    }
    let _component_teardown = ComponentTeardown;

    let sock = socket_path();
    let pid_file = pid_path();

    if let Some(parent) = sock.parent() {
        std::fs::create_dir_all(parent)?;
        ensure_socket_dir_is_trusted(parent)?;
    }

    // Hold the startup lock across cleanup → bind → pid-write so a concurrent
    // client's kill_and_respawn (which also holds this lock) cannot unlink the
    // socket between our bind and our pid-write.  The lock is released once the
    // listener is bound and the PID file is written — at that point any racing
    // client will find a live socket+pid and skip the stale-cleanup path.
    //
    // Deadlock note: the client holds this lock only during kill+spawn and
    // releases it before the spawned daemon process starts (the lock guard is
    // dropped when kill_and_respawn returns, before the readiness probe loop).
    // The daemon holds exactly one handle to this lock for its whole boot
    // sequence (received as `boot_guard`, extended from before `dispatcher`
    // was constructed) — never a second, independently-acquired handle in the
    // same process, which would self-deadlock on `flock`.
    let _startup_lock = boot_guard;

    if !cleanup_stale_daemon(&sock, &pid_file, allow_same_process_incumbent).await {
        tracing::info!("a responsive khived is already running; exiting");
        return Ok(());
    }

    let listener = UnixListener::bind(&sock)?;
    // Fail closed, same reason as the directory above. If this chmod fails the
    // socket is world-reachable in a way the accepted design never covered, so
    // the bound listener is dropped and the entry removed rather than served.
    if let Err(e) = std::fs::set_permissions(&sock, std::fs::Permissions::from_mode(0o600)) {
        drop(listener);
        let _ = std::fs::remove_file(&sock);
        return Err(anyhow::anyhow!(
            "refusing to start: cannot chmod 0600 {}: {e}. The daemon socket must be owner-only \
             — it is half of the single-principal guarantee this daemon enforces.",
            sock.display()
        ));
    }
    // Captured while still holding the startup lock, immediately after
    // bind, so shutdown cleanup can later prove "this is still the same socket
    // I bound" rather than trusting the path alone.
    let bound_identity = socket_identity(&sock);

    if let Err(e) = write_pid_file_exclusive(&pid_file) {
        if e.kind() == std::io::ErrorKind::AlreadyExists {
            // A PID file appeared between our own `cleanup_stale_daemon`
            // removing it and this write — only possible if the boot lock did
            // not actually exclude a concurrent booter (e.g. `acquire_recovery_lock`
            // failed for one side). Never touch the winner's files: drop only
            // the socket entry we ourselves just bound (proven via identity,
            // not path), then decide by checking whether the PID now on disk
            // names a live, reachable daemon.
            if bound_identity.is_some() && socket_identity(&sock) == bound_identity {
                drop(listener);
                let _ = std::fs::remove_file(&sock);
            }
            if pid_file_names_a_reachable_daemon(&pid_file, &sock, allow_same_process_incumbent)
                .await
            {
                tracing::info!(
                    "a replacement khived already claimed the pid/socket rendezvous; exiting"
                );
                return Ok(());
            }
            anyhow::bail!(
                "failed to claim daemon pid file at {pid_file:?}: it already exists \
                 and does not name a reachable daemon"
            );
        }
        return Err(e.into());
    }
    // Release the startup lock now: the listener is bound and the PID file is
    // written.  Any concurrent client or daemon startup will observe a live
    // socket+pid and take the non-recovery path.
    drop(_startup_lock);
    tracing::info!(
        socket = ?sock,
        pid = std::process::id(),
        source_revision = crate::BUILD_INFO.source_revision,
        build_time = crate::BUILD_INFO.build_time,
        "khived listening"
    );

    {
        let warm = dispatcher.clone();
        tokio::spawn(async move {
            warm.warm_all().await;
        });
    }

    // The checkpoint task's own strong-count-based exit is unreachable
    // whenever `event_store_for_checkpoint()` returns `Some` (the ordinary
    // production shape), because the `SqlEventStore` it wraps retains its
    // own clone of the same pool. An explicit watch channel replaces that
    // mechanism: the sender is held for the remainder of this function's
    // scope and signalled as the first action once shutdown is observed,
    // below.
    let (checkpoint_shutdown_tx, checkpoint_shutdown_rx) = tokio::sync::watch::channel(());
    // ADR-091 Amendment 3: one checkpoint task per file-backed backend the
    // dispatcher wired — the primary pool plus every entry
    // `secondary_pools_for_checkpoint` returns — sharing this one shutdown
    // channel (the sender broadcasts to every receiver clone), so the single
    // send below stops every spawned task before `drain()`.
    let checkpoint_tasks = checkpoint_task_specs(
        dispatcher.pool_for_checkpoint(),
        dispatcher.secondary_pools_for_checkpoint(),
        dispatcher.event_store_for_checkpoint(),
        dispatcher.namespace().to_string(),
    );
    if !checkpoint_tasks.is_empty() {
        let cfg = CheckpointConfig::from_env();
        let checkpoint_task_count = checkpoint_tasks.len();
        for task in checkpoint_tasks {
            track_background_task(run_checkpoint_task(
                task.pool,
                cfg.clone(),
                task.lifecycle_owner,
                checkpoint_shutdown_rx.clone(),
                task.is_main,
            ));
        }
        tracing::info!(checkpoint_task_count, "WAL checkpoint task(s) started");
    }

    let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let connection_tasks = Arc::new(std::sync::Mutex::new(
        Vec::<tokio::task::JoinHandle<()>>::new(),
    ));
    let (request_shutdown_tx, request_shutdown_rx) = tokio::sync::watch::channel(false);

    let shutdown = async {
        // REASON: signal handler registration can only fail if the global Tokio runtime
        // is not running or the OS rejects the signal number — both are unrecoverable
        // at this point in startup, so panic is the correct response.
        let mut sigterm = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
            .expect("install SIGTERM handler");
        let mut sigint = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::interrupt())
            .expect("install SIGINT handler");
        tokio::select! {
            _ = sigterm.recv() => tracing::info!("received SIGTERM"),
            _ = sigint.recv() => tracing::info!("received SIGINT"),
        }
    };

    // SAFETY: `geteuid` is always successful and takes no arguments.
    let daemon_euid = unsafe { libc::geteuid() } as u32;

    tokio::select! {
        _ = async {
            loop {
                match listener.accept().await {
                    Ok((stream, _)) => {
                        // Refuse a foreign uid before any frame is read.
                        // Fails CLOSED: an error reading peer credentials is
                        // "cannot prove same-uid", which is the same answer as
                        // "is not same-uid" — never a pass.
                        //
                        // Scope: this is a same-UID check, which is strictly
                        // weaker than same-PRINCIPAL. Several distinct actors
                        // running under one uid all pass here, so this does not
                        // by itself establish that the daemon serves a single
                        // principal, and nothing downstream refuses or degrades
                        // on observing more than one. Do not describe it as a
                        // multi-principal guard — it bounds the process boundary,
                        // not the identity one.
                        match peer_uid(&stream) {
                            Ok(peer) if uid_is_permitted(peer, daemon_euid) => {}
                            Ok(peer) => {
                                tracing::error!(
                                    peer_uid = peer,
                                    daemon_euid,
                                    "refusing connection from a foreign uid: this daemon accepts \
                                     only peers running as its own uid"
                                );
                                drop(stream);
                                continue;
                            }
                            Err(e) => {
                                tracing::error!(
                                    error = %e,
                                    "refusing connection: cannot read peer credentials, so \
                                     same-uid cannot be proven"
                                );
                                drop(stream);
                                continue;
                            }
                        }
                        let d = dispatcher.clone();
                        let shutdown = request_shutdown_rx.clone();
                        let handle = spawn_connection_task(Arc::clone(&active), async move {
                            handle_conn_with_shutdown(stream, d, Some(shutdown)).await;
                        });
                        let mut tasks = connection_tasks
                            .lock()
                            .unwrap_or_else(std::sync::PoisonError::into_inner);
                        tasks.retain(|task| !task.is_finished());
                        tasks.push(handle);
                    }
                    Err(e) => tracing::error!(error = %e, "accept failed"),
                }
            }
        } => {}
        _ = shutdown => {}
    }

    // Signal the checkpoint task to exit before draining, so `drain()`
    // actually waits on it via `track_background_task` rather than the
    // task outliving the drain window (or the process) unsignalled.
    let _ = checkpoint_shutdown_tx.send(());

    // Per-run signal: read scopes stop promptly, admitted writes ignore it and
    // retain the rest of the configured drain window to commit or roll back.
    let _ = request_shutdown_tx.send(true);

    // Same ordering contract for ADR-119 daemon components: cancel before
    // drain, so each component's supervisor (itself a tracked task) can run
    // its bounded shutdown inside the drain wait.
    daemon_shutdown_token().cancel();

    let drained = drain(&active).await;
    let tasks = {
        let mut retained = connection_tasks
            .lock()
            .unwrap_or_else(std::sync::PoisonError::into_inner);
        std::mem::take(&mut *retained)
    };
    finish_connection_tasks(tasks, drained).await;

    // A concurrent client's `kill_and_respawn` may have already decided
    // this daemon looked stale, killed it, and spawned a replacement that
    // bound the same socket/PID paths while this daemon was draining above.
    // Reacquire the recovery lock (the same one that serializes startup) and
    // only unlink if the PID file still names this process AND the socket at
    // `sock` is still the exact one this daemon bound — otherwise a
    // replacement daemon owns those paths now and unlinking would delete its
    // live socket/PID out from under it.
    match acquire_recovery_lock() {
        Some(_shutdown_lock) => {
            shutdown_cleanup_if_owned(&sock, &pid_file, bound_identity);
        }
        None => {
            tracing::warn!(
                "could not acquire recovery lock for shutdown cleanup; \
                 skipping unlink to avoid deleting a replacement daemon's paths"
            );
        }
    }
    tracing::info!("khived stopped");
    Ok(())
}

/// Remove `sock`/`pid_file` only if they still belong to this process: the PID
/// file must name `std::process::id()` AND the socket currently at `sock` must
/// still be the exact one identified by `bound_identity` (dev/ino, not path).
///
/// Returns `true` if cleanup ran, `false` if it was skipped because a
/// replacement daemon already owns those paths. The caller must hold
/// the recovery lock across this call — the same lock daemon startup holds
/// across cleanup+bind+pid-write — so no replacement can bind between this
/// function's checks and its unlinks.
#[cfg(unix)]
fn shutdown_cleanup_if_owned(
    sock: &std::path::Path,
    pid_file: &std::path::Path,
    bound_identity: Option<SocketIdentity>,
) -> bool {
    let pid_is_ours = std::fs::read_to_string(pid_file)
        .ok()
        .and_then(|s| s.trim().parse::<u32>().ok())
        == Some(std::process::id());
    let socket_is_ours = bound_identity.is_some() && socket_identity(sock) == bound_identity;
    if pid_is_ours && socket_is_ours {
        let _ = std::fs::remove_file(sock);
        let _ = std::fs::remove_file(pid_file);
        true
    } else {
        tracing::warn!(
            socket = ?sock,
            pid_file = ?pid_file,
            "skipping shutdown cleanup — a replacement daemon already owns this socket/PID"
        );
        false
    }
}

