readsight 1.0.0

Multilingual readability library — 86 languages, 17 formulas, TeX-based syllable counting via the Frank M. Liang algorithm.
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
# ReadSight → Rust Port Guide (`readsight-rs`)

> An exhaustive, implementation-ready specification for building a **complete and
> byte-accurate Rust port** of ReadSight. This document is written to be handed
> directly to an AI coding agent (or a human) and followed top-to-bottom. Every
> algorithm, constant, coefficient, rounding rule, and edge case that matters for
> output parity is spelled out.

## 0. Source repositories (the ground truth)

You are porting an existing, tested library. **Always cross-check against the
reference implementations** — do not invent behaviour.

| Reference | URL | Role |
|---|---|---|
| **PHP (canonical)** | https://github.com/MADEVAL/ReadSight | Primary source of truth. Class/algorithm layout below mirrors it 1:1. |
| **Python port** | https://github.com/MADEVAL/ReadSightPy | Second reference; a proven, idiomatic port of the same logic. Use it to disambiguate anything unclear in PHP. |
| Python on PyPI | https://pypi.org/project/readsight/ | Installable reference (`pip install readsight`) for generating golden vectors. |
| TeX hyphenation data | https://ctan.org/pkg/hyph-utf8 | Origin of the `.tex` pattern files (version 2026-02-21). |

Key raw files you will re-read constantly:

- PHP source tree: `https://github.com/MADEVAL/ReadSight/tree/main/src`
- PHP data: `https://github.com/MADEVAL/ReadSight/tree/main/data` (`languages/*.json`, `patterns/hyph-*.tex`)
- PHP tests: `https://github.com/MADEVAL/ReadSight/tree/main/tests`
- Python source: `https://github.com/MADEVAL/ReadSightPy/tree/main/src/readsight`
- Python data (identical JSON + `.tex`): `https://github.com/MADEVAL/ReadSightPy/tree/main/src/readsight/data`

**Rule of thumb:** if PHP and Python ever disagree, PHP wins (it is canonical),
but the disagreement is a red flag — investigate before continuing.

---

## 1. What ReadSight does (mental model)

ReadSight measures **text readability** across **86 languages** using **17
readability formulas**, plus a **syllable counter** and **word hyphenator** based
on the **Frank M. Liang (TeX) hyphenation algorithm**. Zero network, zero heavy
deps; all data ships in the package (`data/languages/*.json`,
`data/patterns/hyph-*.tex`).

The pipeline for a single call is:

```
text ──► TextSplitter (words / sentences / letters via language regexes)
     ──► SyllableCounter (per word: tex | heuristic | composite)
     ──► TextStatistics (aggregate metrics)
     ──► Formula.calculate(stats, language) ──► FormulaResult
```

Syllable counting has two independent outputs:
- **count** (`syllable_count`) — a number.
- **split** (`split_word` = exact TeX hyphenation points; `split_syllables` =
  possibly-approximate split depending on mode).

---

## 2. Target crate layout

Create a library crate `readsight` (binary demo optional). Recommended layout —
it mirrors the PHP namespaces so cross-referencing is trivial:

```
readsight-rs/
├── Cargo.toml
├── build.rs                      # optional: embed / precompile data
├── data/                         # COPY verbatim from PHP repo
│   ├── languages/*.json          # 86 files
│   └── patterns/hyph-*.tex       # 86 files
├── src/
│   ├── lib.rs                    # re-exports; crate docs
│   ├── engine.rs                 # ReadSight facade (PHP Engine.php)
│   ├── config.rs                 # Config (paths / data source)
│   ├── error.rs                  # Error enum (PHP Exception/*)
│   ├── language/
│   │   ├── mod.rs
│   │   ├── language.rs           # Language struct (parsed JSON)
│   │   ├── code.rs               # LanguageCode::normalize
│   │   ├── script.rs             # Script enum
│   │   └── repository.rs         # JsonLanguageRepository
│   ├── hyphenation/
│   │   ├── mod.rs
│   │   ├── hyphenator.rs         # trait Hyphenator + LiangHyphenator
│   │   ├── pattern.rs            # Pattern + PatternsCollection
│   │   ├── exceptions.rs         # HyphenationOverride + collection
│   │   ├── tex_source.rs         # TeX .tex parser
│   │   └── cache.rs              # optional JSON pattern cache
│   ├── syllable/
│   │   ├── mod.rs
│   │   ├── counter.rs            # trait SyllableCounter
│   │   ├── tex.rs                # TexSyllableCounter
│   │   ├── heuristic.rs          # HeuristicSyllableCounter (+ vowelMode)
│   │   └── composite.rs          # CompositeSyllableCounter
│   ├── text/
│   │   ├── mod.rs
│   │   ├── splitter.rs           # TextSplitter
│   │   ├── analyzer.rs           # TextAnalyzer
│   │   └── statistics.rs         # TextStatistics
│   └── formula/
│       ├── mod.rs                # trait Formula + FormulaResult + registry
│       ├── registry.rs
│       ├── grade_level.rs        # GradeLevelInterpretation
│       ├── helper.rs             # TextStatisticsHelper
│       └── impls/                # one file per formula (17)
└── tests/
    ├── golden/                   # golden vectors dumped from PHP/Python
    ├── hyphenation.rs
    ├── syllable.rs
    ├── formula.rs
    └── multilingual.rs
```

### 2.1 `Cargo.toml`

```toml
[package]
name = "readsight"
version = "1.0.0"
edition = "2021"
rust-version = "1.74"
license = "MIT"
description = "Multilingual readability library — 86 languages, 17 formulas, TeX-based syllable counting via the Frank M. Liang algorithm."

[dependencies]
regex = "1"                 # Unicode \p{L}; NOTE: no look-around (not needed)
serde = { version = "1", features = ["derive"] }
serde_json = "1"
include_dir = "0.7"         # embed data/ into the binary (recommended)

[dev-dependencies]
serde_json = "1"
```

**Do not** pull in an ICU crate for the hyphenation/formula logic. PHP uses
`mbstring` + PCRE; Python uses `regex`. The `regex` crate with default Unicode
features is the correct analogue. See §12 for the exact string-op mapping.

---

## 3. Data embedding

PHP resolves data via `Config { patternsDir, languagesDir, cacheDir }` (see
`src/Config.php`). Mirror that, **and** allow embedded data as the default so the
crate works with no filesystem.

```rust
// config.rs
pub struct Config {
    pub source: DataSource,
    pub cache_dir: Option<PathBuf>,   // caching is OPTIONAL in Rust (see §7.4)
}

pub enum DataSource {
    Embedded,                  // default: include_dir!("$CARGO_MANIFEST_DIR/data")
    Filesystem { patterns_dir: PathBuf, languages_dir: PathBuf },
}
```

- Copy the **entire** `data/` directory from the PHP repo verbatim (86 JSON + 86
  `.tex`). Do not regenerate or "clean up" — file contents (including the Cyrillic
  vowel-mode configs) are part of the contract.
- Embed with `include_dir!` (or `rust-embed`). Look-ups: languages by
  `"{code}.json"`, patterns by `"hyph-{code}.tex"`.

