Cycle 7 Stage B: high-density durable catalog v3

docs/cycle7/README.md

@@ -6,9 +6,10 @@ This directory starts Cycle 7 development. It does not retune Cycle 6 and does n
 
 ## Live state (verified 2026-08-25)
 
-See `docs/cycle7/live-state-audit.md`. Independent GitHub inspection matched the Cycle 6 freeze:
+See `docs/cycle7/live-state-audit.md`. Independent GitHub inspection:
 
-- `main` HEAD `ddccd74a9e0e710085b385fed98959f2730b9d60`
+- `main` HEAD `af341884fff211caeb24649a6ce79305c3491c6c` (PR #95 squash-merge)
+- Cycle 6 freeze parent `ddccd74a9e0e710085b385fed98959f2730b9d60`
 - PR #93 and PR #94 merged
 - Recovery run `32886342498` success
 - Aggregate hash `30577aafaffd0c50f0ddb384a4509eb0bb93e4374bf39704869ddbf5053186a4`
@@ -28,8 +29,11 @@ See `docs/cycle7/seed-ledger.md` and `fuckmark/cycle7/ledger.py`.
 
 | Role | Seed base | Status |
 | --- | --- | --- |
-| Exploratory development (Stage A) | `810000` | in use |
-| Validation development (Stage B) | `820000` | reserved, unused |
+| Exploratory development (Stage A) | `810000` | used; do not keep expanding rules against it |
+| Exploratory development / rule construction (Stage B1) | `860000` | used; do not keep expanding rules against it |
+| Exploratory development / rule construction (Stage C1) | `870000` | active; topic frozen as `measurement protocol` |
+| Validation development (Stage B3) | `820000` | used as Stage B3 disjoint validation; do not retune on it |
+| Validation development (Stage C, reserved) | `880000` | unseen until a v4 freeze |
 | Confirmation reserved | `830000`, `840000`, `850000` | must not be inspected |
 
 Confirmation seeds were chosen as unused 10k blocks before any Cycle 7 detector look.
@@ -46,13 +50,29 @@ The Cycle 6 frozen sanitizer module is unchanged.
 
 ## Durable transform families
 
-See `docs/cycle7/durable-families.md`. Catalog `cycle7-durable-rule-catalog-v2`.
+See `docs/cycle7/durable-families.md`. Catalog `cycle7-durable-rule-catalog-v4`.
+
+Stage A (`v2`) families:
 
 1. Unambiguous contractions and closed orthography (Family 1).
 2. Attested open ↔ hyphenated compounds (Family 2).
 3. In-word ASCII apostrophe ↔ U+2019 (Family 3).
 
-These change tokens without relying on repeated U+0020. They survive ASCII whitespace collapse by construction. Family 3 also survives NFKC and Cf-strip.
+Stage B (`v3`) additions, aimed at natural site density:
+
+4. Sentence-boundary newline (layout; survives `whitespace-collapse-v1` because that sanitizer keeps LF).
+5. Bounded optional complementizer / object-relative `that`.
+6. Sentence-initial discourse comma.
+7. Attested prenominal hyphen modifiers.
+8. Parenthetical conjunctive adverb.
+9. Coordinating-conjunction comma (`and`/`but`/`or`), with determiner-gated insert.
+
+Stage C (`v4`) additions, aimed at higher natural site density after Stage B showed families 5–9 were sparse:
+
+10. Clause punctuation newline (`,` / `;` / `:` + space ↔ mark + LF).
+11. Optional `of` after `all` / `both` / `half` before a closed determiner/possessive.
+
+These change tokens without relying on repeated U+0020. They survive ASCII whitespace collapse by construction. Family 3 also survives NFKC and Cf-strip. Family 4 is a formatting channel: VERIFIED against the six Cycle 7 sanitizer variants, not against an unlisted reflow sanitizer.
 
 Quote interiors may receive these durable edits under `quote-container-durable-v1` without changing quote delimiters. Cycle 6 `quote-container-surface-spacing-v1` still admits only spacing inside quotes.
 
@@ -64,6 +84,16 @@ Families 2–3 raise candidate density on the same frozen texts (see `evidence/c
 
 Details: `docs/cycle7/stage-a-decision.md`.
 