// ── helpers ───────────────────────────────────────────────────────────────────

/// Liveness verdict for a `kill(pid, 0)` probe.
#[cfg(unix)]
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum PidLiveness {
    /// errno 0 — signal delivery succeeded, the process exists and this
    /// caller may signal it.
    Alive,
    /// ESRCH (or any other non-EPERM errno) — no such process.
    Dead,
    /// EPERM — the process exists but this caller lacks permission to
    /// signal it. Unknown-safe: treated as running so stale-daemon cleanup
    /// never unlinks a live daemon's socket/PID file just because it is
    /// owned by a different user/uid.
    PermissionDenied,
}

#[cfg(unix)]
impl PidLiveness {
    fn is_running(self) -> bool {
        !matches!(self, PidLiveness::Dead)
    }
}

/// Maps a `kill(pid, 0)` outcome (return code + errno) to a [`PidLiveness`].
/// Pure and side-effect-free so the errno mapping can be unit tested without
/// a real process probe.
#[cfg(unix)]
fn classify_kill_result(rc: i32, errno: i32) -> PidLiveness {
    if rc == 0 {
        return PidLiveness::Alive;
    }
    match errno {
        libc::EPERM => PidLiveness::PermissionDenied,
        _ => PidLiveness::Dead,
    }
}

#[cfg(unix)]
fn is_process_running(pid: u32) -> bool {
    let Ok(pid) = i32::try_from(pid) else {
        return false;
    };
    if pid <= 0 {
        return false;
    }
    // SAFETY: signal 0 is an existence/permission probe with no side effects.
    let rc = unsafe { libc::kill(pid, 0) };
    let errno = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
    classify_kill_result(rc, errno).is_running()
}

/// Whether a PID may identify an incumbent from this candidate's point of view.
///
/// Production rejects the current PID so a stale rendezvous left by a prior
/// process whose PID was reused cannot protect an unrelated socket. The
/// in-process daemon harness opts in to the same-PID case because all of its
/// otherwise independent boot candidates necessarily share one OS process.
#[cfg(unix)]
fn pid_can_name_incumbent(pid: u32, current_pid: u32, allow_same_process_incumbent: bool) -> bool {
    allow_same_process_incumbent || pid != current_pid
}

#[cfg(unix)]
async fn cleanup_stale_daemon(
    sock: &std::path::Path,
    pid_file: &std::path::Path,
    allow_same_process_incumbent: bool,
) -> bool {
    if let Ok(pid_str) = std::fs::read_to_string(pid_file) {
        if let Ok(pid) = pid_str.trim().parse::<u32>() {
            if pid_can_name_incumbent(pid, std::process::id(), allow_same_process_incumbent)
                && is_process_running(pid)
                && sock.exists()
                && UnixStream::connect(sock).await.is_ok()
            {
                return false;
            }
        }
    }
    if sock.exists() {
        if let Err(e) = std::fs::remove_file(sock) {
            tracing::warn!(error = %e, path = ?sock, "failed to remove stale socket");
        }
    }
    if pid_file.exists() {
        if let Err(e) = std::fs::remove_file(pid_file) {
            tracing::warn!(error = %e, path = ?pid_file, "failed to remove stale PID file");
        }
    }
    true
}

/// Create `pid_file` exclusively (`O_EXCL`) and write this process's PID.
///
/// Uses `create_new(true)` rather than `create(true).truncate(true)` so
/// this can never silently overwrite a PID file another process created —
/// under normal operation the boot lock already serializes cleanup → bind →
/// pid-write across processes, but that guarantee depends on
/// `acquire_recovery_lock` succeeding for every party. Exclusive creation is
/// the defense that holds even if the lock itself is unavailable to one side:
/// the loser observes `ErrorKind::AlreadyExists` instead of clobbering the
/// winner's PID out from under it.
#[cfg(unix)]
fn write_pid_file_exclusive(pid_file: &std::path::Path) -> std::io::Result<()> {
    use std::os::unix::fs::OpenOptionsExt;

    let mut opts = std::fs::OpenOptions::new();
    opts.write(true).create_new(true).mode(0o600);
    let mut f = opts.open(pid_file)?;
    f.write_all(std::process::id().to_string().as_bytes())?;
    Ok(())
}

/// Return `true` if `pid_file` currently names an eligible live process that
/// still answers on `sock` — i.e. a daemon already owns this rendezvous and it
/// is safe to defer to it rather than treat the `AlreadyExists` PID-file
/// collision as a boot failure. Eligibility requires a different PID in
/// production; the explicit in-process harness may allow the current PID.
#[cfg(unix)]
async fn pid_file_names_a_reachable_daemon(
    pid_file: &std::path::Path,
    sock: &std::path::Path,
    allow_same_process_incumbent: bool,
) -> bool {
    let Ok(pid_str) = std::fs::read_to_string(pid_file) else {
        return false;
    };
    let Ok(pid) = pid_str.trim().parse::<u32>() else {
        return false;
    };
    pid_can_name_incumbent(pid, std::process::id(), allow_same_process_incumbent)
        && is_process_running(pid)
        && sock.exists()
        && UnixStream::connect(sock).await.is_ok()
}

#[cfg(unix)]
async fn drain(active: &std::sync::atomic::AtomicUsize) -> bool {
    drain_with_timeout(active, drain_timeout()).await
}

#[cfg(unix)]
async fn drain_with_timeout(
    active: &std::sync::atomic::AtomicUsize,
    timeout: std::time::Duration,
) -> bool {
    use std::sync::atomic::Ordering;
    // One sequentially-consistent order spans connection handoff and tracked
    // task publication. Drain must not observe an ended connection and a
    // stale pre-publication background count as two simultaneous zeroes.
    let remaining = || active.load(Ordering::SeqCst) + background_task_count();
    if remaining() == 0 {
        return true;
    }
    let deadline = tokio::time::Instant::now() + timeout;
    while remaining() > 0 {
        if tokio::time::Instant::now() >= deadline {
            tracing::warn!(
                remaining_connections = active.load(Ordering::SeqCst),
                remaining_background_tasks = background_task_count(),
                "drain timeout reached; forcing shutdown"
            );
            return false;
        }
        tokio::select! {
            _ = tokio::time::sleep(std::time::Duration::from_millis(100)) => {}
            _ = tokio::time::sleep_until(deadline) => {}
        }
    }
    true
}

#[cfg(unix)]
async fn finish_connection_tasks(tasks: Vec<tokio::task::JoinHandle<()>>, drained: bool) {
    if !drained {
        for task in &tasks {
            if !task.is_finished() {
                task.abort();
            }
        }
    }
    for task in tasks {
        let _ = task.await;
    }
}

/// The bound `drain()` waits for tracked background tasks at daemon shutdown
/// (`KHIVE_DRAIN_TIMEOUT_SECS`, default 10s). Public so component supervision
/// can clamp per-component shutdown timeouts against it — a component timeout
/// longer than the drain bound could never complete its abort/state
/// transition before the daemon returns.
pub fn drain_timeout() -> std::time::Duration {
    let secs = std::env::var("KHIVE_DRAIN_TIMEOUT_SECS")
        .ok()
        .and_then(|v| v.parse::<u64>().ok())
        .unwrap_or(DEFAULT_DRAIN_TIMEOUT_SECS);
    std::time::Duration::from_secs(secs)
}

/// Returns `true` for non-empty env values that are not `"0"` or `"false"`.
#[cfg(unix)]
pub fn env_truthy(key: &str) -> bool {
    std::env::var(key)
        .map(|v| {
            let v = v.trim();
            !v.is_empty() && v != "0" && !v.eq_ignore_ascii_case("false")
        })
        .unwrap_or(false)
}

#[cfg(test)]
mod khive_root_tests {
    use super::khive_root_from;
    use std::path::PathBuf;

    #[test]
    fn home_wins_when_set() {
        assert_eq!(
            khive_root_from(Some("/base/home".into()), Some("/other".into())),
            PathBuf::from("/base/home/.khive")
        );
    }

    #[test]
    fn userprofile_backfills_missing_home() {
        assert_eq!(
            khive_root_from(None, Some("/profile/home".into())),
            PathBuf::from("/profile/home/.khive")
        );
    }

    #[test]
    fn blank_values_are_skipped() {
        assert_eq!(
            khive_root_from(Some("  ".into()), Some("/profile/home".into())),
            PathBuf::from("/profile/home/.khive")
        );
    }

    #[cfg(not(unix))]
    #[test]
    fn fallback_is_never_cwd_relative() {
        // On non-unix the lock anchor must not depend on the caller's working
        // directory even when no home variable is set — a relative path here
        // would give the same database different lock files per cwd.
        assert!(khive_root_from(None, None).is_absolute());
    }

    #[cfg(unix)]
    #[test]
    fn unix_fallback_stays_historical_not_shared_tmp() {
        // On unix the no-HOME last resort deliberately stays "./.khive"
        // rather than a shared world-writable anchor like /tmp, which a
        // local attacker could pre-claim to intercept the daemon socket.
        assert_eq!(khive_root_from(None, None), PathBuf::from("./.khive"));
    }
}

#[cfg(all(test, unix))]
mod tests {
    use super::*;
    use serial_test::serial;

    #[derive(Debug)]
    struct DrainBlockingBlobStore {
        started: std::sync::Mutex<Option<tokio::sync::oneshot::Sender<()>>>,
        release: Arc<tokio::sync::Semaphore>,
    }

