Rosetta convergence arc + rung-stratified standing wave + anti-eigenvalue guards (36 commits)

.claude/board/INTEGRATION_PLANS.md

@@ -1,3 +1,22 @@
+## 2026-07-26 — rosetta-codebook-convergence v1 — PROPOSED (D-RCC-1 calibrator runnable today) — main thread
+
+**Plan:** `.claude/plans/rosetta-codebook-convergence-v1.md`
+The Bible Rosetta SoA + multi-codebook qualia agreement: frozen verse address
+= external key ⇒ all translations are LANES on one row (~31k rows, in-memory);
+language = lane discriminant resolved from classid (no LanguageDto — the
+witness facet is the shipped `ClauseSignature`); codebooks never merge — the
+qualia of a word is a VECTOR OF AGREEMENTS (rank stats, i4, support-masked,
+POS-routed hydration); sense resolution = Schnittpunkt of cross-language
+sense-set intersection (extensional) × CLAM/HHTL cascade descent
+(intensional); cross-lane constraint propagation to fixpoint (facts cross,
+form never; propagated ≠ attested); Czech/Kralická as the aspect-marked
+lane; bake as PD-texts-only Release with NC treebanks as local oracle
+(libtesseract pattern). Deliverables D-RCC-1 (lanes-to-singleton
+CALIBRATOR — per-item anchors are the value, blocks nothing; operator-corrected
+from a kill gate) + D-RCC-2..8. Absorbs task-board W4/W7/W15/W17/W18;
+un-gates W11 via D-RCC-4. Blockers §4: per-treebank licence re-verify,
+Kralická edition provenance, versification map source, Greek PD edition.
+
 ## 2026-07-23 — scientific-kg-substrate v1 — PROPOSED (scoping; outward-facing crawl BLOCKED) — main thread
 
 **Plan:** `.claude/plans/scientific-kg-substrate-v1.md`

.claude/board/PR_ARC_INVENTORY.md

@@ -405,7 +405,7 @@
 
 **Added:** `.claude/knowledge/ast-as-partof-isa-address.md` corrections — Status CONJECTURE → PARTIALLY MEASURED (carrier #613/#614/#615 shipped AND the rank-minter brick-3 has RUN: `ruff_csharp_spo` harvest → `ruff_spo_address::{mint, mint_factored}` → `medcare_probe`); "The missing brick" → "The brick that ran"; Next-bricks checkmarked with the real open bricks (reroute *execution* in the mint pipeline, probe re-run blocked on harvest input data, classid Canon:Custom half-order, LSP serve end). New `.claude/knowledge/do-arm-triage-3-bucket.md` — the operator's 3-bucket DO-residue triage (fuzzy/order-varying → canonicalize-first; anticipated standard DO → ontologically-shaped landing zone as ONE DTO adapter + codebook swiss-knife `open`/`filter`/`reorder`/`apply_mask`; truly random → hand-port, recurrences graduate to bucket 2), refining OGAR's 85/15 split; records the C#/C++ DO-extractor gap (`ruff_python_dto_check` is Python-only). Lock: ogar pin `597ecb1 → a0c7936` (post-OGAR-#146). EPIPHANIES `E-BRICK3-RAN-TRUNCATION-DISALLOWED` (same-commit hygiene).
 
-**Locked:** the naive fixed-width 6-tier `(part_of:is_a)` mint is FALSIFIED at real-corpus scale (mass truncation + god-class collisions); `mint_factored` (base-255 positional path + `is_a`-from-`inherits`-only) is the corrected minter. **Truncation is DISALLOWED by policy** — bucket overflow (256-cap / 6-tier depth) is a separation-of-concerns REROUTE trigger (split the god-class or escalate a cascade level; never truncate, never field-widen — the OGAR `256-cap-is-a-lint` law #130/#140 made operational). Division of labour: `mint_factored` = addressing precision; overflow→SoC-reroute = structure. Overflow *classification* SHIPPED upstream as `ruff_spo_address::soc` (`soc_findings` → `SocVerdict`, `law_holds`; class-conditioned cascade shapes Rails `6×2` / other `4×3` / canonical GUID `3×4`, all `G·D = 12`). **Privacy split:** proprietary corpus numbers live ONLY in private MedCare-rs `.claude/archive/ruff-spo-address-medcare-probe/`; the lance-graph tree carries the design consequence, zero corpus identifiers.
+**Locked:** the naive fixed-width 6-tier `(part_of:is_a)` mint is FALSIFIED at real-corpus scale (mass truncation + god-class collisions); `mint_factored` (base-255 positional path + `is_a`-from-`inherits`-only) is the corrected minter. **Truncation is DISALLOWED by policy** — bucket overflow (256-cap / 6-tier depth) is a separation-of-concerns REROUTE trigger (split the god-class or escalate a cascade level; never truncate, never field-widen — the OGAR `256-cap-is-a-lint` law #130/#140 made operational). Division of labour: `mint_factored` = addressing precision; overflow→SoC-reroute = structure. Overflow *classification* SHIPPED upstream as `ruff_spo_address::soc` (`soc_findings` → `SocVerdict`, `law_holds`; class-conditioned cascade shapes Rails `6×2` / other `4×3` / canonical GUID `3×4`, all `G·D = 12`). **Privacy split:** the proprietary corpus numbers are recorded nowhere in this repo; the lance-graph tree carries the design consequence, zero corpus identifiers.
 