---

## 4. Language layer

### 4.1 `LanguageCode::normalize` (`src/Language/LanguageCode.php`)
```rust
pub fn normalize(code: &str) -> String { code.trim().to_lowercase() }
```
Unicode-aware lowercase (PHP `mb_strtolower`). Rust `str::to_lowercase` matches.

### 4.2 `Script` enum (`src/Language/Script.php`)
String-backed enum, exact variants (value == name):
`Latin, Cyrillic, Arabic, Hebrew, Devanagari, Bengali, Greek, Armenian, Georgian,
Thai, Tamil, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi, Odia, Ethiopic,
Coptic, CJK, Other`. Deserialize from the JSON `script` string.

### 4.3 `Language` struct (`src/Language/Language.php`)
Parsed from a language JSON file. Fields (defaults in parentheses):

| Field | JSON key | Type | Default |
|---|---|---|---|
| `code` | `code` | String ||
| `name` | `name` | String ||
| `native_name` | `nativeName` | String ||
| `script` | `script` | Script ||
| `min_hyphen_left` | `hyphenMins.left` | i32 ||
| `min_hyphen_right` | `hyphenMins.right` | i32 ||
| `letter_pattern` | `letterPattern` | String (regex body) ||
| `word_split_pattern` | `wordSplitPattern` | String (regex body) ||
| `sentence_boundary_pattern` | `sentenceBoundaryPattern` | String (regex body) ||
| `formula_configs` | `formulas` | Map<String, Map<String, Value>> | `{}` |
| `syllable_heuristics` | `syllableHeuristics` | Option<Map<String, Value>> | `None` |
| `syllable_mode` | `syllableMode` | String | `"tex"` |

Methods: `supports_formula(name) -> formula_configs.contains_key(name)`,
`get_formula_config(name) -> Option<&Map>`, `get_supported_formulas() ->
Vec<String>` (= keys of `formula_configs`).

> ⚠️ **Critical distinction (do not conflate):** `Language::get_supported_formulas`
> (JSON keys) is **NOT** the list used by `Engine::get_supported_formulas`. The
> Engine list comes from each `Formula`'s own `supported_languages()` (see §9.2).
> The JSON `formulas` block only supplies **coefficients** (FRE `base/aslMult/
> aswMult`) and the **LIX `longWordThreshold`**.

### 4.4 `JsonLanguageRepository` (`src/Language/JsonLanguageRepository.php`)
- `find(code)`: normalize → look up `"{code}.json"` in the data source → parse →
  memoize in an internal cache map. Missing file ⇒ `UnsupportedLanguage` error.
- `list_codes()`: enumerate `*.json`, strip extension, **sorted ascending**.
- `exists(code)`: normalize then check presence.

---

## 5. Hyphenation layer (the Liang algorithm) — port with surgical precision

This is the highest-risk area for divergence. Port it **line by line** from
`src/Hyphenation/`, then validate against golden vectors (§10).

### 5.1 `Pattern` (`Pattern.php`)
```rust
pub struct Pattern { pub chars: Vec<String>, pub weights: Vec<i32>, pub length: usize }
// length = chars.len()
```
`chars` are Unicode "characters" (scalar values as 1-char strings); `weights.len()
== chars.len() + 1`.

### 5.2 `PatternsCollection` (`PatternsCollection.php`)
Stores a map **key → weights-as-digit-string**:
- `add(pattern)`: `key = pattern.chars.concat()`, `value = pattern.weights.map(to_string).concat()`; track `max_pattern_length = max(len, chars.len())`.
- `get_weights(subword) -> Option<&str>`.
- `max_length() -> usize`.

In Rust prefer `HashMap<String, Vec<u8>>` (weights as small ints) but keep the
**digit-string semantics** in mind when reading PHP: `weights[offset]` indexes a
byte/char that is a single ASCII digit.

### 5.3 `HyphenationOverride` + `HyphenationExceptionsCollection`
DTO `{ word, hyphenated }`. Collection is a `HashMap<String,String>`:
key = word (lowercased, hyphens removed), value = hyphenated form (with `-`).
Methods `add/has/get/all/count/is_empty`.

### 5.4 TeX parser `TexSource` (`Source/TexSource.php`)
Parses a hyph-utf8 `.tex` file. Read the file as UTF-8 **line by line**. Maintain
state `command: Option<String>` and `in_braces: bool`. For each line iterate
**byte offset** `offset` while `offset < line.len()` (bytes):

1. `%` and not `in_braces` ⇒ rest of line is a comment; `break` to next line.
2. `\` and not `in_braces` ⇒ match regex `^\\([a-zA-Z]+)` on the remaining slice;
   if it matches, set `command = group1`, advance by full match length; else
   advance 1.
3. `{` ⇒ if `command.is_some()` set `in_braces = true`; advance 1.
4. `}` and `in_braces``in_braces=false; command=None`; advance 1.
5. If `in_braces`:
   - command `"patterns"`: match `^(\S+)` (Unicode) on remaining slice ⇒ token;
     `parse_pattern_token(token)`; if `Some`, add to patterns; advance by match
     byte-length.
   - command `"hyphenation"`: match `^(\S+)` ⇒ token; `word = token.replace('-',"")`
     lowercased; `hyphenated = token.lowercased()`; add override; advance.
6. Otherwise advance 1.

Return `{ patterns, exceptions, max_pattern_length }`.

Only the `\patterns{…}` and `\hyphenation{…}` blocks matter; everything else
(comments, `\message`, `\endinput`, etc.) is ignored by the state machine above.

**`parse_pattern_token(token)`** (the trickiest bit — reproduce exactly):
- Split `token` into Unicode characters.
- `numbers = String::new()`, `chars: Vec<String>`, `expect_number = true`,
  `has_digit = false`.
- For each `ch`:
  - if `ch` is a digit (PHP `is_numeric` on a single char): `numbers.push(ch)`,
    `has_digit = true`, `expect_number = false`.
  - else: if `expect_number` then `numbers.push('0')`; `chars.push(ch)`;
    `expect_number = true`.
- After loop: if `expect_number` then `numbers.push('0')`.
- If `chars.is_empty() || !has_digit` ⇒ return `None`.
- `weights = numbers.chars().map(|d| d.to_digit(10))` as `Vec<i32>`.
- Return `Some(Pattern { chars, weights })`.

Worked example: token `a2ch` → chars `[a,c,h]`, numbers `"0200"` →
weights `[0,2,0,0]`, key `"ach"`, weight-string `"0200"`.
Token `.ach4` → chars `[.,a,c,h]`, weights `[0,0,0,0,4]`.