    #[async_trait]
    impl khive_storage::BlobStore for DrainBlockingBlobStore {
        async fn put(
            &self,
            _bytes: Vec<u8>,
        ) -> khive_storage::StorageResult<khive_storage::ContentRef> {
            panic!("put is not used by the hydration drain test")
        }

        async fn get_bounded_verified(
            &self,
            _content_ref: &khive_storage::ContentRef,
            _max_bytes: u64,
        ) -> khive_storage::StorageResult<Vec<u8>> {
            if let Some(started) = self
                .started
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .take()
            {
                let _ = started.send(());
            }
            self.release
                .clone()
                .acquire_owned()
                .await
                .expect("test release semaphore remains open")
                .forget();
            Ok(b"late result".to_vec())
        }

        async fn exists(
            &self,
            _content_ref: &khive_storage::ContentRef,
        ) -> khive_storage::StorageResult<bool> {
            panic!("exists is not used by the hydration drain test")
        }

        async fn size(
            &self,
            _content_ref: &khive_storage::ContentRef,
        ) -> khive_storage::StorageResult<Option<u64>> {
            panic!("size is not used by the hydration drain test")
        }

        async fn delete(
            &self,
            _content_ref: &khive_storage::ContentRef,
        ) -> khive_storage::StorageResult<bool> {
            panic!("delete is not used by the hydration drain test")
        }
    }

    struct AppendCompletionEventStore {
        inner: Arc<dyn khive_storage::EventStore>,
        first_append: std::sync::Mutex<Option<tokio::sync::oneshot::Sender<()>>>,
    }

    #[async_trait]
    impl khive_storage::EventStore for AppendCompletionEventStore {
        async fn append_event(
            &self,
            event: khive_storage::Event,
        ) -> khive_storage::StorageResult<()> {
            self.inner.append_event(event).await?;
            if let Some(completed) = self
                .first_append
                .lock()
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .take()
            {
                let _ = completed.send(());
            }
            Ok(())
        }

        async fn append_events(
            &self,
            events: Vec<khive_storage::Event>,
        ) -> khive_storage::StorageResult<khive_storage::BatchWriteSummary> {
            self.inner.append_events(events).await
        }

        async fn get_event(
            &self,
            id: uuid::Uuid,
        ) -> khive_storage::StorageResult<Option<khive_storage::Event>> {
            self.inner.get_event(id).await
        }

        async fn query_events(
            &self,
            filter: khive_storage::EventFilter,
            page: khive_storage::PageRequest,
        ) -> khive_storage::StorageResult<khive_storage::Page<khive_storage::Event>> {
            self.inner.query_events(filter, page).await
        }

        async fn count_events(
            &self,
            filter: khive_storage::EventFilter,
        ) -> khive_storage::StorageResult<u64> {
            self.inner.count_events(filter).await
        }
    }

    #[tokio::test]
    #[serial(checkpoint_skip_metrics)]
    async fn secondary_only_checkpoint_topology_emits_lifecycle_outcome() {
        let main_backend = khive_db::StorageBackend::memory().expect("in-memory main backend");
        let inner_event_store = main_backend.events().expect("main event store");
        let (first_append_tx, first_append_rx) = tokio::sync::oneshot::channel();
        let event_store: Arc<dyn khive_storage::EventStore> =
            Arc::new(AppendCompletionEventStore {
                inner: inner_event_store,
                first_append: std::sync::Mutex::new(Some(first_append_tx)),
            });
        let secondary_dir = tempfile::tempdir().expect("secondary tempdir");
        let secondary_backend =
            khive_db::StorageBackend::sqlite(secondary_dir.path().join("secondary.db"))
                .expect("file-backed secondary backend");

        let mut tasks = checkpoint_task_specs(
            None,
            vec![secondary_backend.pool_arc()],
            Some(Arc::clone(&event_store)),
            "local".to_string(),
        );
        assert_eq!(tasks.len(), 1);
        let task = tasks.pop().expect("one secondary checkpoint task");
        assert!(!task.is_main, "the only checkpoint task must be secondary");
        assert!(
            task.lifecycle_owner.is_some(),
            "the secondary task must own lifecycle emission when no main task exists"
        );

        let config = CheckpointConfig {
            interval: std::time::Duration::from_millis(10),
            warn_pages: 0,
            ..CheckpointConfig::default()
        };
        let (shutdown_tx, shutdown_rx) = tokio::sync::watch::channel(());
        let handle = tokio::spawn(run_checkpoint_task(
            task.pool,
            config,
            task.lifecycle_owner,
            shutdown_rx,
            task.is_main,
        ));

        // The scheduler intentionally aborts its bounded append worker during
        // shutdown. Wait on the append's completion edge before requesting
        // shutdown, so the observation cannot race that deliberate abort.
        tokio::time::timeout(std::time::Duration::from_secs(10), first_append_rx)
            .await
            .expect("secondary checkpoint owner did not complete an append within 10s")
            .expect("checkpoint lifecycle append completion sender dropped");

        let events = event_store
            .query_events(
                khive_storage::EventFilter::default(),
                khive_storage::PageRequest {
                    limit: 100,
                    offset: 0,
                },
            )
            .await
            .expect("query lifecycle events");

        shutdown_tx.send(()).expect("send checkpoint shutdown");
        tokio::time::timeout(std::time::Duration::from_secs(1), handle)
            .await
            .expect("checkpoint task should exit within 1s")
            .expect("checkpoint task panicked");
        assert!(
            !events.items.is_empty()
                && events
                    .items
                    .iter()
                    .all(|event| event.kind == khive_types::EventKind::CheckpointOutcomeRecorded),
            "the designated secondary owner must emit CheckpointOutcomeRecorded"
        );

        let file_main_dir = tempfile::tempdir().expect("file-backed main tempdir");
        let file_main = khive_db::StorageBackend::sqlite(file_main_dir.path().join("main.db"))
            .expect("file-backed main backend");
        let tasks = checkpoint_task_specs(
            Some(file_main.pool_arc()),
            vec![secondary_backend.pool_arc()],
            Some(event_store),
            "local".to_string(),
        );
        assert!(tasks[0].is_main && tasks[0].lifecycle_owner.is_some());
        assert!(!tasks[1].is_main && tasks[1].lifecycle_owner.is_none());
    }

    // Focused regression tests for the unsafe process probe (SAFETY: signal 0
    // is an existence check with no side effects; see is_process_running).

    #[test]
    fn current_process_is_running() {
        // The current PID is always alive.
        let pid = std::process::id();
        assert!(
            is_process_running(pid),
            "current process {pid} should be detected as running"
        );
    }

    #[test]
    fn pid_zero_is_not_running() {
        // PID 0 is the process group; kill(0, 0) sends to the group,
        // which we treat as invalid — the guard `pid <= 0` must block it.
        assert!(
            !is_process_running(0),
            "pid 0 must be rejected by the guard before the unsafe call"
        );
    }

    #[test]
    fn very_large_pid_is_not_running() {
        // u32::MAX overflows i32 — try_from returns Err, guard returns false.
        assert!(
            !is_process_running(u32::MAX),
            "u32::MAX should fail i32 conversion and return false"
        );
    }

    // EPERM (process exists, no permission to signal it) must not be
    // misread as "not running" during stale-daemon cleanup.

    #[test]
    fn classify_kill_result_zero_is_alive() {
        assert_eq!(classify_kill_result(0, 0), PidLiveness::Alive);
        assert!(classify_kill_result(0, 0).is_running());
    }

    #[test]
    fn classify_kill_result_esrch_is_dead() {
        assert_eq!(classify_kill_result(-1, libc::ESRCH), PidLiveness::Dead);
        assert!(!classify_kill_result(-1, libc::ESRCH).is_running());
    }

    #[test]
    fn classify_kill_result_eperm_is_permission_denied_and_counts_as_running() {
        assert_eq!(
            classify_kill_result(-1, libc::EPERM),
            PidLiveness::PermissionDenied
        );
        assert!(
            classify_kill_result(-1, libc::EPERM).is_running(),
            "EPERM must be unknown-safe: treated as running, never as a basis \
             for stale cleanup to unlink a live daemon's rendezvous files"
        );
    }

    #[test]
    fn same_process_pid_requires_explicit_in_process_harness_opt_in() {
        let current = std::process::id();
        assert!(
            !pid_can_name_incumbent(current, current, false),
            "production startup must not trust a same-PID stale rendezvous"
        );
        assert!(
            pid_can_name_incumbent(current, current, true),
            "the in-process harness must let a responsive same-PID owner win"
        );
        // Keep the probe two away from `current` so the fixture preserves the
        // off-by-one regression check for adjacent PIDs; wrapping_add avoids
        // making the test overflow-sensitive at the u32 boundary.
        let distinct_probe_pid = current.wrapping_add(2);
        assert_ne!(
            distinct_probe_pid, current,
            "probe PID must differ from this process's PID"
        );
        assert!(
            pid_can_name_incumbent(distinct_probe_pid, current, false),
            "a distinct PID remains eligible under ordinary production rules"
        );
    }

    #[test]
    fn pid_1_probe_is_running_regardless_of_permission_outcome() {
        // PID 1 (init/launchd) always exists. An unprivileged process gets
        // EPERM signaling it (never ESRCH); running as root would get 0
        // instead. Either way `is_process_running` must report true — this
        // is the live regression guard for EPERM being misread as dead;
        // `classify_kill_result` above is the pure-function unit coverage
        // for the same mapping, kept independent of process ownership so
        // it is never flaky in CI.
        assert!(
            is_process_running(1),
            "PID 1 always exists; EPERM must not read as dead"
        );
    }

    #[test]
    fn env_truthy_recognises_set_values() {
        assert!(!env_truthy("__KHIVE_TEST_ABSENT_VAR_XYZ__"));

        // env_truthy with a live value — set and unset atomically to avoid
        // cross-test pollution (not parallel-safe without serial_test, but these
        // are fast unit tests and the variable name is unique).
        let key = "__KHIVE_TEST_TRUTHY_ABC__";
        std::env::set_var(key, "1");
        assert!(env_truthy(key));
        std::env::set_var(key, "false");
        assert!(!env_truthy(key));
        std::env::set_var(key, "0");
        assert!(!env_truthy(key));
        std::env::remove_var(key);
    }

    #[tokio::test(flavor = "current_thread", start_paused = true)]
    #[serial(background_tasks)]
    async fn accepted_connection_is_counted_before_first_poll_and_drain_waits() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let started = Arc::new(std::sync::atomic::AtomicBool::new(false));
        let (release_tx, release_rx) = tokio::sync::oneshot::channel::<()>();
        let started_in_task = Arc::clone(&started);

        let handle = spawn_connection_task(Arc::clone(&active), async move {
            started_in_task.store(true, Ordering::Relaxed);
            let _ = release_rx.await;
        });

        assert_eq!(active.load(Ordering::Relaxed), 1);
        assert!(
            !started.load(Ordering::Relaxed),
            "the current-thread runtime must leave the spawned handler unpolled"
        );

        let drain_fut = drain(active.as_ref());
        tokio::pin!(drain_fut);
        let too_early =
            tokio::time::timeout(std::time::Duration::from_millis(150), &mut drain_fut).await;
        assert!(
            too_early.is_err(),
            "drain must wait for a connection claimed before its task's first poll"
        );
        assert!(started.load(Ordering::Relaxed));

        release_tx.send(()).expect("handler still waiting");
        tokio::time::timeout(std::time::Duration::from_secs(1), handle)
            .await
            .expect("handler should finish promptly")
            .expect("handler should not panic");
        assert_eq!(active.load(Ordering::Relaxed), 0);
        tokio::time::timeout(std::time::Duration::from_secs(1), drain_fut)
            .await
            .expect("drain should finish once the handler releases its claim");
    }