 **Deferred:** overflow-reroute *execution* inside the mint pipeline (the lint flags, a human splits); `medcare_probe` re-run on the current minter (needs corpus/NDJSON access); C#/C++ DO harvester (bucket-2 landing-zone extractor); classid Canon:Custom half-order (superseded mid-arc by the #627 canon:custom flip plan).
 
@@ -415,7 +415,7 @@
 
 **Confidence (2026-07-02):** HIGH for the knowledge content (brick-3 findings mirror the private archive's measured record; re-fetch diff against ruff `b459ec3` confirmed `lib.rs` byte-identical + `soc.rs` as the movement). Lock verified: contract `ogar_codebook` 8/8 + `lance-graph-ogar` standalone green at `a0c7936` (compile-time COUNT_FUSE holds); consumer medcare-bridge (`--features ontology`) compiles clean. NOTE: merged BEFORE #628's P1 flip — the ast-address doc's classid examples predate `ClassidOrder::CanonHigh`; interpret per the flip (canon `domain:appid` = HIGH u16).
 
-**Cross-ref:** #616/#617 (the ast-address doc lineage); #623 (OGAR sink-in plan); OGAR #145/#146 (OSINT mint + zero-rows ruling); ruff `b459ec3` (`ruff_spo_address` + `soc`); `E-CODEBOOK-MINT-IS-A-CROSS-REPO-ARC`; MedCare-rs private archive (RESTORE-STATUS re-fetch log).
+**Cross-ref:** #616/#617 (the ast-address doc lineage); #623 (OGAR sink-in plan); OGAR #145/#146 (OSINT mint + zero-rows ruling); ruff `b459ec3` (`ruff_spo_address` + `soc`); `E-CODEBOOK-MINT-IS-A-CROSS-REPO-ARC`.
 
 **Correction (2026-07-02, from the OGAR+ruff review fan-out):** the Locked line's "`mint_factored` … is the corrected minter" overstates its location — on public ruff main only `mint` + `soc.rs` exist; `mint_factored`+`RadixCodebook` are stranded on branch `claude/medcare-ruff-csharp-sync-4iahey` (`505fdc4`), which predates and LACKS `soc.rs`. The corrected-minter state exists only as the union of two diverged branches. See `E-BRICK3-CORRECTION-MINT-FACTORED-IS-SPLIT-BRAIN`.
 