### 5.5 `LiangHyphenator` (`LiangHyphenator.php`)
Constructor: `patterns`, `exceptions`, `min_hyphen_left` (default 2),
`min_hyphen_right` (default 2), plus a mutable `user_hyphenations: HashMap<String,String>`.

`add_hyphenations(map)`: insert each `(word.to_lowercase(), hyphenated.to_lowercase())`.

**`hyphenate(word) -> Vec<String>`** — operate on Unicode chars (`mb_strlen`/
`mb_substr` == iterate `chars()`):
```
let n = word.chars().count();
if n == 0 { return vec![]; }
if n < min_hyphen_left + min_hyphen_right { return vec![word.to_string()]; }
let lower = word.to_lowercase();
if let Some(h) = user_hyphenations.get(&lower) { return split_by_hyphenation(h, word); }
if let Some(h) = exceptions.get(&lower)         { return split_by_hyphenation(h, word); }
split_by_patterns(word, n, &lower)
```

**`count_syllables(word)`** = `let p = hyphenate(word); if p.is_empty() {0} else {p.len()}`.

**`split_by_hyphenation(hyphenated, original)`**: walk `hyphenated` char by char;
on `-` push the accumulated part and reset; otherwise append the **original**
word's next char (preserving original case). Push trailing non-empty part. (This
maps hyphen positions from the lowercased override onto the original-cased word.)

**`split_by_patterns(word, word_len, word_lower)`** — reproduce indices EXACTLY:
```
text = format!(".{}.", word_lower);       // note leading & trailing '.'
let text_chars: Vec<char> = text.chars().collect();
let text_length = word_len + 2;
let mut pattern_length = patterns.max_length();
if pattern_length > text_length { pattern_length = text_length; }

let mut scores: HashMap<usize,i32> = {};   // (PHP uses a sparse array)
let end = text_length - min_hyphen_right;
for start in 0..end {
    let max_len = min(pattern_length, text_length - start);
    for len in 1..=max_len {
        let subword: String = text_chars[start..start+len].iter().collect();
        let Some(weights) = patterns.get_weights(&subword) else { continue };
        for (offset, w) in weights.iter().enumerate() {   // weights.len() == its char count
            let idx = start + offset;
            let score = *w as i32;
            if scores.get(&idx).map_or(true, |s| score > *s) { scores.insert(idx, score); }
        }
    }
}

// Reconstruct parts. `word` chars are 0-indexed; boundary '.' shifts by 1.
let word_chars: Vec<char> = word.chars().collect();
let mut parts: Vec<String> = vec![];
let mut part: String = word_chars[0..min_hyphen_left].iter().collect();
let break_end = text_length - min_hyphen_right;
for i in (min_hyphen_left + 1)..break_end {
    if let Some(score) = scores.get(&i) {
        if score & 1 != 0 { parts.push(std::mem::take(&mut part)); }
    }
    part.push(word_chars[i - 1]);       // (i-1) maps text index → word index
}
for i in break_end..(text_length - 1) {
    part.push(word_chars[i - 1]);
}
if !part.is_empty() { parts.push(part); }
parts
```
Notes:
- The **weight length equals the stored pattern's char count** (i.e. `chars.len()+1`
  from parsing → but stored as the digit string; iterate its digits).
- Odd score ⇒ hyphenation point (`score & 1`).
- The two trailing loops append the final `min_hyphen_right` region without
  considering break points.
- Do not "optimize" the index arithmetic; replicate it and lean on golden tests.

`trait Hyphenator { fn hyphenate(&self, w:&str)->Vec<String>; fn count_syllables(&self, w:&str)->usize; }`.

---

## 6. Syllable layer (`src/Syllable/`)

`trait SyllableCounter { fn count_syllables(&self,w:&str)->i64; fn split_syllables(&self,w:&str)->Vec<String>; }`

### 6.1 `TexSyllableCounter`
Thin delegate to the hyphenator: `count = hyphenator.count_syllables(w)`,
`split = hyphenator.hyphenate(w)`.

### 6.2 `HeuristicSyllableCounter` (`HeuristicSyllableCounter.php`) — includes v1.0.7 `vowelMode`
Config fields (all from `syllableHeuristics`):
`problem_words: Map<String,i64>`, `subtract_patterns: Vec<String>`,
`add_patterns: Vec<String>`, `prefixes: Map<String,i64>`, `suffixes: Map<String,i64>`,
`vowel_pattern: String` (default `"[aeiouy]"`), `vowel_mode: "cluster"|"individual"`
(default `"cluster"`; any value other than `"individual"` ⇒ `"cluster"`).
Precompute `vowel_chars = vowel_pattern.trim_matches(['[',']'])`.

**`count_syllables(word)`** — exact order:
1. `word = word.trim()`; if empty ⇒ `0`.
2. `lower = word.to_lowercase()`. If `problem_words` contains `lower` ⇒ return it.
3. `clean = lower` with all non-letters removed (`\p{L}` kept). If empty ⇒ `1`.
4. `affix = 0`. For each `(prefix, n)` in `prefixes` **in map order**: if `clean`
   starts with `prefix`, strip it and `affix += n`.
5. For each `(suffix, n)` in `suffixes`: if `clean` ends with `suffix`, strip it and `affix += n`.
6. `vowel_groups = count_vowel_groups(clean)` (see below).
7. `count = vowel_groups + affix`.
8. For each `p` in `subtract_patterns`: `count -= if regex(p).is_match(clean) {1} else {0}`.
9. For each `p` in `add_patterns`: `count += if regex(p).is_match(clean) {1} else {0}`.
10. Return `max(count, 1)`.

> ⚠️ Steps 8–9 mirror PHP `preg_match` which returns **0 or 1** (found / not
> found), *not* the number of occurrences. So each pattern adjusts the count by
> at most ±1. Do **not** use `find_iter().count()`.

**`count_vowel_groups(clean)`**:
- `"individual"`: number of characters in `clean` matching `[vowel_chars]`
  (i.e. `regex("[{vowel_chars}]").find_iter(clean).count()`).
- `"cluster"` (default): split `clean` by `[^{vowel_chars}]+` and count non-empty
  pieces (== number of maximal vowel runs).

Example: `дыхание` with Russian vowels → cluster `3` (`ы,а,ие`), individual `4`
(`ы,а,и,е`). Russian/Ukrainian/Belarusian/Bulgarian ship `vowelMode:"individual"`.

**`split_syllables(word)`** — arithmetic (approximate) split into
`count_syllables(word)` roughly-equal char chunks:
- `count = count_syllables(word)`; if `count <= 1``[word]` (or `[]` if empty).
- `len = word.chars().count()`; if `count >= len` ⇒ one part per char.
- else split into `count` chunks: base `len/count`, the first `len % count` chunks
  get one extra char. Clamp the last chunk to not overrun. Use `char`-based
  substrings.