    #[tokio::test(flavor = "current_thread")]
    async fn cancelled_connection_releases_count_before_first_poll() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let started = Arc::new(std::sync::atomic::AtomicBool::new(false));
        let started_in_task = Arc::clone(&started);
        let handle = spawn_connection_task(Arc::clone(&active), async move {
            started_in_task.store(true, Ordering::Relaxed);
            std::future::pending::<()>().await;
        });

        assert_eq!(active.load(Ordering::Relaxed), 1);
        handle.abort();
        let error = handle.await.expect_err("aborted handler must be cancelled");
        assert!(error.is_cancelled());
        assert!(!started.load(Ordering::Relaxed));
        assert_eq!(active.load(Ordering::Relaxed), 0);
    }

    #[tokio::test(flavor = "current_thread")]
    async fn panicked_connection_releases_count() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let handle = spawn_connection_task(Arc::clone(&active), async move {
            panic!("intentional connection-handler panic");
        });

        assert_eq!(active.load(Ordering::Relaxed), 1);
        let error = handle
            .await
            .expect_err("panicked handler must fail its join");
        assert!(error.is_panic());
        assert_eq!(active.load(Ordering::Relaxed), 0);
    }

    #[test]
    fn connection_claim_releases_if_spawn_panics() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            drop(spawn_connection_task(Arc::clone(&active), async {}));
        }));

        assert!(result.is_err(), "tokio::spawn outside a runtime must panic");
        assert_eq!(active.load(Ordering::Relaxed), 0);
    }

    #[tokio::test]
    #[serial(background_tasks)]
    async fn drain_returns_promptly_with_no_accepted_connection() {
        let active = std::sync::atomic::AtomicUsize::new(0);
        tokio::time::timeout(std::time::Duration::from_secs(1), drain(&active))
            .await
            .expect("empty drain should return immediately");
    }

    #[tokio::test(start_paused = true)]
    #[serial(background_tasks)]
    async fn graceful_drain_has_a_hard_upper_bound_with_stuck_work() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let task = spawn_connection_task(Arc::clone(&active), async {
            std::future::pending::<()>().await;
        });
        assert_eq!(active.load(Ordering::Relaxed), 1);
        let started = tokio::time::Instant::now();

        let drained = drain_with_timeout(&active, std::time::Duration::from_millis(250)).await;

        assert!(!drained, "stuck work must exhaust the drain bound");
        assert!(
            started.elapsed() >= std::time::Duration::from_millis(250)
                && started.elapsed() < std::time::Duration::from_millis(350),
            "graceful shutdown exceeded its configured bound: {:?}",
            started.elapsed()
        );
        finish_connection_tasks(vec![task], drained).await;
        assert_eq!(
            active.load(Ordering::Relaxed),
            0,
            "hard-bound escalation must abort, await, and release the handler"
        );
    }

    #[tokio::test(start_paused = true)]
    #[serial(background_tasks)]
    async fn admitted_work_finishes_inside_drain_window_without_abort() {
        use std::sync::atomic::Ordering;

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let committed = Arc::new(std::sync::atomic::AtomicBool::new(false));
        let committed_in_task = Arc::clone(&committed);
        let task = spawn_connection_task(Arc::clone(&active), async move {
            tokio::time::sleep(std::time::Duration::from_millis(100)).await;
            committed_in_task.store(true, Ordering::SeqCst);
        });

        let drained = drain_with_timeout(&active, std::time::Duration::from_millis(250)).await;
        assert!(
            drained,
            "admitted work should finish inside the drain window"
        );
        finish_connection_tasks(vec![task], drained).await;
        assert!(committed.load(Ordering::SeqCst));
        assert_eq!(active.load(Ordering::Relaxed), 0);
    }

    // `drain()` must wait for tracked background tasks (e.g. memory.recall's
    // serve-ledger append), not just in-flight connections, or a SIGTERM
    // lands mid-flight with no log and no row.
    //
    // `#[serial(background_tasks)]`: this test reads/asserts on the
    // process-wide `BACKGROUND_TASKS` static shared with the two counter
    // tests below. Under default parallel execution one test's increment
    // leaks into another's snapshot-then-assert window (reproduced: both
    // counter tests failed together, passed with `--test-threads=1`).
    // Serializing just this named group isolates them from each other
    // without forcing the whole test binary (including unrelated
    // `#[serial]` tests elsewhere in this crate) onto one thread.
    #[tokio::test]
    #[serial(background_tasks)]
    async fn drain_waits_for_tracked_background_tasks_before_returning() {
        let active = std::sync::atomic::AtomicUsize::new(0);
        let (tx, rx) = tokio::sync::oneshot::channel::<()>();

        track_background_task(async move {
            let _ = rx.await;
        });
        assert!(
            background_task_count() >= 1,
            "track_background_task must make the in-flight task visible immediately"
        );

        let drain_fut = drain(&active);
        tokio::pin!(drain_fut);

        // Must NOT resolve while the tracked task is still pending.
        let too_early =
            tokio::time::timeout(std::time::Duration::from_millis(150), &mut drain_fut).await;
        assert!(
            too_early.is_err(),
            "drain() must not return while a tracked background task is still running"
        );

        // Completing the task must let drain() proceed promptly.
        tx.send(())
            .expect("tracked task still awaiting the oneshot");
        let done = tokio::time::timeout(std::time::Duration::from_secs(5), drain_fut).await;
        assert!(
            done.is_ok(),
            "drain() must return once the tracked background task finishes"
        );
    }

    #[tokio::test]
    #[serial(background_tasks)]
    async fn drain_waits_for_hydration_after_its_last_request_waiter_is_cancelled() {
        let before = background_task_count();
        let active = std::sync::atomic::AtomicUsize::new(0);
        let (started_tx, started_rx) = tokio::sync::oneshot::channel();
        let release = Arc::new(tokio::sync::Semaphore::new(0));
        let store = Arc::new(DrainBlockingBlobStore {
            started: std::sync::Mutex::new(Some(started_tx)),
            release: Arc::clone(&release),
        });
        let hydrator = Arc::new(
            crate::BlobHydrator::new(
                store as Arc<dyn khive_storage::BlobStore>,
                khive_storage::MAX_BLOB_WHOLE_BYTES,
            )
            .expect("minimum hydration budget is valid"),
        );
        let content_ref =
            khive_storage::ContentRef::from_hex("a".repeat(64)).expect("fixture content ref");

        let request_hydrator = Arc::clone(&hydrator);
        let request = tokio::spawn(async move {
            request_hydrator
                .hydrate_verified(&content_ref, khive_storage::MAX_BLOB_WHOLE_BYTES)
                .await
        });
        started_rx.await.expect("backend work must begin");
        request.abort();
        assert!(request.await.unwrap_err().is_cancelled());
        assert_eq!(background_task_count(), before + 1);

        let draining = drain_with_timeout(&active, std::time::Duration::from_secs(5));
        tokio::pin!(draining);
        assert!(
            tokio::time::timeout(std::time::Duration::from_millis(150), &mut draining)
                .await
                .is_err(),
            "drain must remain pending while cancelled-request hydration still runs"
        );

        release.add_permits(1);
        assert!(
            tokio::time::timeout(std::time::Duration::from_secs(5), draining)
                .await
                .expect("drain should finish after native hydration ends"),
            "hydration should finish inside the drain window"
        );
        assert_eq!(background_task_count(), before);
    }

    // See the `#[serial(background_tasks)]` note on
    // `drain_waits_for_tracked_background_tasks_before_returning` above —
    // this test shares the same process-wide `BACKGROUND_TASKS` static and
    // races it (and the panic test below) under default parallelism.
    #[tokio::test]
    #[serial(background_tasks)]
    async fn track_background_task_count_returns_to_zero_after_completion() {
        // Sanity check on the counter's own bookkeeping, independent of drain().
        let before = background_task_count();
        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
        track_background_task(async move {
            let _ = rx.await;
        });
        assert_eq!(background_task_count(), before + 1);
        tx.send(()).expect("still awaiting");
        // Yield until the spawned task's decrement has actually run.
        for _ in 0..100 {
            if background_task_count() == before {
                break;
            }
            tokio::time::sleep(std::time::Duration::from_millis(10)).await;
        }
        assert_eq!(background_task_count(), before);
    }

    // See the `#[serial(background_tasks)]` note above — shares
    // `BACKGROUND_TASKS` with the other two tests in this group.
    #[tokio::test]
    #[serial(background_tasks)]
    async fn track_background_task_count_returns_to_baseline_after_panic() {
        // A panic inside the tracked future must still decrement the
        // counter (via BackgroundTaskGuard's Drop), not leak it forever.
        // `track_background_task` discards the spawned
        // task's `JoinHandle` (it is fire-and-forget by design — the caller
        // never awaits it), so this test does not await the panic directly;
        // tokio isolates the panic to the spawned task instead of aborting
        // the process, and we observe the recovery purely through the
        // shared counter returning to baseline after the guard's `Drop`
        // fires during that task's unwind.
        let before = background_task_count();

        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
        track_background_task(async move {
            let _ = rx.await;
            panic!("intentional panic to exercise the Drop-guard decrement path");
        });
        assert_eq!(background_task_count(), before + 1);

        tx.send(()).expect("still awaiting");
        for _ in 0..100 {
            if background_task_count() == before {
                break;
            }
            tokio::time::sleep(std::time::Duration::from_millis(10)).await;
        }
        assert_eq!(
            background_task_count(),
            before,
            "background task counter must return to baseline after the tracked future panics"
        );
    }

    // ── active background phase names (ADR-103) ──────────────────────────

    // `#[serial(active_phases)]`: these tests read/assert on the process-wide
    // `ACTIVE_PHASES` static. No other test in this crate touches it today,
    // but the group mirrors the `background_tasks` precedent above so a
    // future addition does not silently reintroduce the same interleaving
    // hazard that motivated it there.
    #[test]
    #[serial(active_phases)]
    fn register_active_phase_appears_and_disappears_with_the_guard() {
        assert!(
            !active_phase_names().contains(&"adr103_test_phase".to_string()),
            "must start absent (leaked from a prior failed run would poison this test)"
        );

        let guard = register_active_phase("adr103_test_phase");
        assert!(active_phase_names().contains(&"adr103_test_phase".to_string()));

        drop(guard);
        assert!(
            !active_phase_names().contains(&"adr103_test_phase".to_string()),
            "the phase name must drop out of the gauge once its guard is dropped"
        );
    }