.claude/board/STATUS_BOARD.md

@@ -1,3 +1,19 @@
+## rosetta-codebook-convergence-v1 — Bible Rosetta SoA + qualia agreement (ACTIVE)
+
+Plan: `.claude/plans/rosetta-codebook-convergence-v1.md` (operator convergence arc, 3 same-day corrections baked). Absorbs W4/W15/W17/W18.
+
+| D-id | Deliverable | Repo | Status | Evidence |
+|---|---|---|---|---|
+| D-RCC-1 | lanes-to-singleton probe (calibrator) | lance-graph | **v1 RUN** (`build_rosetta_probe.py`; 4 PD lanes kjv/luther1545/elberfelder1905/bkr; census 31,103 union / 31,097 common; swallow+grape receipts incl. LIVE Ps-84 versification-offset + two-lane rescue; en→de split census 48.9% of 3,071 mid-freq words; `tongue→Zunge/Sprache` real sense split) | `E-RCC-1-FOUR-LANES-ONE-KEY-1`; report in local `out/` |
+| D-RCC-2 | Rosetta SoA shape (contract) | lance-graph | Queued | plan §2 |
+| D-RCC-3 | corpus-derived word alignment | lance-graph | **In flight** (`build_alignment.py`; PMI v0 proven inside D-RCC-1; also the successor to the failed monolingual closed-class detector — labels TRANSFER through alignment) | plan §2 |
+| D-RCC-4 | qualia-agreement vector (POS-routed) | lance-graph | Queued — un-gates W11 | plan §2 |
+| D-RCC-5 | CLAM/WordNet probe + CHAODA lane-anomaly read | lance-graph | Queued (taxonomic arm runnable now) | plan §2 |
+| D-RCC-6 | cross-lane constraint propagation to fixpoint | lance-graph | Queued | plan §2 |
+| D-RCC-7 | Czech lane | lance-graph | **Data landed** (bkr lane fetched + in census) | plan §2 |
+| D-RCC-9 | **Greek SOURCE lane (new)** | lance-graph | **Shipped** — Tischendorf 8th ed., stated `Public Domain`; TR + WH are both CC BY-NC-SA (the age-implies-PD assumption is false). 27 books / 7,895 verses; 62 KJV verses `TextAbsent`. Unblocks *source outranks translation*. | `E-PD-GREEK-LANE-ACQUIRED-TISCHENDORF-1`; `fetch_greek_lane.py` |
+| D-RCC-8 | Rosetta package Release | lance-graph | Licence blocker §4.4 **DISCHARGED** (PD Greek acquired); remaining: per-treebank re-verify, Kralická edition provenance, versification map source | plan §4 |
+
 ## dialectic-engine-v1 — the reasoning cathedral (ACTIVE)
 
 Plan: `.claude/plans/dialectic-engine-v1.md` (six operator pillars + S1-S12 synthesis). V0-V1 SHIPPED; V2-V5 queued.

.claude/board/TECH_DEBT.md

@@ -3189,3 +3189,50 @@ bijection needs re-seeding. Pair: D-IDENTITY-4.
 - **TD-CI-EXCLUDED-FUSE (F5):** main's CI never compiles the workspace-`exclude`d `lance-graph-ogar`, so its compile-time `COUNT_FUSE` fires only in *consumers'* builds (medcare hit E0080 twice). Fix: one CI job `cargo check`-ing the excluded crate against OGAR main. Effort S.
 - **TD-OGAR-LOCK-UNDECIDED (F3):** `lance-graph-ogar/Cargo.lock` is gitignored → fresh checkouts build the fuse against OGAR HEAD (floating canary) while the workspace lock pins. Decide canary-vs-pin; document in one sentence. Effort S.
 - **TD-BOARD-PREPEND-CONFLICTS (F6):** at fleet cadence the append-only board files (EPIPHANIES/LATEST_STATE/PR_ARC) are the only recurring rebase-conflict cost (~30-60 min/day/session). Per-entry board files + generated index make the conflict structurally impossible. Council-sized; forwarded to the V3/coordination session.
+
+## TD-COCA-LEXICON-RANK-UNRELIABLE-AT-HEAD (2026-07-26)
+
+`crates/lance-graph-planner/examples/data/coca/lexicon.tsv` (20,004 rows) is a
+PRE-FILTERED list, not raw COCA rank 1..N, and its `rank` column is wrong in the
+high-frequency band: `the` = 5645 (pos `i`), `and` = 3584 (pos `r`/adverb — also
+a wrong tag), while `of` = 5. Pure function words `a, I, you, he, she, they, not,
+what, which` are absent entirely.
+
+Impact: any consumer treating `rank` as a frequency proxy gets a scrambled head of
+the distribution; any coverage/decile analysis silently excludes function words.
+Surfaced by `E-COVERAGE-INVERSION-CLAIM-REFUTED-1` (the decile analysis is sound
+for content vocabulary only, for this reason).
+
+Fix options: (a) re-derive rank from a real COCA frequency list; (b) rename the
+column `list_order` and add a MANIFEST note that it is not a frequency rank;
+(c) supersede with the verse-attested lane codebooks (`rosetta/build_lane_codebooks.py`),
+whose `freq`/`rank`/`dispersion` ARE corpus-measured — at the cost of Bible-domain bias.
+
+## TD-COCA-LEXICON-POS-UNUSABLE-FOR-FUNCTION-WORDS (2026-07-26)
+
+Extends `TD-COCA-LEXICON-RANK-UNRELIABLE-AT-HEAD`: the same file's **`pos`
+column is wrong for exactly the words a POS column is most needed for.**
+Spot-checked on the main thread:
+
+| word | recorded pos | correct |
+|---|---|---|
+| `the` | `i` (preposition) | determiner |
+| `and` | `r` (adverb) | coordinating conjunction |
+| `it` | `n` (**noun**) | pronoun |
+| `that` | `r` (adverb) | determiner/conjunction/pronoun |
+| `of` | `i` (preposition) | ✓ correct |
+| `a` | ABSENT | determiner |
+
+3 of 5 function words checked are mis-tagged and 1 is missing; the one
+correct tag is the one that is unambiguously a preposition.
+
+Impact: any consumer treating this column as POS ground truth for
+closed-class vocabulary is reading noise. Surfaced by the closed-class
+transfer work (`E-BOTH-CLOSED-CLASS-METHODS-LOSE-TO-RANK-150-1`), which had
+to substitute a curated word list. **Do not use `coca/lexicon.tsv` `pos` for
+function words.** For open-class content words it was not audited — unknown,
+not vindicated.
+
+Fix options: (a) re-derive POS from a real tagged corpus (the German lane
+already uses UD-derived tags and they are sound); (b) rename the column to
+mark it unreliable; (c) supersede with UD-derived English tags.