+## Stage B
+
+Stage B1 on seed `860000` looked **`PROMISING_DEVELOPMENT`** on a tiny detector snapshot (durable 2/4 after collapse). Disjoint validation seed `820000` reproduced density/geometry but **not** the detector reduction (durable 4/4 after collapse). Overall Stage B decision: **`INSUFFICIENT_EVIDENCE`**. Details: `docs/cycle7/stage-b-decision.md`.
+
+Seed `820000` is spent as validation. Seed `880000` is the next unused validation split and has not been inspected. Seeds `830000` / `840000` / `850000` remain unseen confirmation reserves.
+
+## Stage C
+
+Stage C1 on seed `870000` (topic frozen as `measurement protocol`) added clause-punctuation newlines and optional quantifier `of`. Mean durable candidates: **4.75** (still the Stage B four-site regime). Quantifier `of` mean **0.00**. Collapse-surviving intact/root **0.715**. Detector: durable **3/4** after collapse; Cycle 6 spacing **4/4** after collapse. Overall: **`INSUFFICIENT_EVIDENCE`**. Seed `880000` was not inspected. Details: `docs/cycle7/stage-c-decision.md`.
+
 Not claimed:
 
 - Cycle 7 formal confirmation

docs/cycle7/durable-families.md

@@ -1,6 +1,6 @@
 # Cycle 7 durable families
 
-Catalog identity: `cycle7-durable-rule-catalog-v2`.
+Catalog identity: `cycle7-durable-rule-catalog-v4` (Stage C development). Families 1–3 remain the frozen Stage A `v2` set; families 4–9 remain the frozen Stage B `v3` set. Cycle 6 `quote_safe_zrd` ruleset hash is unchanged.
 
 All families below are deterministic, detector-blind, key-blind, and non-neural. Cycle 6 `quote_safe_zrd` ruleset hash is unchanged.
 
@@ -88,3 +88,142 @@ Conflict geometry: an apostrophe edit overlaps contraction expansion of the same
 text change → tokenization change → n-gram/context change → observation replacement → g-value drift → score drop.
 
