sherd 0.6.0

two-axis federation of SPEC.md and source over a dir DAG, for small-context local models
Documentation
V1|-
V2|-
V3|-
V4|-
V5|-
V6|-
V7|-
V8|-
V9|-
V10|-
V11|-
V12| — insertion moves every citation below it
V13|-
V14|-
V15|-
V16|-
V17| — 15-30% off
V18|-
V19|-
V20|-
V21|-
V24|-
V25|-
V26|-
V27|-
V30|-
V34|-
V36|-
V39|-
V40|-
V42| MEASURED 3x that horizontal ⊥ the lever: itok 59% of §V bytes stay @ root · itok 66% of stmts · sherd own 64%
V43| MEASURED: rationale = 80% of `itok` §V, 73% of `microlith` ∴ vertical buys 5x vs horizontal 1.7x
V44| ∴ dropping it is the failure both sibling repos already guard
V45| entry cost is re-billed EVERY turn
V46|-
V47|-
V48| a pass on a section the parser cannot see is indistinguishable from a real pass
V49| per-file losslessness is blind to content vanishing BETWEEN files
V50|-
V51| `session/mod.rs` = 4,946 tok / 0% test = a default that grew a body
V52| a ceiling set a hair above current turns every addition into a raise & trains the reflex it exists to catch
V53| a split that lowers bytes & raises cross-module refs is a regression
V54| an agent applying Judgment blindly silences the guard
V55|-
V56|-
V57|-
V58|-
V59| the deletion commit is where nobody re-checks
V60| — 2.1x @ impl 1,118, 2.9x @ 5,218, widening w/ node size
V61|-
V62| V13 w/o a detector is a comment, ⊥ a guard
V63|-
V64| else "matches no child" ⊥ mean "⊥ in subtree" & reader ! open all of them
V65|-
V66| a positive lens decides only DESCEND; a negative one decides STOP ∴ negative space is the byte that prevents loading. also guards the "rule never carried to a sibling path" class (5 of 15 `itok` bugs)
V67| `parses SPEC.md` ⊥ answer "why are bulleted ids rejected"
V68| the row is resident @ that altitude EVERY turn; the child costs only when opened
V69| MEASURED: itok 656B/rule × 103 = 67.8KB; same COUNT @ 200B = 20.6KB ∴ per-row fatness (~5x) is the multiplier, ⊥ rule count
V70| a spec compacted to satisfy a number trains the same reflex as a ceiling raised to satisfy one
V71| reader asks "count tokens" ⊥ "itok" (V67), & a swapped dep makes a vendor name lie.
V72| VERIFIED cargo 1.96.1. guards the "rule never carried to a sibling path" class @ its source
V77| a monolith edited near its top pays full re-prefill EVERY turn (R15/R16)
V82| a guard the agent cannot SEE is B1
V83| a hand-drawn architecture diagram is a second reading of what `§F` declares — microlith's founding defect
V84|-
V100| MEASURED: structure never violated · gate evaded twice · prompt ignored entirely
V96| supervisor text in a worker prompt is wasted tokens AND instructions aimed at the wrong reader — "revert this" means nothing to a model writing one function
V97|-
V92| MEASURED: gates passed wrong code 3x (stub w/ a doc comment saying so · data-laundering `sanitize_first_cell` · wrong-field test) & a human-level reader caught all 3
V98| — a run killed between `apply` and its review leaves generated code uncommitted, which already happened once (`src/fed:B7`)
V99|-
V95|-
V93| an "infinite loop" w/o stop conditions optimizes for whatever the gate rewards
V94|-
V91|-
V74| `fed::walk` contradicted §C for a whole session & no gate could see it (B1)
V75| the local copy was 6 sections while the dep shipped 7 (B2)
V101| a sibling PATH dep is a shared working tree ∴ the gate's green is true only for the INSTANT it ran & expires silently when the sibling moves (B5).
V102| green names only what RAN ∴ an op nobody declared is invisible, ⊥ merely absent — fmt & clippy were missing for the project's whole life & every verdict looked identical (B6)
V103| a granularize-until-success loop converges on TRIVIA by construction (`src/tdd:B2` = a judge loosening from "proves the invariant" to "would compile").
V104|-
V105| `Depth::Rule` is §I's default & selected nothing for the project's whole life ∴ every pack shipped archive & every ceiling was measured against it (B8).
V106| (a) alone is what `budget` checks & is ⊥ SUFFICIENT — after T41 every chain is ≤10% of working & 0 merit wins remain.