**`has_rules()`** = config present AND (`problem_words` OR `subtract_patterns` OR
`add_patterns` OR `prefixes` OR `suffixes` non-empty). Used by composite.
**`has_word(w)`** = `problem_words` contains `w.trim().to_lowercase()`.

### 6.3 `CompositeSyllableCounter` (`CompositeSyllableCounter.php`)
Holds an ordered `chain: Vec<Box<dyn SyllableCounter>>` (built as `[heuristic, tex]`).
For both `count_syllables` and `split_syllables`, iterate the chain:
- If the element is a `HeuristicSyllableCounter`: if `has_rules()` ⇒ use it and
  **return**; else skip to next.
- Otherwise ⇒ use it and return.
- Fallback: last element, else `1` / `[word]`.

> Consequence to internalize: when the heuristic has rules (e.g. `en-us`, which
> ships a big `problemWords` list), **composite behaves as pure heuristic** for
> `count`/`split_syllables`. TeX is still used by `split_word` (which bypasses the
> counter and calls the hyphenator directly). Implement the trait-object downcast
> via an `enum SyllableCounterKind { Tex(..), Heuristic(..) }` chain rather than
> `dyn` + `Any` if you prefer — the observable behaviour is what matters.

### 6.4 Engine wiring (`Engine::loadSyllableCounter`)
```
tex = TexSyllableCounter(hyphenator)
mode = language.syllable_mode
if mode == "tex" || language.syllable_heuristics.is_none() { return tex }
heuristic = HeuristicSyllableCounter(language.syllable_heuristics)
if mode == "heuristic" { return heuristic }
return Composite([heuristic, tex])         // mode == "composite" (or anything else)
```

---

## 7. Text layer (`src/Text/`)

### 7.1 `TextSplitter` (`TextSplitter.php`)
Built from a `Language`. Compile the three regexes once (with Unicode):
- words: `Regex::new(&language.word_split_pattern)` used as a **splitter**.
- sentences / letters: `Regex::new(&format!("{}", language.sentence_boundary_pattern))`
  and `Regex::new(&language.letter_pattern)` (PHP wraps them as `/pat/u`).

Methods:
- `split_words(text)`: `trim`; if empty ⇒ `[]`. Split text by the word regex,
  **filter out empty strings**, collect. (PHP uses `mb_split`; the `regex` crate's
  `Regex::split` is the analogue — remember to drop empties, incl. leading/trailing.)
- `split_sentences(text)`: `trim`; split by sentence-boundary regex with "no empty"
  semantics, then `trim` each piece. (Only used indirectly; keep for parity.)
- `count_letters(text)`: `trim`; if empty ⇒ 0; else number of matches of the
  letter regex (`find_iter().count()`).
- `count_words(text)`: `split_words(text).len()`.
- `count_sentences(text)`: `trim`; if empty ⇒ 0; else `m = matches of sentence
  regex`; return `if m == 0 {1} else {m}`.
- `count_long_words(text, threshold)`: number of words whose `count_letters(word)
  > threshold`.

### 7.2 `TextStatistics` (`TextStatistics.php`) — plain data struct
```rust
pub struct TextStatistics {
    pub letter_count: i64,
    pub word_count: i64,
    pub sentence_count: i64,
    pub syllable_count: i64,
    pub polysyllable_count: i64,
    pub average_syllables_per_word: f64,
    pub average_words_per_sentence: f64,
    pub long_word_count: i64,
    pub syllable_histogram: BTreeMap<i64, i64>, // syllables -> #words, ascending keys
}
```

### 7.3 `TextAnalyzer` (`TextAnalyzer.php`)
Holds `hyphenator`, `syllable_counter`, `text_splitter`, `language`.

- `split_word(w)``hyphenator.hyphenate(w)`.
- `split_syllables(w)``syllable_counter.split_syllables(w)`.
- `syllable_count(w)``syllable_counter.count_syllables(w)`.
- `word_count / sentence_count / letter_count` → delegate to splitter.
- `total_syllables(text)` → sum of `count_syllables` over `split_words`.
- `average_syllables_per_word(text)``words = split_words`; if empty ⇒ `0.0`;
  else `sum(count_syllables)/words.len()`.
- `average_words_per_sentence(text)``w = count_words`, `s = count_sentences`;
  if `s == 0``w as f64`; else `w / s`.
- `words_with_more_than_n_syllables(text, n, count_proper_nouns=true)`: over
  `split_words`, count words with `count_syllables > n`; when `count_proper_nouns
  == false`, only count a word if its first char is **not** uppercase
  (`first_char != first_char.to_uppercase()`).
- `polysyllable_count(text, count_proper_nouns=true)` = `words_with_more_than_n_syllables(text, 2, …)`.
- `histogram_syllables(text)`: over words, `s = count_syllables`; skip `s == 0`;
  `histogram[s] += 1`; return ascending-key map.
- **`analyze(text) -> Result<TextStatistics, Error>`** (this is the hot path;
  reproduce exactly):
  1. `text = text.trim()`.
  2. `words = split_words(text)`; `word_count = words.len()`; if `0``Err(EmptyText)`.
  3. `letter_count = count_letters(text)`; `sentence_count = count_sentences(text)`.
  4. Loop words: `s = count_syllables(word)`; `total_syllables += s`; if `s > 2`
     `polysyllable_count += 1`; if `s > 0``histogram[s] += 1`.
  5. `sentence_for_avg = if sentence_count == 0 {1} else {sentence_count}`.
  6. LIX threshold: read `language.get_formula_config("lix")["longWordThreshold"]`
     if numeric else `6`; `long_word_count = count_long_words(text, threshold)`.
  7. Build `TextStatistics { …, average_syllables_per_word: total_syllables/word_count,
     average_words_per_sentence: word_count/sentence_for_avg, … }` (float division).
- `add_hyphenations(map)`: forward to the hyphenator if it is a `LiangHyphenator`.

> Note: `analyze` computes `polysyllable_count` as **> 2** syllables. So does
> `words_with_more_than_n_syllables(_,2)`. Keep both consistent.

### 7.4 Caching (optional)
PHP caches parsed patterns as JSON (`Cache/JsonPatternCache.php`, version `"2.0"`,
file `syllable.{code}.json`). In Rust this is a **pure performance optimization**
and may be omitted for v1. If you implement it, keep the same on-disk JSON shape
for interop:
```json
{ "version":"2.0",
  "patterns":[{"chars":["a","c","h"],"weights":[0,2,0,0]}, …],
  "exceptions":{"word":"hy-phen-ated", …},
  "maxPatternLength": <int> }
```
A better Rust-native approach: precompile all 86 pattern sets at build time in
`build.rs` into a static structure. Optional; not required for correctness.