.claude/board/exec-runs/aligner-lemma-key.txt

@@ -0,0 +1,117 @@
+Task #34 — shared lemma key for build_alignment.py (D-RCC-3 aligner)
+Agent: Sonnet grindwork exec. File owned/touched: crates/lance-graph-planner/examples/data/rosetta/build_alignment.py ONLY.
+No other rosetta/*.py touched. No commit/push performed (per brief).
+
+WHAT WAS BUILT
+- Added `normalize_de`/`DE_SUFFIXES` (copied with attribution from
+  build_rosetta_probe.py, not imported — that module is a separate
+  deliverable another agent owns concurrently) and a NEW `normalize_en`/
+  no-EN_SUFFIXES-table-needed (hand-written suffix cascade: -ies->y, -eth/
+  -est/-ing, -ed, -es (orthography-aware: sibilant stem -> full -es strip,
+  else keep the silent e and strip only -s), plain -s), both min-stem-len=4
+  guarded, both explicitly NOT lemmatisers (no dict, no irregulars, no
+  ablaut/umlaut, no compounds).
+- New CLI flag `--lemma-key`, default OFF. Verified: with or without the
+  flag, `alignment_en-de.tsv` and `alignment_en-el.tsv` (the files the
+  closed_class_transfer.py consumer reads) are byte-identical (md5 diff
+  empty both before and after the normalizer bugfix). When the flag IS
+  passed, an ADDITIONAL `alignment_<pair>_lemmakey.tsv` is written (new
+  file, same 5 columns, nothing renamed/reordered) and the report gains a
+  "Lemma-key pass" section per pair.
+- Fixed a bug I introduced myself before reporting: initial normalize_en
+  stripped "grapes" -> "grap" (wrong; -es branch always chopped 2 chars).
+  Fixed with an orthography-aware branch (sibilant-final stem -> chop -es;
+  else chop only trailing -s so grapes->grape, tongues->tongue). Verified
+  with unit spot-checks before regenerating the real report.
+- Data for real run: bible_{kjv,luther1545,tischendorf}.json are
+  gitignored + were not present in the fresh checkout; found cached copies
+  in this session's scratchpad (left over from the prior D-RCC-3 session)
+  and copied them into the rosetta dir (also gitignored, `git status`
+  confirms zero untracked noise) to run the REAL aligner end-to-end rather
+  than reasoning about it unverified.
+
+MEASURED RESULTS (real KJV/Luther1545/Tischendorf corpus, MIN_COOC=5, topk=3)
+
+en-de (12,266 src tokens, 28,636 shared verses):
+  overall coverage PMI: raw 39.2% (4806/12266) -> lemma-key 43.0% (4103/9545)
+  vocab: 12266 distinct src tokens (raw) -> 9545 (lemma-key), i.e. folding
+  did fire (2721 keys merged away).
+  band coverage (own-vocab, raw -> lemma-key):
+    hapax(1):   3872 tok, 0.0%   -> 2845 tok, 0.0%   (hard floor unmoved)
+    rare(2-4):  3286 tok, 0.0%   -> 2415 tok, 0.0%   (hard floor unmoved)
+    low(5-19):  2943 tok, 89.7%  -> 2285 tok, 92.0%
+    mid(20-99): 1518 tok, 100%   -> 1335 tok, 100%
+    high(100+):  647 tok, 100%   ->  665 tok, 100%
+  LOW-BAND LIFT (direct flip measurement: raw-unaligned token, does its
+  normalised key become aligned?):
+    hapax:  3872 considered, 581 flipped  = 15.0%
+    rare:   3286 considered, 866 flipped  = 26.4%
+    low:     302 considered, 165 flipped  = 54.6%
+    ALL THREE: 7460 considered, 1612 flipped = 21.6%
+  -> normalisation DOES lift low-frequency coverage via merging, and it is
+     quantified directly (not inferred from two independently-banded
+     tables): ~1 in 5 raw-unaligned hapax/rare/low tokens gains an aligned
+     target purely from folding onto a normalised key that already cleared
+     cooc>=5 elsewhere.
+
+en-el (5,960 src tokens, 7,895 shared verses; NO German normalizer applies