 Byte change or a one-off score drop is not enough. Family 1–3 have VERIFIED tokenization change and collapse survival on fixtures. Detector attachment on TinyDev 64-token text is a separate density question.
+
+## Family 4 — sentence-boundary newline
+
+Construction: `sentence_boundary_newline`.
+
+Precondition: `. ` / `? ` / `! ` immediately before an uppercase letter. Exclusions: digits before the mark, single-letter initials, a closed abbreviation list (`Dr.`, `Fig.`, `U.S.`, months, `etc.`), already-newline after the mark, and ellipsis/multi-punctuation (`... Next`).
+
+Rewrite: swap the single ASCII space after sentence punctuation for a single LF, or the reverse.
+
+Inverse: the opposite direction.
+
+Tokenizer: GPT-2 `Hello. World` vs `Hello.\nWorld` is `[15496, 13, 2159]` vs `[15496, 13, 198, 10603]`. VERIFIED on frozen revision `607a30d783dfa663caf39e06633721c8d4cfcd7e`.
+
+Whitespace collapse: `whitespace-collapse-v1` preserves newlines. The rewrite still differs after `ws_collapse` and `ws_collapse_nfkc_cf_strip`. VERIFIED on fixtures. LF is category Cc, not Cf.
+
+Sanitizer note: this is a layout/formatting channel. A wrap/reflow sanitizer is not in the Cycle 7 frozen sanitizer suite. Survival is VERIFIED only against the six listed Cycle 7 variants. HYPOTHESIS: a paragraph-reflow sanitizer would erase it.
+
+Conflict geometry: adjacent sentence boundaries do not overlap. Format edits can still compete with spacing for the B14 budget in the combined arm.
+
+## Family 5 — bounded optional complementizer / object-relative `that`
+
+Construction: `bounded_complementizer_that_drop`, `bounded_complementizer_that_insert`, `bounded_object_relative_that_drop`.
+
+Precondition (drop): pronoun + closed attitude/report verb immediately before ` that `, and a closed clause-starter after it. Object-relative drop: determiner + noun of length at least 3 before ` that `, then a subject pronoun + lowercase word.
+
+Precondition (insert): the same pronoun+verb pattern with a clause-starter and no `that` already present. The stored source `I think ` is a sentinel; matching uses the shared complementizer pattern.
+
+Rewrite: delete or insert optional `that`.
+
+Inverse: insert vs drop on the complementizer pair. Object-relative drop is one-way in v3.
+
+Tokenizer: GPT-2 `I think that the protocol works` versus `I think the protocol works` differs by the `that` token. VERIFIED on the frozen GPT-2 revision.
+
+Whitespace collapse: the edit is a word, not spaces. Survives.
+
+False positives blocked: `I found that book`, `so that the logs`, all-caps.
+
+EXTERNAL-VALIDATION-ONLY: optional complementizer *that* is a high-frequency English alternation in the UID/Jaeger literature. Local admission still requires the closed verb list.
+
+## Family 6 — sentence-initial discourse comma
+
+Construction: `sentence_initial_discourse_comma`.
+
+Closed markers include `However`, `Therefore`, `Moreover`, `In fact`, and other listed conjunctive adverbs.
+
+Precondition: sentence-initial (start or after `.?!`). Insert is blocked before degree-adverb followers (`much`, `many`, `long`, and the listed set) so `However much evidence` is not rewritten as `However, much evidence`.
+
+Rewrite: drop or insert the comma after the marker.
+
+Tokenizer: comma token appears or disappears. HYPOTHESIS on TinyDev density: sparse unless the sample uses discourse markers.
+
+Whitespace collapse: comma is not space. Survives.
+
+## Family 7 — attested prenominal hyphen modifiers
+
+Construction: `attested_prenominal_hyphen_modifier`.
+
+Closed list includes `well known` / `well-known`, `long term` / `long-term`, `open source` / `open-source`, and other listed prenominal compounds.
+
+Precondition: same hyphen-chain/path guards as Family 2. Hyphenation additionally requires a following lowercase noun so predicative `The method is well known.` does not fire.
+
+Rewrite: open / hyphenated.
+
+Tokenizer: hyphenation splits GPT-2 pieces. VERIFIED locally for `well known` / `well-known`.
+
+Whitespace collapse: hyphens are not spaces. Survives.
+
+EXTERNAL-VALIDATION-ONLY: CMOS hyphenates compound modifiers before a noun. Predicative forms stay open.
+
+## Family 8 — parenthetical conjunctive adverb
+
+Construction: `parenthetical_conjunctive_adverb`.
+
+Closed adverbs: `however`, `therefore`, `moreover`, and the listed set.
+
+Precondition: previous character alphabetic; drop matches `, however, `; insert matches ` however ` and does not fire on already-comma forms.
+
+Rewrite: `, however, ` / ` however `.
+
+Whitespace collapse: commas remain. Survives.
+
+## Family 9 — coordinating-conjunction comma
+
+Construction: `coordinating_conjunction_comma`.
+
+Closed conjunctions: `and`, `but`, `or`.
+
+Precondition: previous character alphabetic and next character alphabetic. Insert additionally requires a determiner follower (`the|a|an|this|these|those`) so `you and I` and `cats and dogs` do not gain a comma. Digit-comma lists such as `1, and 2` are blocked.
+
+Rewrite: `, and ` / ` and ` (same for `but` / `or`).
+
+Tokenizer: GPT-2 `failed and the replica` versus `failed, and the replica` differs by the comma token. VERIFIED on the frozen GPT-2 revision.
+
+Whitespace collapse: comma is not space. Survives.
+
+Safety: insert is a style comma before a determiner, not a full independent-clause parser. NP coordinations of the form `X and the Y` can still receive a comma. That is semantically conservative and style-marked. Stage B treats density as the first metric and may reject the insert direction if natural prose over-fires.
+
+Not admitted: `nor` / `yet` / `so` (adverb/conjunction ambiguity). Running-text `and` to `&` remains REJECTED (CMOS/AP: ampersand is not prose-equivalent).
+
+## Family 10 — clause punctuation newline
+
+Construction: `clause_punctuation_newline`.
+
+Precondition: `, ` / `; ` / `: ` immediately before an alphabetic word of length at least 2. Exclusions: start of string, digits immediately before the mark (thousands separators, `v2, and`), other punctuation immediately before the mark, already-newline after the mark, and single-letter followers (`Hello, X` / `a, b`).
+
+Rewrite: swap the single ASCII space after the clause mark for a single LF, or the reverse.
+
+Inverse: the opposite direction.
+
+Tokenizer: GPT-2 `Hello, World` versus `Hello,\nWorld` differs. VERIFIED on frozen revision `607a30d783dfa663caf39e06633721c8d4cfcd7e` when transformers is present.
+
+Whitespace collapse: same as Family 4. `whitespace-collapse-v1` preserves LF. VERIFIED on fixtures against all six Cycle 7 sanitizer variants.
+
+Sanitizer note: this is still a layout/formatting channel. HYPOTHESIS: a wrap/reflow sanitizer would erase it. It is admitted because Stage B showed sentence-boundary newlines were the only family that raised natural site density, and clause commas are much more common than sentence terminals on ordinary prose.
+
+Conflict geometry: `, and ` overlaps Family 9 coordinating-comma drop. Cover-greedy v4 remains detector-blind.
+
+Not admitted: generic extra spaces, NBSP, ZWSP, or converting commas to other Unicode spaces.
+
+## Family 11 — quantifier `of` before a determiner
+
+Construction: `quantifier_of_determiner`.
+
+Closed quantifiers: `all`, `both`, `half`. Closed followers: `the`, `this`, `that`, `these`, `those`, and the possessives `my` / `our` / `your` / `their` / `his` / `her`. `both` does not pair with singular `this` / `that`.
+
+Precondition: whole-word match plus the same hyphen/slash/alnum adjacency guard as Family 2. All-caps blocked.
+
+Rewrite: `all of the` ↔ `all the` (and the other closed pairs).
+
+Inverse: the opposite direction in the same closed list.
+
+Tokenizer: GPT-2 `All of the replicas failed` versus `All the replicas failed` differs. VERIFIED on the frozen GPT-2 revision when transformers is present.
+
+Whitespace collapse: the edit is the word `of`, not spaces. Survives.
+
+False positives blocked: `all of them` / `both of us` / `half of it` (pronouns require `of`), `all of a sudden`, `some/most/each of the` (those quantifiers are not optional-`of` in the same way).
+
+EXTERNAL-VALIDATION-ONLY: Cambridge Grammar of English / English Grammar Today: `all`, `both`, and `half` may appear with or without `of` before articles, demonstratives, and possessives, with no meaning change. Pronoun objects keep obligatory `of`. That literature is not a local TinyDev density measurement.
+