---

## 8. Error type (`src/Exception/`)
PHP has an abstract `ReadabilityEngineException` and concrete subclasses. Model as
one enum:
```rust
#[derive(Debug)]
pub enum Error {
    UnsupportedLanguage(String),                 // withCode
    UnsupportedFormula { formula: String, language: String },
    EmptyText,
    PatternFileNotFound(String),
    PatternParse { token: String, line: usize, file: String },
    Io(std::io::Error),
    Json(serde_json::Error),
}
```
Match the message formats where practical, e.g. UnsupportedFormula:
`"Formula \"{formula}\" is not supported for language \"{language}\"."`

---

## 9. Formula layer (`src/Formula/`) — the numeric heart

### 9.1 Contracts
```rust
pub struct FormulaResult {
    pub formula_name: String,
    pub language_code: String,
    pub score: f64,
    pub grade_level: Option<f64>,
    pub interpretation: String,
    pub inputs: BTreeMap<String, f64>,   // debug values; ints stored as f64
}

pub trait Formula {
    fn name(&self) -> &'static str;
    fn description(&self) -> &'static str;
    fn supported_languages(&self) -> &'static [&'static str]; // ["*"] == all
    fn calculate(&self, stats: &TextStatistics, language: &Language) -> FormulaResult;
}
```

### 9.2 `FormulaRegistry` (`FormulaRegistry.php` + `FormulaRegistryFactory.php`)
Register all 17 (order below is the PHP factory order; it affects
`list_names()`/`list_for_language()` order, which some tests check):
`AutomatedReadabilityIndex, ColemanLiau, Crawford, DaleChall, FernandezHuerta,
FleschKincaidGradeLevel, FleschReadingEase, FogPL, Gulpease, GunningFog,
GutierrezPolini, Lix, Osman, SmogIndex, Spache, SzigrisztPazos, WienerSachtextformel`.

- `list_for_language(lang)`: for each formula in registration order, include if
  `supported_languages() == ["*"]` or contains `lang.code`.
- `calculate(name, lang, stats)`: look up; if missing or not supported for lang ⇒
  `Err(UnsupportedFormula)`; else `formula.calculate(stats, lang)`.

### 9.3 Shared helpers
- **`GradeLevelInterpretation::for_score(score)`** (`grade_level.rs`): match on
  ascending thresholds (`<=`):
  `≤1 Kindergarten, ≤2 1st Grade, ≤3 2nd Grade, ≤4 3rd Grade, ≤5 4th Grade,
  ≤6 5th Grade, ≤7 6th Grade, ≤8 7th Grade, ≤9 8th Grade, ≤10 9th Grade,
  ≤11 10th Grade, ≤12 11th Grade, ≤13 12th Grade, ≤16 College, else Graduate`.
- **`TextStatisticsHelper::estimate_difficult_percentage(stats)`** (`helper.rs`):
  if `word_count == 0``0.0`; `easy = histogram.get(&1).copied().unwrap_or(0)`;
  `difficult = max(word_count - easy, 0)`; return `difficult/word_count * 100.0`.

### 9.4 Rounding — **read this twice**
PHP `round($x, n)` and `round($x)` use **round-half-away-from-zero**. Rust
`f64::round()` is *also* half-away-from-zero, so:
```rust
fn round_to(x: f64, places: i32) -> f64 {
    let f = 10f64.powi(places);
    (x * f).round() / f
}
```
Use this for every `round($x, n)`. `round($x, 1)` ⇒ `round_to(x,1)`,
`round($x, 2)` ⇒ `round_to(x,2)`, `round($x, 4)` ⇒ `round_to(x,4)`.

**Where interpretation reads the raw vs rounded score:** in almost every formula,
`interpret()` receives the **raw (unrounded)** score while the `score` field is the
rounded value. The **one exception is Szigriszt-Pazos**, which rounds first and
then interprets the rounded value. Reproduce each formula's exact call.

`grade_level` clamps happen **after** rounding: `min(max(round_to(score,1), lo), hi)`.

### 9.5 The 17 formulas — exact definitions

Let `ASL = stats.average_words_per_sentence`, `ASW = stats.average_syllables_per_word`,
`W = max(word_count,1)`, `S = max(sentence_count,1)`, `L = letter_count`,
`poly = polysyllable_count`.

**Universal (`supported_languages() == ["*"]`):**

1. **Gunning Fog** (`gunning_fog`)
   - `poly_pct = if word_count>0 { poly/word_count*100 } else {0}`
   - `score = 0.4 * (ASL + poly_pct)`
   - `grade_level = clamp(round1(score), 0, 19)`; `interpret(raw)`:
     `<6 Very Easy, <8 Easy, <12 Standard, <14 Hard, <17 Very Hard, else Extremely Hard`
   - inputs: `asl, polysyllablePct, polysyllableCount, wordCount`

2. **SMOG** (`smog`)
   - `score = 1.0430 * sqrt(poly * (30.0 / S)) + 3.1291`
   - `grade_level = clamp(round1, 0, 18)`; `interpret = GradeLevelInterpretation::for_score(raw)`
   - inputs: `polysyllableCount, sentenceCount`

3. **Coleman-Liau** (`coleman_liau`)
   - `Lv = (L / W) * 100`, `Sv = (S / W) * 100`
   - `score = 0.0588*Lv - 0.296*Sv - 15.8`
   - `grade_level = clamp(round1, 0, 18)`; `interpret = for_score(raw)`
   - inputs: `L=round2(Lv), S=round2(Sv), letterCount, wordCount, sentenceCount`

4. **ARI** (`ari`)
   - `score = 4.71*(L/W) + 0.5*(W/S) - 21.43`
   - `grade_level = clamp(round1, 0, 18)`; `interpret = for_score(raw)`
   - inputs: `charsPerWord=round2(L/W), wordsPerSentence=round2(W/S)`

5. **LIX** (`lix`)
   - `threshold` = lix config `longWordThreshold` if numeric else 6 (only echoed in inputs)
   - `long_pct = if word_count>0 { long_word_count/word_count*100 } else {0}`
   - `score = ASL + long_pct`; **round to 2 places**; `grade_level = None`
   - `interpret(raw)`: `<25 Children's Books, <30 Simple Texts, <40 Normal / Fiction,
     <50 Factual Information, <60 Specialized Texts, else Research / Advanced`
   - inputs: `asl, longWordPct=round2, threshold, longWordCount, wordCount`

