BFM-decomp/phase-ends/CURRENT_PHASE.md

CURRENT_PHASE — Phase 33.5: Docs consolidation, the wiki as source of truth, the tracked-.run/ prune, and the day-one decomp kit

Gate 1 approved 2026-09-07 (Drew, plan mode, Max, Fable 5.1). Rules R74–R83 ratified at gate 1 (the Phase 33 candidates (a)–(j); binding; full text in phase-ends/DIGEST.md §3). The approved plan is reproduced VERBATIM at the end of this file (§"Approved plan") — its ~/.claude/plans/ copy is not part of the repo. This is a sub-phase between Phase 33 (closed at v1.32.0) and Phase 34 (the flip, docs/phase34-seed.md, untouched); it closes at v1.32.1. It changes no build input: the fleet's 218/218 is asserted by make tools-health (report BINARY=main + audit-digest), not by an R22 rebuild. Baseline HEAD at open: 80d45b29b (Drew's Phase-33 CLOSE commit; tree clean).

Milestone (gate 2 — what Drew confirms)

1. The wiki is the source of truth: every file under docs/ (outside wiki/, how-to-ai-decomp/, sunset/) is either a row in docs/wiki/Reference-index.md or README-linked (the ⊆ check in tools/doc_links.py); the five new pages exist and are reachable from _Sidebar.md; no wiki/how-to page links a docs/ file outside the index; nothing links into docs/sunset/; tools/doc_links.py --strict = 0 broken / 0 pending / 0 untracked-cited; a real tools/wiki_render.py run exits 0.
2. No information lost: every sunset file is under docs/sunset/ with history (git log --follow), a row in docs/wiki/Archive-index.md (what · outcome · where it lives now) and a review row in docs/sunset/README.md (comm -3 empty).
3. Tracked .run/ is flip-ready: git ls-files -ci --exclude-standard .run = 0; the finished-work groups untracked (the list in the task-8 log == git diff --diff-filter=D); the two disassembly files untracked and named in tools/public_rewrite/untracked_after_rewrite.txt; tools/audit_public.py OK with its new content check and both R39 controls recorded; purge_set.txt unchanged so gate_scan.py --all stays PASS.
4. The kit exists and installs: decomp-architect/ (README, intake, SETUP, methodology, templates, corpus, memory-seed, MANIFEST) passes tools/kit_lint.py, and the dry-run install in a throwaway repo produced the expected manifest with the guardrail assertions unchanged — or, past the cut line, KNOWN-DEFECTS.md says exactly what did not (R66).
5. Memory reconciled: the off-project memory moved, the seven stale ones updated, the portable set present as the kit's seed.
6. make tools-health OK; PhaseEnd_Phase33.5.md + DIGEST §0/§2 append + CURRENT_PHASE.md archived; left for Drew's close commit.

Effort / model (R7/R26/R27 — every transition is PROMPTED, never assumed)

• Per task (the plan's table): Max for tasks 0, 3, 10, 12, 15; xHigh for the rest. Task 5's three document diffs may fan out to read-only agents (breadth); no Ultracode/Workflow otherwise. The dry-run (task 13) is ONE subagent under D6's guardrails.
• Drew-only (R6): every git push; the gate-2 confirmation; the eventual deletion of docs/sunset/.

Owner decisions (binding — AskUserQuestion, 2026-09-07)

1. Kit in-tree now (decomp-architect/), split into its own repository after the flip as xsig was. 2. Phase 33.5, v1.32.1, (a)–(j) ratified now as R74–R83. 3. Sunset = git mv → docs/sunset/ + git rm --cached of finished tracked .run/; Drew reviews and deletes later. 4. The kit's SETUP.md opens with "install ProjectArchitect 2.0 first, then this". 5. BFM's own memories stay in ~/.claude (kit seed only; no .claude-state/ retrofit). 6. The two fable_cd4 disassembly files: untrack + an audit-only rule + a content check; no second history rewrite (the tracked C + the pinned compiler reproduce those bytes). 7. Link form: relative in-tree links to docs/wiki/<Page>.md. 8. .run/: only what git tracks; no scan of the untracked 29 GB. 9. README's docs/ links stay.

Tasks (strict order; one commit per task after this file is updated; harness task list = the same rows, R28 — 16 at gate 1, 17 since the 2026-09-07 amendment, 18 since amendment 3)