    #[test]
    #[serial(active_phases)]
    fn register_active_phase_counts_concurrent_occurrences_of_the_same_name() {
        let first = register_active_phase("adr103_concurrent_phase");
        let second = register_active_phase("adr103_concurrent_phase");
        assert!(active_phase_names().contains(&"adr103_concurrent_phase".to_string()));

        drop(first);
        assert!(
            active_phase_names().contains(&"adr103_concurrent_phase".to_string()),
            "one of two concurrent occurrences ending must not remove the name early"
        );

        drop(second);
        assert!(
            !active_phase_names().contains(&"adr103_concurrent_phase".to_string()),
            "the name must be removed only once every concurrent occurrence has ended"
        );
    }

    // ── metrics-only frame (load/perf harness read-surface) ────────────────

    /// Minimal `DaemonDispatch` for the metrics tests: `dispatch` just counts
    /// how many times it was called (so tests can assert the ops path was
    /// never reached) and `pool_for_checkpoint` returns whatever pool the
    /// test wired in (or `None`, matching an in-memory/poolless dispatcher).
    #[derive(Clone)]
    struct MockDispatch {
        namespace: String,
        config_id: String,
        dispatch_calls: Arc<std::sync::atomic::AtomicUsize>,
        pool: Option<Arc<ConnectionPool>>,
        /// When `Some(msg)`, `dispatch` returns `Err(msg)` instead of the
        /// default `Ok("{}")` — lets a test drive `handle_conn`'s real
        /// dispatch-error arm (khive#948 request_id echo coverage).
        dispatch_err: Option<String>,
    }

    #[derive(Clone)]
    struct CancellationAwareDispatch {
        started: Arc<tokio::sync::Notify>,
        cancellation_observed: Arc<std::sync::atomic::AtomicBool>,
    }

    #[async_trait]
    impl DaemonDispatch for CancellationAwareDispatch {
        async fn dispatch(
            &self,
            _ops: String,
            _presentation: Option<String>,
            _presentation_per_op: Option<Vec<Option<String>>>,
            _format: Option<String>,
            _format_per_op: Option<Vec<Option<String>>>,
            _from_wire: bool,
            _identity: Option<RequestIdentity>,
        ) -> Result<String, String> {
            self.started.notify_one();
            khive_storage::wait_for_request_read_cancellation().await;
            self.cancellation_observed
                .store(true, std::sync::atomic::Ordering::SeqCst);
            Ok("{}".to_string())
        }

        async fn warm_all(&self) {}

        fn namespace(&self) -> &str {
            "local"
        }

        fn config_id(&self) -> &str {
            "disconnect-test"
        }
    }

    #[async_trait]
    impl DaemonDispatch for MockDispatch {
        async fn dispatch(
            &self,
            _ops: String,
            _presentation: Option<String>,
            _presentation_per_op: Option<Vec<Option<String>>>,
            _format: Option<String>,
            _format_per_op: Option<Vec<Option<String>>>,
            _from_wire: bool,
            _identity: Option<RequestIdentity>,
        ) -> Result<String, String> {
            self.dispatch_calls
                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
            match &self.dispatch_err {
                Some(msg) => Err(msg.clone()),
                None => Ok("{}".to_string()),
            }
        }

        async fn warm_all(&self) {}

        fn namespace(&self) -> &str {
            &self.namespace
        }

        fn config_id(&self) -> &str {
            &self.config_id
        }

        fn pool_for_checkpoint(&self) -> Option<Arc<ConnectionPool>> {
            self.pool.clone()
        }
    }

    fn base_request_frame(config_id: &str) -> DaemonRequestFrame {
        DaemonRequestFrame {
            ops: String::new(),
            presentation: None,
            presentation_per_op: None,
            namespace: "local".to_string(),
            actor_id: None,
            process_ref: None,
            visible_namespaces: Vec::new(),
            config_id: config_id.to_string(),
            protocol_version: PROTOCOL_VERSION,
            probe_only: false,
            metrics_only: false,
            format: None,
            format_per_op: None,
            from_wire: false,
            request_id: None,
        }
    }

    /// Drive `handle_conn` over an in-process `UnixStream::pair()` (no real
    /// socket file needed) and decode the response frame it writes back.
    async fn round_trip(dispatcher: MockDispatch, req: &DaemonRequestFrame) -> DaemonResponseFrame {
        let (mut client, server) = UnixStream::pair().expect("unix stream pair");
        let payload = serde_json::to_vec(req).expect("encode request frame");
        let handle = tokio::spawn(async move {
            handle_conn(server, dispatcher).await;
        });
        write_frame(&mut client, &payload)
            .await
            .expect("write request frame");
        let raw = read_frame(&mut client).await.expect("read response frame");
        handle.await.expect("handle_conn task panicked");
        serde_json::from_slice(&raw).expect("decode response frame")
    }

    #[tokio::test]
    async fn daemon_peer_disconnect_signals_request_read_cancellation() {
        let started = Arc::new(tokio::sync::Notify::new());
        let cancellation_observed = Arc::new(std::sync::atomic::AtomicBool::new(false));
        let dispatcher = CancellationAwareDispatch {
            started: Arc::clone(&started),
            cancellation_observed: Arc::clone(&cancellation_observed),
        };
        let (mut client, server) = UnixStream::pair().expect("unix stream pair");
        let request = base_request_frame("disconnect-test");
        let payload = serde_json::to_vec(&request).expect("encode request frame");
        let handler = tokio::spawn(async move { handle_conn(server, dispatcher).await });
        write_frame(&mut client, &payload)
            .await
            .expect("write request frame");
        started.notified().await;

        drop(client);
        tokio::time::timeout(std::time::Duration::from_millis(500), handler)
            .await
            .expect("daemon handler ignored peer disconnect")
            .expect("daemon handler panicked");
        assert!(
            cancellation_observed.load(std::sync::atomic::Ordering::SeqCst),
            "peer loss did not reach the request-scoped read cancellation signal"
        );
    }

    /// Protocol v4 makes `process_ref` part of dispatch semantics. A still-warm
    /// v3 daemon/client pairing must fail before the verb runs; otherwise the
    /// older peer can ignore the unknown field, persist a message without the
    /// requested provenance, and leave the caller unable to retry safely.
    #[tokio::test]
    async fn protocol_v3_frame_is_rejected_before_process_ref_dispatch() {
        assert_eq!(PROTOCOL_VERSION, 4, "process_ref is the protocol-v4 change");
        let dispatch_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-v4".to_string(),
            dispatch_calls: Arc::clone(&dispatch_calls),
            pool: None,
            dispatch_err: None,
        };
        let mut request = base_request_frame("cfg-v4");
        request.protocol_version = 3;
        request.process_ref = Some("worker/legacy-rollout".to_string());

        let response = round_trip(dispatcher, &request).await;
        assert!(!response.ok);
        assert!(response.version_mismatch);
        assert_eq!(response.daemon_protocol_version, 4);
        assert_eq!(
            dispatch_calls.load(std::sync::atomic::Ordering::SeqCst),
            0,
            "a v3 frame must be rejected before a provenance-bearing mutation dispatches"
        );
        let error = response.error.expect("mismatch explains both versions");
        assert!(
            error.contains("client=3") && error.contains("daemon=4"),
            "mismatch must identify the exact rollout boundary; got {error:?}"
        );
    }

    /// Test 1: a `metrics_only: true` request
    /// returns `metrics: Some(_)` and never reaches the ops-dispatch path; a
    /// normal request (the default `metrics_only: false`) still dispatches
    /// exactly as before and carries no metrics. Also proves `metrics_only`
    /// bypasses the `config_id` equality reject (a gauge read is
    /// process-global, not namespaced to a particular client config).
    #[tokio::test]
    async fn metrics_only_frame_returns_snapshot_without_dispatching() {
        let dispatch_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-a".to_string(),
            dispatch_calls: Arc::clone(&dispatch_calls),
            pool: None,
            dispatch_err: None,
        };

        let mut metrics_req = base_request_frame("cfg-a");
        metrics_req.metrics_only = true;
        let metrics_resp = round_trip(dispatcher.clone(), &metrics_req).await;

        assert!(metrics_resp.ok, "metrics_only response must be ok=true");
        assert!(
            metrics_resp.metrics.is_some(),
            "metrics_only=true must return Some(snapshot)"
        );
        assert_eq!(
            dispatch_calls.load(std::sync::atomic::Ordering::SeqCst),
            0,
            "metrics_only must never reach the ops-dispatch path"
        );

        // metrics_only bypasses the config_id equality reject.
        let mut mismatched_req = base_request_frame("some-other-config");
        mismatched_req.metrics_only = true;
        let mismatched_resp = round_trip(dispatcher.clone(), &mismatched_req).await;
        assert!(mismatched_resp.ok);
        assert!(mismatched_resp.metrics.is_some());
        assert!(!mismatched_resp.config_mismatch);
        assert_eq!(
            dispatch_calls.load(std::sync::atomic::Ordering::SeqCst),
            0,
            "a mismatched-config metrics_only request must still skip dispatch"
        );

        // A normal request (default metrics_only=false) is unaffected: it
        // still dispatches and carries no metrics.
        let normal_req = base_request_frame("cfg-a");
        let normal_resp = round_trip(dispatcher, &normal_req).await;
        assert!(normal_resp.ok);
        assert!(normal_resp.metrics.is_none());
        assert_eq!(dispatch_calls.load(std::sync::atomic::Ordering::SeqCst), 1);
    }

    /// Group/other-writable socket directories are refused whether or not the
    /// sticky bit is set, and — the part that matters — are left exactly as
    /// they were found.
    ///
    /// The sticky `/tmp` shape (1777) is in the refusal set deliberately: the
    /// sticky bit restricts unlinking, not creating, so a shared directory
    /// lets another user pre-bind the predictable socket path before this
    /// daemon starts; and a sticky directory's owner may unlink and rebind
    /// regardless. Re-permissioning someone else's directory on the way past
    /// is what the second assertion pins against: run as a user who *can*
    /// chmod it, the old unconditional call succeeded and revoked access for
    /// every other process using that directory.
    #[test]
    fn shared_writable_socket_dirs_are_refused_without_being_modified() {
        let dir = tempfile::tempdir().expect("tempdir");
        for (name, mode) in [("open", 0o777u32), ("sticky-tmp", 0o1777u32)] {
            let shared = dir.path().join(name);
            std::fs::create_dir(&shared).expect("create");
            std::fs::set_permissions(&shared, std::fs::Permissions::from_mode(mode))
                .expect("chmod");

            let err = ensure_socket_dir_is_trusted(&shared)
                .expect_err("group/other-writable must be refused, sticky or not");
            assert!(
                err.to_string().contains(&format!("{:04o}", mode & 0o7777)),
                "the refusal should name the mode it saw, got: {err}"
            );
            let after = std::fs::metadata(&shared)
                .expect("stat")
                .permissions()
                .mode()
                & 0o7777;
            assert_eq!(
                after, mode,
                "refusing must not re-permission a directory khive does not own"
            );
        }
    }