**Language-specific** (`supported_languages()` lists exact codes):

6. **Flesch Reading Ease** (`flesch_reading_ease`) — langs: `en-us, en-gb, de-1996,
   de-1901, de-ch-1901, ru, es, it, fr, nl, pt, tr`
   - coefficients from lang JSON `flesch_reading_ease`: `base` (def 206.835),
     `aslMult` (def 1.015), `aswMult` (def 84.6)
   - `score = base - aslMult*ASL - aswMult*ASW`; round1; `grade_level = None`
   - `interpret(raw)`: `≥90 Very Easy, ≥80 Easy, ≥70 Fairly Easy, ≥60 Standard,
     ≥50 Fairly Hard, ≥30 Hard, else Very Hard`
   - inputs: `asl=ASL, asw=ASW`

7. **Flesch-Kincaid Grade Level** (`flesch_kincaid_grade_level`) — same lang list
   - `score = 0.39*ASL + 11.8*ASW - 15.59`; round1;
     `grade_level = clamp(round1, 0, 18)`
   - `interpret(raw)`: `≤1 1st Grade, ≤2 2nd Grade, …, ≤12 12th Grade, ≤16 College,
     else Graduate` (its own table — 1st..12th, no Kindergarten)
   - inputs: `asl, asw`

8. **Wiener Sachtextformel** (`wiener_sachtextformel`) — langs: `de-1996, de-1901, de-ch-1901`
   - `ms = poly/W * 100`, `sl = ASL`, `iw = if word_count>0 { long_word_count/word_count*100 } else {0}`
     (⚠ the `threshold=6` argument in PHP is **ignored**; `iw` uses the already-computed
     `long_word_count`, which used the language's LIX threshold), `es = one-syllable %
     = histogram[1]/W*100`.
   - Variants (default variant 1):
     - v1: `0.1935*ms + 0.1672*sl + 0.1297*iw - 0.0327*es - 0.875`
     - v2: `0.2007*ms + 0.1682*sl + 0.1373*iw - 2.779`
     - v3: `0.2963*ms + 0.1905*sl - 1.1144`
     - v4: `0.2744*ms + 0.2656*sl - 1.693`
     - other ⇒ error
   - `score = round1`; `grade_level = clamp(raw_score, 4, 15)` (**clamp uses raw, not rounded**);
     `formula_name = "wiener_sachtextformel_{variant}"`; `interpret(raw)`:
     `<5 Very Easy, <7 Easy, <9 Standard, <11 Fairly Hard, <13 Hard, else Very Hard`
   - inputs: `ms, sl, iw, es, variant`
   - Expose `calculate_variant(stats, lang, variant)`; default `calculate` uses variant 1.

9. **Gulpease** (`gulpease`) — lang: `it`
   - `score = 89.0 + (300.0*sentence_count - 10.0*letter_count) / W`; round1;
     `grade_level = None`
   - `interpret(raw)`: `≥80 Easy for elementary school, ≥60 Easy for middle school,
     ≥40 Easy for high school, else Difficult for high school`
   - inputs: `letterCount, wordCount, sentenceCount`

10. **Fernández-Huerta** (`fernandez_huerta`) — lang: `es`
    - `score = 206.84 - 1.02*ASL - 60.0*ASW`; round1; `grade_level = None`
    - `interpret(raw)`: `≥90 Very Easy, ≥80 Easy, ≥70 Fairly Easy, ≥60 Standard,
      ≥50 Fairly Difficult, ≥30 Difficult, else Very Difficult`
    - inputs: `asl, asw`

11. **Szigriszt-Pazos** (`szigriszt_pazos`) — lang: `es`
    - `spw = syllable_count / W`; `syllables_per_100 = round1(spw*100)`;
      `score = round1(206.835 - 62.3*spw - ASL)`; `grade_level = None`
    - **`interpret(rounded score)`** (the exception): `≥85 Very Easy, ≥75 Easy,
      ≥65 Fairly Easy, ≥55 Standard, ≥40 Fairly Difficult, ≥30 Difficult, else Very Difficult`
    - inputs: `syllablesPer100=syllables_per_100, wordsPerSentence=round1(ASL)`

12. **Gutiérrez-Polini** (`gutierrez_polini`) — lang: `es`
    - `score = 95.2 - 9.7*(L/W) - 0.35*ASL`; round1; `grade_level = None`
    - `interpret(raw)`: `≥80 Very Easy, ≥70 Easy, ≥50 Standard, ≥30 Difficult, else Very Difficult`
    - inputs: `lettersPerWord=round2(L/W), wordsPerSentence=round2(ASL)`

13. **Crawford** (`crawford`) — lang: `es`
    - `avg_letters = L/W`, `sent_per_100 = (S/W)*100`
    - `score = -0.205*avg_letters + 0.049*sent_per_100 - 3.407`; round1; `grade_level = None`
    - `interpret(raw)`: `≥9 Very Easy, ≥7 Easy, ≥5 Standard, ≥3 Difficult, else Very Difficult`
    - inputs: `avgLettersPerWord=round2, sentencesPer100Words=round2`

14. **FOG-PL** (`fog_pl`) — lang: `pl`
    - `hard_pct = poly/W*100`, `asl = W/S`
    - `score = 0.4*(asl + hard_pct)`; round1; `grade_level = clamp(round1, 0, 19)`
    - `interpret(raw)`: `<6 Very Easy, <8 Easy, <12 Standard, <14 Hard, else Very Hard`
    - inputs: `asl=round2, hardWordsPct=round2, polysyllableCount`

15. **Dale-Chall** (`dale_chall`) — langs: `en-us, en-gb`
    - `diff = estimate_difficult_percentage(stats)`
    - `raw = 0.1579*diff + 0.0496*ASL`; `adjusted = if diff>5.0 { raw+3.6365 } else { raw }`
    - `score = round1(adjusted)`; `grade_level = None`; `interpret(adjusted /*raw*/)`:
      `≤4.9 4th grade or below, ≤5.9 5th-6th grade, ≤6.9 7th-8th grade,
      ≤7.9 9th-10th grade, ≤8.9 11th-12th grade, ≤9.9 College, else Graduate`
    - inputs: `difficultWordPct=round1(diff), rawScore=round4(raw), averageWordsPerSentence=ASL`

16. **Spache** (`spache`) — langs: `en-us, en-gb`
    - `diff = estimate_difficult_percentage(stats)`
    - `score = 0.121*ASL + 0.082*diff + 0.659`; round1;
      `grade_level = clamp(round1, 0, 5)`
    - `interpret(raw)`: `≤2 1st Grade, ≤2.5 2nd Grade, ≤3 3rd Grade, ≤3.5 4th Grade,
      else Above 4th Grade`
    - inputs: `averageWordsPerSentence=ASL, difficultWordPct=round2(diff)`

17. **OSMAN** (`osman`) — lang: `ar`
    - `asl = W/S`, `avg_letters = L/W`, `hard_pct = poly/W*100`
    - `score = 200.0 - 2.0*asl - 1.5*avg_letters - 0.4*hard_pct`; round1; `grade_level = None`
    - `interpret(raw)`: `≥90 Very Easy, ≥70 Easy, ≥50 Standard, ≥30 Difficult, else Very Difficult`
    - inputs: `asl=round2, avgLetters=round2, hardWordsPct=round2`