• 0 Gate 1 close-out: this file (ledgers + checkpoint), R74–R83 into DIGEST §3 (+ §0), phase34-seed.md amended, TaskCreate ×16 — Max
• 1 Unblock, code side: tool docstrings re-pointed (decl_from_use.py, integration_resolver.py, jtbl_carve.py, autopsy.py, residual_class.py); CLAUDE.md lines 11 / 61 / the effort-default sentence — xHigh — see Log 2026-09-07 Task 1
• 2 Unblock, doc side: banners (concurrency-design, gen2-mips-matching-model, psyq-worklist, distill/*, hindsight-study HISTORICAL + supersession table); SETUP §1a rewritten; the pin count re-derived (one dated number + its command, both gen3 docs) — xHigh — see Log 2026-09-07 Task 2
• 3 Wiki — conventions + firewall: Docs-and-scratch-conventions.md, The-ROM-firewall.md (the gitignore template block); sidebar rows; Repository-layout/Contributing/ch.11 links; forward links in doc_links_pending.txt — Max — see Log 2026-09-07 Task 3
• 4 Wiki — reference + direction: Reference-index.md (list-driven), Where-the-project-goes-next.md, Contributing "AI use — conduct", Home rewritten, sidebar sections — xHigh — see Log 2026-09-07 Task 4 — P6 rules check done after it
• 5 Integrate the six (hand-matching-process; wave-metrics + the s37w block → wave-playbook; portable-decomp-workflow vs the 13 chapters; generic-decomp-package attributed; hindsight §2.1/§7 → ch.03 + permuter-ils; the gen3 docs → the direction page); re-point every referrer incl. retrospective.md:10–11 — xHigh — see Log 2026-09-07 Task 5
• 6 Sunset moves (59 files — family-hseq.md stays, see the log) with the referrer census by command; docs/sunset/README.md; Archive-index.md rows (backticked paths + v1.32.1); DIGEST §4 + docs/history/README.md refreshed; coverage: 0 docs/ files outside wiki/how-to/sunset uncovered — xHigh — see Log 2026-09-07 Task 6
• 7 Checkers: doc_links.py (sunset refusal, index-derived allowlist, wiki-first warnings, TRACKED/UNTRACKED citation classifier + --disk, coverage ⊆); reachability in wiki_render --selftest; timeline.py regen in report BINARY=main + --check in audit-digest; the 13 dangling cookbook cites; the gitignore-template diff; tools-health wiring; SETUP rows — xHigh — see Log 2026-09-07 Task 7
• 8 Tracked .run/ prune (218 paths untracked: the 172 inertia files + 44 finished logs + the 2 firewall listings; untracked_after_rewrite.txt; audit_public check 4 with both controls; runbook §11; the four tool notes) — xHigh — see Log 2026-09-07 Task 8 — P6 rules check done after it
• 9 Memory reconciliation (the off-project file parked; the 7 stale updated in place; bfm-decomp-context-system refreshed; the seed set — 16 files — under decomp-architect/memory-seed/ with upstream: PA tags) — xHigh — see Log 2026-09-07 Task 9
• 10 Kit part 1: README.md, intake.decomp.md, decomp-architect.md, templates/registry-E.decomp.md, corpus/decomp-kernels.md, templates/PLACEHOLDERS.md — Max — see Log 2026-09-07 Task 10
• 11 Kit part 2: the firewall pack, docs/run READMEs, ops-setup, bootstrap, CLAUDE overlay, pa-overlays.md, LICENSE/NOTICE/README/CONTRIBUTING skeletons, .clang-format + format snippet, tools/MANIFEST.md; tools/kit_lint.py in tools-health — xHigh — see Log 2026-09-07 Task 11
• 12 Kit part 3: SETUP.md (Step 0–Step 10, answers:, the PA-2.0 version pin, the honesty section, Path A only) — Max — then P6 rules check — see Log 2026-09-07 Task 12
• 13 Dry-run install under D6's guardrails; fix; re-run to green; fixture + expected manifest + logs tracked under .run/P33.5/kit-dryrun/ — xHigh — see Log 2026-09-07 Task 13
• 13.5 (added 2026-09-07 by Drew — plan amendment, P5d; AMENDED the same day after the dry-run: the dictionaries, in-tree) Tools audit + the derived tool index + the two corpora as dictionaries: tools/tool_census.py — for every file under tools/ outside submodules/vendored code: the docstring line, its consumers (Makefile/.mk targets, CI, the wave playbook, its SETUP row, other tools by import or subprocess), its class (LIVE · REFERENCED · ORPHAN; PORTABLE vs PROJECT-ONLY) and its ladder phase; coverage-asserted against find (R32) → the generated docs/tool-index.md (KEEP-GEN, --check in tools-health, a Reference-index row, linked from The-matching-workflow + Toolchain-setup, one pointer line in how-to ch.04/05/10) whose rows are dictionary entries keyed by the NEED they answer (what · when the need shows up · why — the SETUP row's incident · what proved it — its consumer/control · what to adapt — the hard-codes); the kit's tools/MANIFEST.md RE-DERIVED from the census (--manifest, byte-equality asserted); the hand review of the no-consumer class (91 of 326 at S91) under Drew's criterion: a tool that served a purpose and was REQUIRED to complete the decomp is judged on its value to a FUTURE decomp — still needed (stays live and/or becomes a port task), superseded by a better tool (retired, the row NAMING the successor), or a one-off (retired, the row naming what it produced); retirement = git mv → tools/sunset/ only after the census proves 0 consumers, an archive-index-style row per tool, its SETUP row moved to a retired section; Drew reviews and deletes later. The dictionaries (Drew, 2026-09-07, confirmed in-tree now): tool_census.py --corpus materialises the tools VERBATIM into decomp-architect/corpus/tools/<phase>/ (LIVE + REFERENCED portable tools