V107| MEASURED: where the type holds the rule (`enum Verdict{Fits{slack},Over{by}}`), vague wording costs 0; `Option<u64>` holds only `can fail` & its magic number still cost 100% (R45). prose precision is the fallback for what a type cannot hold, & it is worth 3.3x (R44)
V108| the R44 → `src/tdd:T13` gap holds FIVE at once — pack size · signature given vs invented · tests hidden vs model-authored · judge in the loop · repair steps — ∴ ⊥ result across it is attributable & the honest form is one row per variable (R46). a 5-variable comparison that CONFIRMS a hypothesis is worth as little as one that refutes it
V109|-
V110| a SIBLING @ depth 2 pays root + `src` + itself & adds NOTHING to any existing chain ∴ promoting sideways LOWERS per-node cost & raises only the total-if-all-loaded, which nothing ever loads. a CHILD adds a level EVERY descendant pays EVERY turn (R6: hubs absorb 9.5%) ∴ depth is where splitting can RAISE the cost of working. — this repo can MEASURE that curve where ordinary refactoring cannot
V111| MEASURED anti-correlated (R51): 0 of 3 wrong impls caught, 7 of 30 correct ones rejected. `src/tdd` step 1 authors the test that step 3's RED gate then requires ∴ the loop's red is satisfied by a test that does ⊥ measure the invariant, & `src/tdd:B2`/`B12` are that shape reported one at a time
V112|-
V113|-
V114| pre-1.0 SemVer lets ANY minor break ∴ `0.4` vs `0.5` otherwise carries NOTHING a consumer can act on, & parity is the 1 bit the number carries FREE, before a changelog is read. ⊥ novel: Linux 2.x & GNOME shipped it ∴ a reader may already know it. `microlith` V34 states the same rule ∴ one reading serves both.
V115| — an interface section that advertises 5 absent verbs & hides 10 present ones is worse than none, ∵ a reader trusts it (B17)
V116|-
V117| the 2 halves answer different KINDS of question: whether a dir DAG parses, budgets & validates is settled by TESTS, while whether a 20B writes code that survives review is a RESEARCH result that may take months ∴ tying a release to the 2nd holds the 1st hostage, & the 1st is the half a consumer can reuse.
V118|-
V119| the 2nd file also gives `plan::structure` the `Declared` evidence it asks for: `split src/assay` \& `split src/cli` both answer "no module declarations found" ∵ nobody drew a line, & the tool refuses to guess (`src/split:V1`). ∴ the fix to a node's SIZE is the same edit as the fix to the tool being BLIND to it.
V120| MEASURED: node impl 903 (`lens`) → 13,413 (`assay`), a 14x spread, while the gated number — the CHAIN — spans 9,922 → 12,480, a 26% spread ∵ every chain pays root + parent + self ∴ node variance is DAMPED by construction. `lens` @ 903 is one of the cleanest nodes here & equalising would merge it into something. equal size is a CONSEQUENCE of equal-grained concerns, ⊥ a cause of health. — `src/tdd` 31 bug rows vs `src/code` 0 says where defects live; size alone said nothing
V121| MEASURED ceiling on the pipeline half: of 265 warnings, 111 are `too_many_lines` (a judgement over SEAMS, ⊥ a transform) & 47 arithmetic that is correct code ∴ the automatable residue is the indexing/duplication/status family, ~40-60 sites.
V122| MEASURED: 9 pedantic findings across 2 files `actionlint` had passed on EVERY commit of their lives (B28). the ledger is empty here by MEASUREMENT (`sherd` has no required status checks ∴ naming a job costs nothing), ⊥ by having no rule to write down
V73| — 30 files → 60 is ceremony
V123| hardcoding the writer ships 2 decisions as 1 ∴ neither measures alone.
V124| a FACET is a property of the FILE (`src/lens:V8`) ∴ a file mixing code & tests has none & V50 must GUESS the cut (B29).
V125| the 3 recordings in B34 were each found by an OLD OUTPUT FORMAT, ⊥ by a check ∴ the case nothing caught is the one whose format had ⊥ changed. the dep-info is the evidence cargo itself trusts, ∴ it is the 1 place the question can be answered before the binary runs
V126| owner 2026-10-24 (`paradigm/T18`, `paradigm/T20`): `.cargo/config.toml` ∵ cargo reads it in ∀ 4 build places — pre-push, CI, `buildRustPackage`, bare cargo. MEASURED (pklith #111): an unforced `"15.0"` built minos 14.0 under nix's stdenv. MEASURED here: forced `26.0` + a warm build dir → `cargo-universal` said `each minos 14.0` from binaries linked before the edit; from clean → `arm64 slice has minos 26.0` ∴ the release step empties its own dirs first. release ONLY ∵ 2 full release builds (one per Mac, the 2nd needing the other Mac's std) per push is too slow. Mac ONLY ∵ pklith's check PASSES w/ a notice off macOS (no `lipo`), right for pre-push & vacuous for a release (`.:B6`)