---

## 10. Public API — `ReadSight` facade (`src/Engine.php`)

Name the struct `ReadSight` (matches Python) or `Engine` (matches PHP) — expose
both via `pub use`. Constructor mirrors PHP `Engine::__construct`:

```rust
impl ReadSight {
    pub fn new(language: &str) -> Result<Self, Error>;                 // embedded data
    pub fn with_config(language: &str, config: Config) -> Result<Self, Error>;

    // static
    pub fn supported_languages(config: Option<&Config>) -> Vec<String>; // sorted codes

    // accessors
    pub fn language(&self) -> &Language;
    pub fn supported_formulas(&self) -> Vec<String>; // registry.list_for_language

    // text / syllable
    pub fn split_word(&self, w:&str) -> Vec<String>;
    pub fn split_syllables(&self, w:&str) -> Vec<String>;
    pub fn syllable_count(&self, w:&str) -> i64;
    pub fn word_count(&self, t:&str) -> i64;
    pub fn sentence_count(&self, t:&str) -> i64;
    pub fn letter_count(&self, t:&str) -> i64;
    pub fn total_syllables(&self, t:&str) -> i64;
    pub fn average_syllables_per_word(&self, t:&str) -> f64;
    pub fn average_words_per_sentence(&self, t:&str) -> f64;
    pub fn polysyllable_count(&self, t:&str, count_proper_nouns: bool) -> i64;
    pub fn words_with_more_than_n_syllables(&self, t:&str, n:i64, count_proper_nouns:bool) -> i64;
    pub fn histogram_syllables(&self, t:&str) -> BTreeMap<i64,i64>;
    pub fn analyze(&self, t:&str) -> Result<TextStatistics, Error>;
    pub fn add_hyphenations(&mut self, map: HashMap<String,String>);

    // formulas
    pub fn score(&self, formula:&str, text:&str) -> Result<FormulaResult, Error>;
    // convenience wrappers, each == self.score("<name>", text):
    pub fn flesch_reading_ease / flesch_kincaid_grade_level / gunning_fog /
        smog_index("smog") / coleman_liau / automated_readability_index("ari") / lix /
        gulpease / fernandez_huerta / szigriszt_pazos / gutierrez_polini / crawford /
        fog_pl("fog_pl") / dale_chall / spache / osman(&self, text:&str) -> Result<FormulaResult,Error>;
    pub fn wiener_sachtextformel(&self, text:&str, variant: i32 /*default 1*/) -> Result<FormulaResult,Error>;
}
```

Construction steps (`Engine::__construct`):
1. Resolve `Config` (defaults ⇒ embedded data).
2. `language = JsonLanguageRepository::find(language)`.
3. `hyphenator = load_hyphenator(language)` — from cache if present & version
   matches, else parse `.tex` via `TexSource` (and populate cache if enabled).
4. `syllable_counter = load_syllable_counter()` (§6.4).
5. `text_analyzer = TextAnalyzer::new(hyphenator, syllable_counter, TextSplitter::new(language), language)`.
6. `formula_registry = registry_factory()`.

`score(formula, text)` = `formula_registry.calculate(formula, &language, &self.analyze(text)?)`.
For `wiener_sachtextformel` route to `calculate_variant`.

> ⚠️ `analyze` errors on **empty text** (`EmptyText`). All formula convenience
> methods therefore return `Result`. Preserve that (Python raises; PHP throws).

---

## 11. Idiomatic Rust vs. literal port

Follow these substitutions faithfully; they change **behaviour**, not just style:

| PHP idiom | Rust equivalent | Note |
|---|---|---|
| `mb_strlen($s)` | `s.chars().count()` | code points, not bytes/graphemes |
| `mb_substr($s,$i,$n)` | `s.chars().skip(i).take(n).collect::<String>()` | code points |
| `mb_strtolower` / `mb_strtoupper` | `to_lowercase` / `to_uppercase` | Unicode |
| `preg_match('/p/u', s)` (bool) | `Regex::new("p")?.is_match(s)` | returns 0/1 in PHP → bool |
| `preg_match_all('/p/u', s)` (count) | `re.find_iter(s).count()` | |
| `mb_split(pat, s)` | `re.split(s).filter(!empty)` | filter empties |
| `preg_split('/(?<!^)(?!$)/u', s)` | `s.chars()` | split into chars |
| `is_numeric($ch)` (1 char) | `ch.is_ascii_digit()` | tokens use ASCII digits |
| `round($x, n)` | `round_to(x, n)` (§9.4) | half-away-from-zero |
| `json_decode(..., true)` | `serde_json::from_str::<Value>` / typed | |
| `array<string,int>` insertion order | `IndexMap` if order matters | `problemWords`, prefixes/suffixes iterate in file order; use `serde_json::Map` (preserves insertion order with the `preserve_order` feature) or `IndexMap` |

> **Ordering caveat:** PHP associative arrays and Python dicts preserve insertion
> order, and the heuristic counter iterates `prefixes`/`suffixes` in that order.
> Enable `serde_json`'s `preserve_order` feature (or use `indexmap`) so
> `syllableHeuristics` maps keep JSON order. For the current data this rarely
> changes results, but keep parity.

The `regex` crate does not support look-around, but **none of the language
patterns need it** (they are character classes like `[^\p{L}'’-]+`, `[.!?…]+`,
`[А-Яа-яЁёҐ-ӿЀ-ӿ]`). If a pattern ever fails to compile, compare against the raw
JSON and the Python port's handling before altering it.

---

## 12. Unicode / regex parity checklist

- Enable `regex` default features (Unicode on). `\p{L}` must match the same set
  PHP/PCRE matches for the test corpus.
- Word/sentence/letter regexes come **verbatim** from language JSON — compile the
  string body as-is (PHP adds `/…/u` delimiters; you just enable Unicode).
- Lowercasing a word before hyphenation must be locale-independent (both PHP
  `mb_strtolower` and Rust `to_lowercase` are; do not use `to_ascii_lowercase`).
- `.tex` files are UTF-8. Read as UTF-8; iterate the parser by **byte offset** for
  the `%`, `\`, `{`, `}` scanning but capture tokens as full Unicode `\S+`.

---

## 13. Verification strategy (make parity provable)

Correctness = **matching the reference outputs**, not "looks right". Build a
golden-vector harness.

### 13.1 Generate golden vectors from the references
Write a tiny script against **PHP** (canonical) and/or **Python** (`pip install
readsight`) that dumps JSON for a battery of inputs, then commit the JSON under
`tests/golden/`.