+here — Greek target side has no normalizer, English source side does):
+  overall coverage PMI: raw 30.7% (1831/5960) -> lemma-key 36.3% (1679/4620)
+  LOW-BAND LIFT: hapax 12.9%, rare 25.6%, low 36.9%, ALL THREE 20.4%
+  (independent corpus, same order of magnitude as en-de -- the lift is not
+  an en-de-specific artifact).
+
+ANCHOR RECEIPTS — the falsifier, both ways (en-de pair)
+  grape:
+    raw (8 verses):    die(cooc=5,.76); der(cooc=5,.54); und(cooc=8,.37)   <- stopword noise
+    lemma (41 verses): herling(cooc=5,9.45); traub(cooc=14,9.09); beer(cooc=5,8.77)
+    -> YES, grape now finds trauben (traub, unripe-grape herling, beere/berry)
+       instead of stopwords. The predicted fix landed and is measured, not
+       asserted.
+  swallow:
+    raw (17 v):    verschlingen(7,10.36); ihr(5,1.31); nicht(7,1.05)
+    lemma (41 v):  verschlang(6,8.86); verschl(6,8.86); verschling(9,8.53)
+    -> also improved: all three lemma-key targets are the swallow-up verb
+       stem across its inflected forms, replacing 2 of 3 stopword hits.
+  vineyard:
+    raw (58 v):    weinberges(7,8.95); weinberg(36,8.83); weingärtnern(7,8.75)
+    lemma (100 v): weinberg(97,8.08); jesreelit(5,7.48); weingärtn(10,7.02)
+    -> stays correct both ways (was never broken); coverage/verse-count
+       grows via merging plural forms.
+  tongue — THE REGRESSION-CHECK ANCHOR:
+    raw (96 v):    zunge(57,8.08); zungen(14,6.78); sprache(7,6.44)
+    lemma (126 v): zung(88,7.63); lipp(8,5.00); schweig(5,4.79)
+    DID NOT SURVIVE: sprach(e)'s sense drops out of top-3 under lemma-key.
+    ROOT CAUSE (diagnosed, NOT tuned away): the borrowed DE_SUFFIXES
+    "chen" entry (meant for diminutives, e.g. Mädchen) spuriously matches
+    the tail of "sprachen" (spra+chen), folding plural "sprachen" -> "spra"
+    while singular "sprache" folds to "sprach" via the separate bare-"e"
+    suffix rule. Two forms of ONE word land on TWO different normalised
+    keys, so their evidence never merges, and both individually rank below
+    newly-boosted associates (lipp, schweig, falsch) that gained mass from
+    unrelated folding. This is a suffix-table collision in the REUSED
+    German table, not a bug in the new English-side logic. Per the brief's
+    explicit instruction, this was reported as-is and NOT hand-tuned to
+    make the anchor look green again.
+
+SO: lemma-key is a real, measured net positive on coverage (+3.8pp en-de
+overall, +5.6pp en-el overall, ~21% low-band flip rate both corpora,
+grape/swallow both fixed) but is NOT unconditionally safe — the known-good
+tongue regression anchor DOES break under it, root-caused to a specific
+suffix-table collision, and documented in the module docstring, the
+generated report's Limitations section, and this file rather than being
+silently avoided.
+
+IRON-RULE COMPLIANCE
+- stdlib-only Python 3 (argparse/json/math/re/sys/collections/pathlib —
+  unchanged from before).
+- No network access.
+- Every threshold in force (MIN_COOC=5, DE_MIN_STEM_LEN=4, EN_MIN_STEM_LEN=4,
+  topk) is stated in the report.
+- Primary TSV columns/filenames unchanged (verified via md5 before/after,
+  with and without the flag); the lemma-key TSV is a NEW file with the same
+  5-column schema, not a rename/reorder.
+- Reported a regression rather than tuning it away (tongue/sprache).
+
+FILES CHANGED: crates/lance-graph-planner/examples/data/rosetta/build_alignment.py only.
+NOT COMMITTED. Orchestrator to review + commit/push.