    /// A vetted final directory is still refused when an ANCESTOR would let
    /// another local user rename it away and recreate it: a non-sticky
    /// group/other-writable ancestor re-roots the whole socket path without
    /// the final directory's own metadata ever changing. The sticky-ancestor
    /// arm (a root-owned 1777 `/tmp` above a user-owned 0700 directory is
    /// acceptable) is exercised implicitly by every accepting test below,
    /// whose tempdirs live under the platform temp root.
    #[test]
    fn writable_non_sticky_ancestor_is_refused() {
        let dir = tempfile::tempdir().expect("tempdir");
        let open_mid = dir.path().join("open-mid");
        std::fs::create_dir(&open_mid).expect("create mid");
        let inner = open_mid.join("private");
        std::fs::create_dir(&inner).expect("create inner");
        std::fs::set_permissions(&inner, std::fs::Permissions::from_mode(0o700)).expect("chmod");
        std::fs::set_permissions(&open_mid, std::fs::Permissions::from_mode(0o777))
            .expect("chmod mid");

        let err = ensure_socket_dir_is_trusted(&inner)
            .expect_err("a 0777 non-sticky ancestor must be refused");
        assert!(
            err.to_string().contains("ancestor"),
            "the refusal should say it was an ancestor that failed, got: {err}"
        );
        assert!(
            err.to_string().contains("open-mid"),
            "the refusal should name the failing ancestor, got: {err}"
        );
    }

    /// The path walk validates what a symlink component points AT, not just
    /// the symlink node: a trusted (self-owned) link into a group/other-
    /// writable non-sticky directory is refused on the target directory's
    /// metadata, proving the walk keeps traversing — and keeps applying the
    /// directory rules — past the link, exactly as the kernel will at bind
    /// time.
    #[test]
    fn symlink_component_to_untrusted_directory_is_refused() {
        let dir = tempfile::tempdir().expect("tempdir");
        let open = dir.path().join("open-target");
        std::fs::create_dir(&open).expect("create target");
        std::fs::set_permissions(&open, std::fs::Permissions::from_mode(0o777)).expect("chmod");
        let link = dir.path().join("link");
        std::os::unix::fs::symlink(&open, &link).expect("symlink");

        // SAFETY: `geteuid` is always successful and takes no arguments.
        let euid = unsafe { libc::geteuid() } as u32;
        let err = ensure_socket_path_is_swap_resistant(&link, euid)
            .expect_err("a link into a 0777 non-sticky directory must be refused");
        assert!(
            err.to_string().contains("open-target"),
            "the refusal should name the untrusted target directory, got: {err}"
        );
    }

    /// A symlink component owned by neither the daemon euid nor root is
    /// refused on the symlink itself: its owner can retarget it after
    /// validation, so where it currently points is irrelevant. A non-root
    /// test cannot create a foreign-owned symlink, so this injects a
    /// mismatched euid instead — and the fixture must live under an
    /// all-root-owned chain (the platform `/tmp`), because under the test's
    /// own tempdir the self-owned ancestor directories would already fail
    /// the injected euid before the walk ever reached the link.
    #[test]
    fn foreign_owned_symlink_component_is_refused() {
        let link =
            std::path::PathBuf::from(format!("/tmp/khive-swaptest-link-{}", std::process::id()));
        let _ = std::fs::remove_file(&link);
        std::os::unix::fs::symlink("/tmp", &link).expect("symlink");

        // SAFETY: `geteuid` is always successful and takes no arguments.
        let not_my_euid = (unsafe { libc::geteuid() } as u32).wrapping_add(1);
        let result = ensure_socket_path_is_swap_resistant(&link, not_my_euid);
        std::fs::remove_file(&link).expect("cleanup");

        let err = result.expect_err("a symlink owned by another uid must be refused");
        assert!(
            err.to_string().contains("symlink component"),
            "the refusal should strike the symlink itself, got: {err}"
        );
        assert!(
            err.to_string().contains("khive-swaptest-link"),
            "the refusal should name the offending link, got: {err}"
        );
    }

    /// Directories this daemon can trust end to end are served as found:
    /// owner-only, and the umask-default 0755 every `tempfile::tempdir` and
    /// test runner produces (readable/traversable but writable only by the
    /// owner). On macOS these fixtures also implicitly exercise the
    /// symlink-accept arm, since the platform temp root itself resolves
    /// through root-owned symlinks. Foreign ownership of a directory is
    /// refused by the same helper, but a non-root test cannot chown a
    /// directory away from itself, so that arm is exercised by the mode
    /// checks' shared fail-closed path rather than a dedicated fixture.
    #[test]
    fn trusted_socket_dirs_are_accepted_unmodified() {
        let dir = tempfile::tempdir().expect("tempdir");
        for (name, mode) in [("private", 0o700), ("listable", 0o755)] {
            let d = dir.path().join(name);
            std::fs::create_dir(&d).expect("create");
            std::fs::set_permissions(&d, std::fs::Permissions::from_mode(mode)).expect("chmod");

            ensure_socket_dir_is_trusted(&d)
                .unwrap_or_else(|e| panic!("mode {mode:04o} must be accepted, got: {e}"));

            let after = std::fs::metadata(&d).expect("stat").permissions().mode() & 0o7777;
            assert_eq!(
                after, mode,
                "acceptance must not re-permission the directory either"
            );
        }
    }

    /// Test 2: `wal_pages` reflects a real
    /// checkpoint observation after writes, deterministically forced via a
    /// direct `checkpoint_once` call rather than waiting on the async
    /// periodic task.
    #[tokio::test]
    async fn metrics_snapshot_wal_pages_reflects_recent_write() {
        let dir = tempfile::tempdir().expect("tempdir");
        let path = dir.path().join("metrics_wal_test.db");
        let pool = Arc::new(
            ConnectionPool::new(khive_db::PoolConfig {
                path: Some(path),
                ..khive_db::PoolConfig::default()
            })
            .expect("pool open"),
        );

        {
            let writer = pool.try_writer().expect("writer");
            writer
                .conn()
                .execute_batch(
                    "CREATE TABLE t (x INTEGER); \
                     INSERT INTO t VALUES (1); \
                     INSERT INTO t VALUES (2);",
                )
                .expect("seed writes");
        }

        let dedicated_conn = pool
            .open_standalone_writer()
            .expect("open dedicated checkpoint connection");
        khive_db::checkpoint_once(
            &pool,
            &dedicated_conn,
            &CheckpointConfig::default(),
            &mut khive_db::checkpoint::TruncateState::default(),
        )
        .expect("checkpoint_once must observe on a healthy dedicated connection");

        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-wal".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: Some(pool),
            dispatch_err: None,
        };

        let snapshot = build_metrics_snapshot(&dispatcher);
        assert!(
            snapshot.wal_pages.is_some(),
            "wal_pages must be observed after a real checkpoint tick, got {snapshot:?}"
        );
        assert_eq!(snapshot.wal_log_frames, snapshot.wal_pages);
        assert!(snapshot.wal_checkpointed_frames.is_some());
        assert!(snapshot.wal_pending_frames.is_some());
        assert!(snapshot.wal_physical_bytes.is_some());
        assert!(snapshot.wal_observed_at_unix_ms.is_some());
        // The snapshot carries the checkpoint-pressure fields read-only
        // (no mutation path reachable through `MetricsSnapshot`/`DaemonRequestFrame`);
        // an observed tick (not a skip) must report a zero-length skip streak.
        assert_eq!(
            snapshot.wal_checkpoint_consecutive_skips, 0,
            "an observed (non-skipped) tick must report zero consecutive skips, got {snapshot:?}"
        );
    }

    #[tokio::test]
    async fn metrics_snapshot_exposes_decomposed_writer_stages() {
        let dir = tempfile::tempdir().expect("tempdir");
        let path = dir.path().join("metrics_writer_stage_test.db");
        let pool = Arc::new(
            ConnectionPool::new(khive_db::PoolConfig {
                path: Some(path),
                ..khive_db::PoolConfig::default()
            })
            .expect("pool open"),
        );
        {
            let writer = pool.try_writer().unwrap();
            writer
                .conn()
                .execute_batch("CREATE TABLE t (id INTEGER PRIMARY KEY)")
                .unwrap();
        }
        let handle = pool
            .writer_task_handle()
            .unwrap()
            .expect("file-backed default writer task");
        handle
            .send(|conn| {
                std::thread::sleep(std::time::Duration::from_millis(30));
                conn.execute("INSERT INTO t VALUES (1)", [])
                    .map_err(|error| khive_storage::error::StorageError::Pool {
                        operation: "metrics_writer_stage_test".into(),
                        message: error.to_string(),
                    })
            })
            .await
            .unwrap();

        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-writer-stages".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: Some(pool),
            dispatch_err: None,
        };
        let snapshot = build_metrics_snapshot(&dispatcher);
        assert!(snapshot.write_last_queue_wait_micros.is_some());
        assert!(snapshot.write_last_transaction_acquire_micros.is_some());
        assert!(snapshot.write_last_commit_micros.is_some());
        assert!(
            snapshot.write_last_body_micros >= Some(25_000),
            "synthetic delay must be attributed to the body: {snapshot:?}"
        );
        assert!(snapshot.write_last_total_micros >= snapshot.write_last_body_micros);
        assert!(snapshot.write_last_observed_at_unix_ms.is_some());
    }

    /// Test 3: the tx-pin oracle. The registry is process-global, so an
    /// unrelated transaction can depart between snapshots and exactly offset
    /// this test's registration. Keep an owned handle live and assert the
    /// resulting count floor instead of comparing two points in time.
    #[test]
    #[serial(tx_registry)]
    fn metrics_snapshot_reflects_open_transaction_registry() {
        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-tx".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: None,
            dispatch_err: None,
        };

        let departing_handle = khive_storage::tx_registry::register(Some(
            "daemon_metrics_snapshot_departing_test_tx".to_string(),
        ));
        let before = build_metrics_snapshot(&dispatcher).open_tx_count;
        assert!(before >= 1);

        let handle = khive_storage::tx_registry::register(Some(
            "daemon_metrics_snapshot_owned_test_tx".to_string(),
        ));
        drop(departing_handle);

        let during = build_metrics_snapshot(&dispatcher);
        assert!(
            during.open_tx_count >= 1,
            "open_tx_count must reflect the live owned transaction despite registry churn: \
             churn_baseline={before} during={}",
            during.open_tx_count
        );
        assert!(
            during.oldest_pinned_tx_micros.is_some(),
            "oldest_pinned_tx_micros must be Some while a transaction is open"
        );

        drop(handle);
        assert!(
            !khive_storage::tx_registry::snapshot()
                .iter()
                .any(|(_, label)| label.as_deref()
                    == Some("daemon_metrics_snapshot_owned_test_tx")),
            "the owned registry entry must disappear when its handle is dropped"
        );
    }