Python generator sketch (fast to run):
```python
import json
from readsight import ReadSight
CASES = {
  "en-us": ["hyphenation","banana","beautiful","communication","extraordinary","strengths"],
  "ru":    ["дыхание","беззвучная","молоко","красивый","привет"],
  "de-1996": ["Verständlichkeit","Schreiben","Deutsch"],
  # …cover every language you can, plus paragraphs for formulas
}
out = {}
for lang, words in CASES.items():
    rs = ReadSight(lang)
    out[lang] = {
      "syllable_count": {w: rs.syllable_count(w) for w in words},
      "split_word":     {w: rs.split_word(w) for w in words},
      "split_syllables":{w: rs.split_syllables(w) for w in words},
    }
json.dump(out, open("tests/golden/syllable.json","w"), ensure_ascii=False, indent=2)
```
Do the analogous dump for `analyze(text)` and every `score(formula, text)` result
(score, grade_level, interpretation) over a set of representative paragraphs per
language. Cross-check PHP vs Python first — they must already agree.

### 13.2 Port the reference test suites
The PHP tests (`tests/`, ~260 assertions) and Python tests (`tests/`, 133) encode
the intended behaviour. Re-express the important ones as Rust tests:
- **Hyphenation**: `LiangHyphenatorTest`, `PatternTest`, `PatternsCollectionTest`.
- **Syllable**: `HeuristicSyllableCounterTest` (incl. `vowelMode` cluster vs
  individual), `CompositeSyllableCounterTest`, `SyllableConsistencyTest`
  (invariants: `split_word` reconstructs the original word;
  `0 < count <= char_len`; determinism across repeated calls).
- **Text**: `TextSplitterTest`.
- **Formula**: `UniversalFormulaTest`, `LanguageSpecificFormulaTest`,
  `GradeLevelInterpretationTest`, `TextStatisticsHelperTest`.
- **Integration/Multilingual**: load every language; sample words yield `count > 0`;
  specific vectors (`en-us banana → 3`, `ru молоко → 3`, `ru дыхание → 4`).

### 13.3 Float comparison
Compare `score`/`grade_level` with an absolute tolerance of **1e-9** after your
`round_to`. If a value diverges by more than the last displayed decimal, you have
a real bug (usually rounding order or an ASL/ASW division difference) — fix the
port, don't loosen the tolerance.

### 13.4 Full-language smoke test
Iterate all 86 codes from `supported_languages`, construct an engine, run
`analyze` + every `supported_formulas` on a short sample; assert no panics and
finite numbers.

---

## 14. Recommended build order for the AI (dependency-first)

Implement and test each layer before moving on. Do **not** write all 54 files then
test — you will drown in compounding errors.

1. **Scaffold**: crate, `Config`, `DataSource::Embedded`, `Error`. Copy `data/`.
2. **Language**: `Script`, `LanguageCode`, `Language` (serde), repository.
   Test: load all 86 JSON; `supported_languages()` sorted; spot-check fields.
3. **Hyphenation**: `Pattern`, `PatternsCollection`, exceptions, `TexSource`
   parser, `LiangHyphenator`. **Gate on golden `split_word` vectors** before
   proceeding — everything downstream depends on this.
4. **Syllable**: trait, `Tex`, `Heuristic` (with `vowelMode`), `Composite`; engine
   wiring. Gate on golden `syllable_count`/`split_syllables` (esp. `en-us` composite
   and `ru/uk/be/bg` individual-mode vectors, incl. `дыхание → 4`).
5. **Text**: `TextSplitter`, `TextStatistics`, `TextAnalyzer`. Gate on golden
   `analyze` output.
6. **Formula**: helpers (`round_to`, `GradeLevelInterpretation`,
   `TextStatisticsHelper`), then the 5 universal, then the 12 language-specific,
   then registry. Gate each on golden `score` vectors.
7. **Facade** `ReadSight` + convenience methods. Gate on integration/multilingual.
8. **Optional**: JSON pattern cache, `build.rs` precompilation, a `demo` binary
   mirroring `examples/demo.php`.

At each gate: `cargo test`, `cargo clippy -- -D warnings`, `cargo fmt --check`.

---

## 15. Pitfalls checklist (the things that will bite you)

- [ ] **Rounding**: `f64::round()` is half-away-from-zero — matches PHP. Use
      `round_to`; never truncate; round *before* clamping `grade_level`.
- [ ] **Interpretation uses raw score** in every formula **except Szigriszt-Pazos**
      (rounded). WSTF `grade_level` clamps the **raw** score.
- [ ] **`preg_match` returns 0/1** — heuristic subtract/add patterns adjust count by
      at most ±1 each (`is_match`, not occurrence count).
- [ ] **Composite = heuristic when heuristic has rules** (e.g. `en-us`). TeX is only
      reached via `split_word`.
- [ ] **`vowelMode`**: `individual` counts each vowel char; `cluster` counts vowel
      runs. Default `cluster`. `ru/uk/be/bg` ship `individual`.
- [ ] **Supported-formula list** comes from `Formula::supported_languages()`, **not**
      the language JSON `formulas` block. JSON only supplies coefficients + LIX
      threshold.
- [ ] **WSTF `iw`** ignores the literal `6` and uses `long_word_count` computed with
      the language's LIX threshold in `analyze`.
- [ ] **`analyze` throws on empty text**; `polysyllable_count` = words with **>2**
      syllables; histogram excludes `0`-syllable words; keys ascending.
- [ ] **Char vs byte indexing** everywhere in hyphenation/syllable code.
- [ ] **Registration order** of formulas affects `supported_formulas()` ordering.
- [ ] **`split_by_patterns` boundary math** (`.word.`, `i-1` mapping, two trailing
      loops) — port verbatim; verify with golden vectors, don't refactor first.
- [ ] **Map insertion order** for `problemWords`/prefixes/suffixes (`preserve_order`).
- [ ] **Data is the contract**: copy `data/` verbatim; do not regenerate.
- [ ] **`wiener_sachtextformel` result name** is `wiener_sachtextformel_{variant}`.
- [ ] **Language code normalization** (trim + Unicode lowercase) before every lookup.

---

## 16. Definition of done

- All 86 languages load; `supported_languages()` returns 86 sorted codes.
- Golden vectors for syllables, `split_word`, `analyze`, and all 17 formulas match
  the PHP/Python references within 1e-9 (numbers) / exact (strings, splits).
- Ported PHP + Python test cases pass.
- `cargo test`, `cargo clippy -- -D warnings`, `cargo fmt --check` all clean.
- Public API mirrors the facade in §10; docs link back to the two reference repos.

---

*Generated as the porting spec for `readsight` (Rust). Canonical behaviour lives in
the PHP repo (https://github.com/MADEVAL/ReadSight); the Python port
(https://github.com/MADEVAL/ReadSightPy) is the secondary reference. When in doubt,
generate a golden vector from both and match it.*