docs/cycle7/literature.md

@@ -33,4 +33,23 @@ VERIFIED sanitizer fate:
 
 HYPOTHESIS. Unambiguous contractions, attested hyphenation, and typographic apostrophes can destroy some observation windows. Whether they destroy enough windows on 64-token TinyDev GPT-2 text is a Cycle 7 measurement. Family 1 detector attachment on seed `810000` was INSUFFICIENT_EVIDENCE because natural site density was too low.
 
+EXTERNAL-VALIDATION-ONLY. Optional complementizer *that* is a high-frequency English reduction (Jaeger, *Cognitive Psychology* 2010, UID/information-density accounts). Cycle 7 admits only a closed pronoun+verb list plus object-relative `DET+N that PRON`. That literature is not a local density measurement.
+
+EXTERNAL-VALIDATION-ONLY. CMOS hyphenates compound modifiers before a noun (`well-known method`) and leaves predicative forms open. Family 7 encodes that bound, not a generic hyphenation whitelist.
+
+VERIFIED tokenizer effects added in catalog v3 on the same frozen GPT-2 revision:
+
+- `Hello. World` → `[15496, 13, 2159]`; `Hello.\nWorld` → `[15496, 13, 198, 10603]`.
+- `I think that the protocol works` versus `I think the protocol works` differs by the `that` piece.
+- `failed and the replica` versus `failed, and the replica` differs by the comma piece.
+
+HYPOTHESIS. Sentence-boundary newline survives `whitespace-collapse-v1` because that sanitizer preserves LF. A wrap/reflow sanitizer is not in the frozen Cycle 7 suite.
+
 REJECTED as a Cycle 7 direction: adding more U+0020 spaces to chase 7/192 → 0/192 on spent confirmation rows.
+
+REJECTED: running-text `and` to `&` (not prose-equivalent under CMOS/AP). U+2026 ellipsis NFKC-maps to `...`. U+2009 / U+202F NFKC-map to space.
+
+EXTERNAL-VALIDATION-ONLY. Cambridge English Grammar Today: `all`, `both`, and `half` allow optional `of` before articles, demonstratives, and possessives (`all the` / `all of the`) with no meaning change; `of` is obligatory before personal pronouns (`all of them`). Family 11 encodes that bound. That literature is not a local density measurement.
+
+HYPOTHESIS. Clause commas are much more common than sentence terminals on ordinary prose, so Family 10 should raise layout-channel density relative to Family 4 alone. Whether that is enough to replace Cycle 6 spacing after `whitespace-collapse-v1` is a Stage C measurement, not a theorem.
+

docs/cycle7/live-state-audit.md

@@ -62,3 +62,17 @@ Live GitHub still has `main` at `ddccd74a9e0e710085b385fed98959f2730b9d60`. Open
 
 Hygiene-only follow-up `6e587ea016cf5a2f4e7845a5550f97e0b5d92b34` made Release Readiness Baseline and CI succeed after `tests/test_repository_hygiene.py` rejected comments/docstrings in new Python. Cycle 6 sealed confirmation on that SHA succeeded. Codex review was usage-limited, not a code finding. CodeRabbit did not auto-review (repository star threshold).
 