    /// Test 4: write-queue depth is flag-gated
    /// on `PoolConfig::write_queue_enabled` (the `KHIVE_WRITE_QUEUE=1`
    /// setting), never on a specific depth value (racy under concurrency).
    #[tokio::test]
    async fn metrics_snapshot_write_queue_depth_flag_gated() {
        let dir = tempfile::tempdir().expect("tempdir");

        let enabled_pool = Arc::new(
            ConnectionPool::new(khive_db::PoolConfig {
                path: Some(dir.path().join("wq_enabled.db")),
                write_queue_enabled: Some(true),
                ..khive_db::PoolConfig::default()
            })
            .expect("pool open"),
        );
        let enabled_dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-wq-on".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: Some(enabled_pool),
            dispatch_err: None,
        };
        let snapshot_on = build_metrics_snapshot(&enabled_dispatcher);
        assert!(
            snapshot_on.write_queue_depth.is_some(),
            "write_queue_depth must be Some when write_queue_enabled=true, got {snapshot_on:?}"
        );
        assert!(snapshot_on.write_queue_capacity.is_some());

        let disabled_pool = Arc::new(
            ConnectionPool::new(khive_db::PoolConfig {
                path: Some(dir.path().join("wq_disabled.db")),
                write_queue_enabled: Some(false),
                ..khive_db::PoolConfig::default()
            })
            .expect("pool open"),
        );
        let disabled_dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-wq-off".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: Some(disabled_pool),
            dispatch_err: None,
        };
        let snapshot_off = build_metrics_snapshot(&disabled_dispatcher);
        assert!(
            snapshot_off.write_queue_depth.is_none(),
            "write_queue_depth must be None when write_queue_enabled=false, got {snapshot_off:?}"
        );
        assert!(snapshot_off.write_queue_capacity.is_none());

        // No pool at all (in-memory/poolless dispatcher): also None.
        let no_pool_dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-no-pool".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: None,
            dispatch_err: None,
        };
        let snapshot_no_pool = build_metrics_snapshot(&no_pool_dispatcher);
        assert!(snapshot_no_pool.write_queue_depth.is_none());
        assert!(snapshot_no_pool.write_queue_capacity.is_none());
    }

    /// Test 5: serde default back-compat in both directions — a request JSON
    /// without additive request fields deserializes to their defaults, and a
    /// response JSON without `metrics` (an old daemon's shape) deserializes
    /// with it `None`.
    #[test]
    fn frame_serde_defaults_additive_fields_when_absent() {
        let req_json = serde_json::json!({
            "ops": "",
            "presentation": null,
            "presentation_per_op": null,
            "namespace": "local",
            "actor_id": null,
            "visible_namespaces": [],
            "config_id": "cfg",
            "protocol_version": PROTOCOL_VERSION,
            "probe_only": false,
            "format": null,
            "format_per_op": null,
            "from_wire": false
        });
        let frame: DaemonRequestFrame =
            serde_json::from_value(req_json).expect("decode a metrics_only-absent request frame");
        assert!(
            !frame.metrics_only,
            "metrics_only must default to false when absent from the wire payload"
        );
        assert_eq!(
            frame.request_id, None,
            "request_id must default to None when absent from the wire payload (khive#948)"
        );
        assert_eq!(
            frame.process_ref, None,
            "process_ref must default to None when absent from the wire payload (khive#1428)"
        );
        let encoded_frame = serde_json::to_value(&frame).expect("encode request frame");
        assert!(
            encoded_frame.get("process_ref").is_none(),
            "absent provenance must not change the serialized request wire shape"
        );

        let resp_json = serde_json::json!({
            "ok": true,
            "result": null,
            "error": null,
            "namespace_mismatch": false,
            "config_mismatch": false,
            "served_config_id": "cfg",
            "version_mismatch": false,
            "daemon_protocol_version": PROTOCOL_VERSION
        });
        let resp: DaemonResponseFrame =
            serde_json::from_value(resp_json).expect("decode a metrics-absent response frame");
        assert!(
            resp.metrics.is_none(),
            "metrics must default to None when absent from the wire payload"
        );
        assert_eq!(
            resp.request_id, None,
            "request_id must default to None when absent from the wire payload (khive#948)"
        );
    }

    /// khive#948: a request carrying `request_id: Some(n)` gets back a
    /// response with `request_id: Some(n)` on both the success and the
    /// error/denied dispatch arms — the echo must survive every branch of
    /// `handle_conn`, not only the happy path.
    #[tokio::test]
    async fn request_id_echoed_on_success_and_error_arms() {
        let dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-a".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: None,
            dispatch_err: None,
        };
        let mut ok_req = base_request_frame("cfg-a");
        ok_req.request_id = Some(42);
        let ok_resp = round_trip(dispatcher, &ok_req).await;
        assert!(ok_resp.ok, "expected successful dispatch: {ok_resp:?}");
        assert_eq!(
            ok_resp.request_id,
            Some(42),
            "request_id must be echoed back on a successful dispatch response"
        );

        // config_mismatch is a rejection arm that never reaches dispatch —
        // must still echo the id so the client can join the failure.
        let mismatched_dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-a".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: None,
            dispatch_err: None,
        };
        let mut mismatch_req = base_request_frame("cfg-WRONG");
        mismatch_req.request_id = Some(99);
        let mismatch_resp = round_trip(mismatched_dispatcher, &mismatch_req).await;
        assert!(mismatch_resp.config_mismatch);
        assert_eq!(
            mismatch_resp.request_id,
            Some(99),
            "request_id must be echoed on the config_mismatch rejection arm too"
        );

        // The real ops-dispatch error arm (`Err(e)` from `dispatcher.dispatch`)
        // must echo the id as well, not only the pre-dispatch rejection arms.
        let erroring_dispatcher = MockDispatch {
            namespace: "local".to_string(),
            config_id: "cfg-a".to_string(),
            dispatch_calls: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
            pool: None,
            dispatch_err: Some("simulated dispatch error".to_string()),
        };
        let mut err_req = base_request_frame("cfg-a");
        err_req.request_id = Some(7);
        let err_resp = round_trip(erroring_dispatcher, &err_req).await;
        assert!(!err_resp.ok, "expected a dispatch error: {err_resp:?}");
        assert_eq!(
            err_resp.request_id,
            Some(7),
            "request_id must be echoed on the real ops-dispatch error arm"
        );
    }

    // ── owner-checked shutdown cleanup ────────────────────────────────────────
    //
    // A draining daemon must not unlink a socket/PID pair that a replacement
    // daemon has already bound. These tests exercise `shutdown_cleanup_if_owned`
    // directly (the pure decision the caller makes under the recovery lock)
    // rather than driving `run_daemon`'s real SIGTERM shutdown, which would
    // require sending a signal to the whole test process.

    #[test]
    fn shutdown_cleanup_removes_paths_it_still_owns() {
        let dir = tempfile::tempdir().expect("tempdir");
        let sock = dir.path().join("khived.sock");
        let pid_file = dir.path().join("khived.pid");

        let _listener = std::os::unix::net::UnixListener::bind(&sock).expect("bind socket");
        std::fs::write(&pid_file, std::process::id().to_string()).expect("write pid file");
        let identity = socket_identity(&sock);
        assert!(
            identity.is_some(),
            "must read identity of a freshly bound socket"
        );

        let cleaned = shutdown_cleanup_if_owned(&sock, &pid_file, identity);

        assert!(
            cleaned,
            "cleanup must proceed when PID and socket still match"
        );
        assert!(!sock.exists(), "owned socket must be removed");
        assert!(!pid_file.exists(), "owned pid file must be removed");
    }

    #[test]
    fn shutdown_cleanup_skips_when_pid_file_names_a_different_process() {
        let dir = tempfile::tempdir().expect("tempdir");
        let sock = dir.path().join("khived.sock");
        let pid_file = dir.path().join("khived.pid");

        let _listener = std::os::unix::net::UnixListener::bind(&sock).expect("bind socket");
        let identity = socket_identity(&sock);
        // A concurrent client's kill_and_respawn already replaced the PID file
        // with a different (replacement daemon's) PID before this daemon's
        // drain completed.
        std::fs::write(&pid_file, "1").expect("write foreign pid file");

        let cleaned = shutdown_cleanup_if_owned(&sock, &pid_file, identity);

        assert!(
            !cleaned,
            "cleanup must be skipped when the PID file no longer names this process"
        );
        assert!(sock.exists(), "replacement daemon's socket must survive");
        assert!(
            pid_file.exists(),
            "replacement daemon's pid file must survive"
        );
    }

    #[test]
    fn shutdown_cleanup_skips_when_socket_was_rebound_by_a_replacement() {
        let dir = tempfile::tempdir().expect("tempdir");
        let sock = dir.path().join("khived.sock");
        let original_sock = dir.path().join("original.sock");
        let pid_file = dir.path().join("khived.pid");

        // Bind two sockets at DIFFERENT paths, both alive at the same time,
        // so the OS cannot recycle an inode between them the way it could
        // across a bind/drop/rebind cycle at a single path (the flakiness a
        // prior version of this test hit on some filesystems). Both
        // identities are captured through the real production
        // `socket_identity()` path, not a synthetic/sentinel value, so a
        // regression where `socket_identity()` returns a constant identity
        // for every socket makes the `assert!` below fail loudly instead of
        // silently passing.
        let _original_listener =
            std::os::unix::net::UnixListener::bind(&original_sock).expect("bind original socket");
        let _replacement_listener =
            std::os::unix::net::UnixListener::bind(&sock).expect("bind replacement socket");

        let original_identity = socket_identity(&original_sock);
        let replacement_identity = socket_identity(&sock);
        assert!(
            original_identity.is_some(),
            "must read identity of the original socket"
        );
        assert!(
            replacement_identity.is_some(),
            "must read identity of the replacement socket"
        );
        assert!(
            original_identity != replacement_identity,
            "two concurrently bound sockets must have distinct identities"
        );

        std::fs::write(&pid_file, std::process::id().to_string())
            .expect("write pid file matching this process");