+## Continuation (2026-08-25, after PR #95 squash-merge)
+
+PR #95 squash-merged at 2026-08-25 21:23:21 UTC after all 41 head checks succeeded and `mergeStateStatus` was `CLEAN`. No unresolved review threads.
+
+Verified `main` HEAD: `af341884fff211caeb24649a6ce79305c3491c6c`
+
+Subject: `Cycle 7 Stage A: collapse-resistant durable transforms (#95)`
+
+Parent freeze commit remains `ddccd74a9e0e710085b385fed98959f2730b9d60`. Cycle 6 formal result is unchanged: `NONZERO_RESIDUAL` 7/192.
+
+Open PR after the merge: [#96](https://github.com/byte271/FuckMark/pull/96) on `cursor/cycle7-stage-b-density-3fb1`, rebased onto `main` and retargeted. Open issues: none.
+
+Changing the PR base from `research/cycle7-durable-transforms` to `main` does not emit `pull_request` `synchronize`, so Actions on the rebase SHA must be retriggered by a follow-up push. Do not treat a CodeRabbit-only success as full CI.
+

docs/cycle7/seed-ledger.md

@@ -1,8 +1,8 @@
 # Cycle 7 seed ledger
 
-Ledger identity: `cycle7-seed-ledger-v1` (`fuckmark/cycle7/ledger.py`).
+Ledger identity: `cycle7-seed-ledger-v3` (`fuckmark/cycle7/ledger.py`).
 
-Seeds were assigned **before** Cycle 7 detector scoring.
+Seeds were assigned **before** Cycle 7 detector scoring. Seed `860000` and topic `independent replication` were frozen in the ledger **before** any Stage B generation or detector look. Seed `870000` and topic `measurement protocol` were frozen **before** any Stage C generation or detector look. Seed `880000` and topic `independent check` were frozen as the next unused validation split **before** any Stage C validation generation.
 
 ## Spent (forbidden)
 
@@ -17,10 +17,20 @@ Seeds were assigned **before** Cycle 7 detector scoring.
 
 | Role | Seed | May score detectors? |
 | --- | --- | --- |
-| Exploratory development (Stage A) | 810000 | yes, development only |
-| Validation development (Stage B) | 820000 | only after Stage A family freeze; not in this PR's detector run |
+| Exploratory development (Stage A) | 810000 | yes, development only; do not keep expanding rules against it |
+| Exploratory development / rule construction (Stage B1) | 860000 | used; do not keep expanding rules against it |
+| Exploratory development / rule construction (Stage C1) | 870000 | yes, development only; frozen topic `measurement protocol` |
+| Validation development (Stage B3) | 820000 | used as Stage B3 disjoint validation; do not retune on it |
+| Validation development (Stage C, reserved) | 880000 | **not yet inspected**; reserved until catalog v4 is frozen |
 | Confirmation reserved | 830000, 840000, 850000 | **no** until a later frozen confirmation protocol |
 
-Do not promote 810000 or 820000 into confirmation after seeing scores.
+Do not promote 810000, 820000, 860000, 870000, or 880000 into confirmation after seeing scores.
+
+Stage A generation used 810000 with pair stride 32, matching TinyDev pairing, one pair per domain, topic `reproducibility` chosen without a detector look.
+
+Stage B1 generation used 860000 with the same pairing rule and topic `independent replication`. Seed 810000 remains an exploratory seed for Stage A rescoring only.
+
+Validation generation used 820000 with topic `held-out evaluation`. That topic was frozen before validation generation. Do not retune catalog v3 on 820000. A later revised mechanism must not recycle 820000; Stage C reserved 880000 instead.
+
+Stage C1 generation uses 870000 with topic `measurement protocol`. Rule-construction generation must call `assert_rule_construction_seed(870000)`.
 
-Stage A generation uses 810000 with pair stride 32, matching TinyDev pairing, one pair per domain, topic `reproducibility` chosen without a detector look.