        // `sock` (the replacement bind's path) is checked against
        // `original_identity` (a different, concurrently-alive socket's
        // identity) - the mismatch alone must be enough to block cleanup,
        // even though the pid file matches this process.
        let cleaned = shutdown_cleanup_if_owned(&sock, &pid_file, original_identity);

        assert!(
            !cleaned,
            "cleanup must be skipped when the socket at this path is a different \
             inode than the one this daemon originally bound"
        );
        assert!(sock.exists(), "replacement daemon's socket must survive");
        assert!(
            pid_file.exists(),
            "replacement daemon's pid file must survive"
        );
    }

    // ── the recovery lock actually serializes two boot sequences ─────────────
    //
    // Production wiring (`khive_mcp::serve::run` / `serve_server`) now acquires
    // this same lock *before* building a `KhiveMcpServer` (which runs
    // migrations and applies pack schema plans / FTS DDL) and holds it through
    // daemon bind+pid-write, via `run_daemon_with_boot_guard`. That closes the
    // cold-boot race only if `acquire_recovery_lock` genuinely provides mutual
    // exclusion across concurrent boot attempts — this test proves the
    // primitive itself: two "boot sequences" (each holding the lock across a
    // simulated schema-init critical section) must never run their critical
    // sections at the same time.
    #[test]
    #[serial]
    fn recovery_lock_serializes_two_concurrent_boot_sequences() {
        let dir = tempfile::tempdir().expect("tempdir");
        let lock_file = dir.path().join("khived.recovery.lock");
        std::env::set_var("KHIVE_LOCK", &lock_file);

        let active = Arc::new(std::sync::atomic::AtomicUsize::new(0));
        let overlap_detected = Arc::new(std::sync::atomic::AtomicBool::new(false));

        let run_one_boot =
            |active: Arc<std::sync::atomic::AtomicUsize>,
             overlap: Arc<std::sync::atomic::AtomicBool>| {
                move || {
                    let _guard = acquire_recovery_lock().expect("acquire recovery lock");
                    // Enter the "schema-init" critical section.
                    if active.fetch_add(1, std::sync::atomic::Ordering::SeqCst) != 0 {
                        overlap.store(true, std::sync::atomic::Ordering::SeqCst);
                    }
                    std::thread::sleep(std::time::Duration::from_millis(50));
                    active.fetch_sub(1, std::sync::atomic::Ordering::SeqCst);
                    // `_guard` drops here, releasing the lock.
                }
            };

        let t1 = std::thread::spawn(run_one_boot(active.clone(), overlap_detected.clone()));
        let t2 = std::thread::spawn(run_one_boot(active.clone(), overlap_detected.clone()));
        t1.join().expect("boot thread 1 must not panic");
        t2.join().expect("boot thread 2 must not panic");

        assert!(
            !overlap_detected.load(std::sync::atomic::Ordering::SeqCst),
            "two concurrent boot sequences must never hold the schema-init \
             critical section at the same time (#667)"
        );

        std::env::remove_var("KHIVE_LOCK");
    }

    // ── acquire_daemon_boot_guard treats lock failure as fatal ───────────────
    // (unlike best-effort acquire_recovery_lock, whose `None` on failure is
    // correct for its own best-effort callers).

    #[test]
    #[serial]
    fn acquire_daemon_boot_guard_returns_guard_when_lock_available() {
        let dir = tempfile::tempdir().expect("tempdir");
        let lock_file = dir.path().join("khived.recovery.lock");
        std::env::set_var("KHIVE_LOCK", &lock_file);

        let guard = acquire_daemon_boot_guard();
        assert!(
            guard.is_ok(),
            "daemon boot guard must succeed when the lock file can be opened and flocked"
        );
        drop(guard);

        std::env::remove_var("KHIVE_LOCK");
    }

    #[test]
    #[serial]
    fn acquire_daemon_boot_guard_fails_loudly_when_lock_file_cannot_be_opened() {
        let dir = tempfile::tempdir().expect("tempdir");
        // Point KHIVE_LOCK at a directory, not a file: opening a directory
        // with `write(true)` fails (EISDIR), so `acquire_recovery_lock`
        // returns `None` here — the exact failure mode
        // `acquire_daemon_boot_guard` must turn into a hard `Err` instead of
        // silently letting daemon-mode boot proceed unguarded.
        std::env::set_var("KHIVE_LOCK", dir.path());

        let result = acquire_daemon_boot_guard();
        assert!(
            result.is_err(),
            "daemon boot guard must fail loudly, never silently proceed unguarded, \
             when the underlying recovery lock cannot be acquired"
        );

        std::env::remove_var("KHIVE_LOCK");
    }

    // ── write_pid_file_exclusive never truncates a winner's pid file ────────

    #[test]
    fn write_pid_file_exclusive_creates_new_file_with_own_pid() {
        let dir = tempfile::tempdir().expect("tempdir");
        let pid_file = dir.path().join("khived.pid");
        write_pid_file_exclusive(&pid_file).expect("first writer must win");
        let contents = std::fs::read_to_string(&pid_file).expect("read pid file");
        assert_eq!(contents, std::process::id().to_string());
    }

    #[test]
    fn write_pid_file_exclusive_refuses_to_overwrite_an_existing_file() {
        let dir = tempfile::tempdir().expect("tempdir");
        let pid_file = dir.path().join("khived.pid");
        std::fs::write(&pid_file, "999999").expect("seed an existing pid file");

        let err = write_pid_file_exclusive(&pid_file)
            .expect_err("must not silently overwrite an existing pid file");
        assert_eq!(err.kind(), std::io::ErrorKind::AlreadyExists);

        // The existing content must be completely untouched — proving this is
        // `create_new`, not the old `create(true).truncate(true)`.
        let contents = std::fs::read_to_string(&pid_file).expect("read pid file");
        assert_eq!(
            contents, "999999",
            "an existing pid file must never be truncated by a losing writer"
        );
    }

    // Real (not simulated) concurrency: two OS threads race to `create_new`
    // the exact same path, synchronized with a `Barrier` so they genuinely
    // overlap at the syscall rather than relying on a sleep-based ordering
    // guess. This is the deterministic race oracle for the convergence
    // requirement the atomic-creation primitive `write_pid_file_exclusive`
    // is built on: exactly one of two simultaneous daemon starters may claim
    // the pid file, and the loser must see `AlreadyExists`, never silently
    // clobber the winner's content.
    #[test]
    fn two_concurrent_writers_converge_on_exactly_one_pid_file_owner() {
        let dir = tempfile::tempdir().expect("tempdir");
        let pid_file = std::sync::Arc::new(dir.path().join("khived.pid"));
        let barrier = std::sync::Arc::new(std::sync::Barrier::new(2));

        let spawn_writer =
            |pid_file: std::sync::Arc<std::path::PathBuf>,
             barrier: std::sync::Arc<std::sync::Barrier>| {
                std::thread::spawn(move || {
                    barrier.wait();
                    write_pid_file_exclusive(&pid_file)
                })
            };

        let t1 = spawn_writer(pid_file.clone(), barrier.clone());
        let t2 = spawn_writer(pid_file.clone(), barrier.clone());
        let r1 = t1.join().expect("writer 1 must not panic");
        let r2 = t2.join().expect("writer 2 must not panic");

        let results = [&r1, &r2];
        let ok_count = results.iter().filter(|r| r.is_ok()).count();
        let already_exists_count = results
            .iter()
            .filter(|r| matches!(r, Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists))
            .count();
        assert_eq!(
            ok_count, 1,
            "exactly one of two concurrent writers must win the pid file"
        );
        assert_eq!(
            already_exists_count, 1,
            "the other writer must observe AlreadyExists, never a silent overwrite"
        );
        assert!(pid_file.exists(), "the winner's pid file must exist");
        let contents = std::fs::read_to_string(&*pid_file).expect("read pid file");
        assert_eq!(
            contents,
            std::process::id().to_string(),
            "the surviving pid file must contain the winner's pid — both threads \
             share this process's pid, so an unexpected value would also prove a \
             lost/garbled write raced through"
        );
    }

    // ── connection principal (ADR-096 condition 2) ────────────────────────────

    /// The peer-credential syscall must actually work on this platform and
    /// report the real uid, not error or return a placeholder.
    ///
    /// This matters more than it looks because the accept path fails CLOSED: if
    /// `peer_uid` errored unconditionally — wrong syscall, wrong socket option,
    /// an unimplemented platform arm — every connection would be refused and
    /// the daemon would be silently unreachable. A test that only exercised the
    /// decision function would not catch that.
    #[tokio::test]
    async fn peer_uid_reports_the_connecting_process_uid() {
        let dir = tempfile::tempdir().expect("tempdir");
        let sock = dir.path().join("peer.sock");
        let listener = UnixListener::bind(&sock).expect("bind");

        let connect_path = sock.clone();
        let client = tokio::spawn(async move { UnixStream::connect(&connect_path).await });

        let (server_side, _) = listener.accept().await.expect("accept");
        let client_side = client.await.expect("join").expect("connect");

        // SAFETY: `geteuid` is always successful and takes no arguments.
        let expected = unsafe { libc::geteuid() } as u32;

        assert_eq!(
            peer_uid(&server_side).expect("peer_uid must succeed on a live connection"),
            expected,
            "the uid read from the kernel for a same-process connection must be \
             this process's euid"
        );
        // Symmetric: both ends report the same peer on a same-uid connection.
        assert_eq!(
            peer_uid(&client_side).expect("peer_uid must succeed on the client end"),
            expected
        );
    }

    /// The refusal rule itself: the principal is the uid, and only a foreign
    /// uid is refused.
    ///
    /// The second assertion is the load-bearing one and it is a regression
    /// guard, not a formality. ADR-096 shipped many `actor_id`s over one
    /// socket; every seat on a normal host is a distinct attribution at the
    /// same uid. A check that conflated attribution with principal would refuse
    /// them all, so "same uid is permitted" must stay true no matter how the
    /// rule is later tightened.
    #[test]
    fn only_a_foreign_uid_is_refused() {
        // SAFETY: `geteuid` is always successful and takes no arguments.
        let euid = unsafe { libc::geteuid() } as u32;

        assert!(
            uid_is_permitted(euid, euid),
            "a connection from the daemon's own uid must be served — this is \
             every seat on the host, and ADR-096 accepted exactly this shape"
        );
        assert!(
            !uid_is_permitted(euid.wrapping_add(1), euid),
            "a connection from any other uid must be refused"
        );
        assert!(
            !uid_is_permitted(0, euid.wrapping_add(1)),
            "root is not special-cased: the rule is equality with the daemon's \
             euid, not a privilege comparison"
        );
    }
}