and the project-only shape references; retired one-offs out; superseded ones as a pointer file naming the successor) and the knowledge base VERBATIM into decomp-architect/corpus/cookbook/ (the cookbook, its symptom index, the codegen map) behind one front page stating what transfers (gcc 2.7.2 family = direct use, re-proven on the new bytes; other compilers = the symptom → the named pass → read the same pass in your compiler → a five-line reproducer → your lever); both copies regenerated by a make kit-corpus target and asserted EQUAL to the live files in tools-health (no drift, R51); kit_lint exempts corpus/tools/ and corpus/cookbook/ from the de-specialisation check (evidence, verbatim) and asserts coverage (every corpus file has an index entry) — the ROM audit covers them like everything else; the new project installs only the index + the front page + an ops-setup row naming the corpus (SETUP Step 6; its row count derived from the manifest header, no typed 325). Two memory seeds + two registry rules in the kit: consult-the-tool-dictionary-first (before designing or debugging a tool, look the need up; the proven implementation is the jumping-off point, its incident row the failure mode to avoid) and translate-the-source-idiom (a symptom + a named pass + a lever + a byte proof; translate through the pass, never copy the lever; re-prove on your bytes) — G66/G67 in registry-E.decomp.md, cited from the ladder's Phase 7 row and the methodology. A five-tool layout-contract probe (the corpus oracle, match_one, rtu_match, harvest_verify, dedup_propagate: count the layout assumptions — paths, aliases, config names, make targets) → decomp-architect/templates/layout-contract.md (a draft; stamping it at Step 4 is a split-phase decision) — xHigh (the ORPHAN review draft is breadth-shaped: one read-only Explore agent, verified against the consumer list) — verify: rows == find count; tool-index --check green; every retired tool has 0 consumers AND a named successor or product; MANIFEST regenerated byte-equal; the two corpora equal to their sources; kit_lint OK; doc_links --strict rc 0; the dry-run's Step 6 re-verified by a resume on the run-3 throwaway; make tools-health OK (detached) — see Log 2026-09-07 Task 13.5
• 14 Wiki Start-a-new-decomp-project.md + README/Tools-from-this-project.md rows + SETUP rows (R21) + decision-log entry (R31) + accelerators entry if earned — xHigh — AMENDED S91 (Drew): also implement the types hindsight into the kit per docs/decision-log.md "P33.5 S91-b" items (a)–(e): DK-65, the Phase-6 milestone's canonical type layer, G62 extended (or G68), the type tools' need-keys re-pointed at Phase 6, the cookbook front page's width sentence, one paragraph in the methodology; then make kit-corpus + tool_census --check + kit_lint — see Log 2026-09-07 Task 14
• 14.5 (added 2026-09-07 by Drew — plan amendment 3, P5d, S92: "do all three") The record as the third dictionary + the two coverage checks + the phase-log mining pass. (1) decomp-architect/corpus/record/ — the how-to (13 chapters), docs/decision-log.md, docs/accelerators.md, docs/retrospective.md, docs/story.md, docs/wave-playbook.md, docs/effort-map.md, docs/gen3-standards.md, docs/gen3-handoff.md, phase-ends/DIGEST.md and every phase-ends/PhaseEnd_*.md — VERBATIM behind an authored front page (what each is, how to read it, that the phase logs are NOT in it — R19 — and that item 3 is their distillation), regenerated by make kit-corpus, asserted sha1-equal by tool_census --check, lint-exempt like the other two corpora; SETUP Step 6 installs the front page (docs/inherited-record.md); every "two dictionaries" → three (kit README/tree, ops-setup rows, the methodology, the wiki page, the Tools page, README, the SETUP row). (2) tools/kit_coverage.py + config/kit_coverage_map.tsv (the authored residue): every R1–R83 is cited by a registry provenance line OR has a disposition row (the G-ids it became / ENV = environment-specific, not portable / PA = ProjectArchitect's own / FOLDED into a named G); every ## entry of docs/accelerators.md is cited by a kernel provenance line OR has a row; unknown ids refused; denominators printed; wired into tools-health after tool_census --check; SETUP row + dictionary row (R21/R32). (3) the phase-log mining pass — the 26 logs under phase-ends/logs/ (30,510 lines) as 22 read-only slices (the three giants split by line range), one Opus agent each in two batches of ≤11, briefed by .run/P33.5/log-mining/BRIEF.md: extract ONLY lessons that appear in no cookbook section, decision-log entry, accelerator, kernel, rule, playbook or how-to chapter — with the greps that prove it — into .run/P33.5/log-mining/<slice>.md (deliverable early, one JSON line last, R67); my review verifies each against the cited log lines (R14) and banks the real ones as DK-66+ / accelerator entries / decision-log entries, with a harvest table (candidate · verdict · where banked) in this log. (4) Step 6 changed → re-verify by a resume on the run-4 throwaway (.run/P33.5/kit-dryrun/repo/, from Step 6); expected-manifest.txt +1 (docs/inherited-record.md); the judge re-run — xHigh (item 3 is breadth: Opus agents) — verify: tool_census --check OK incl. the record; kit_lint OK; kit_coverage OK with its denominators; the resume PASS + judge PASS; doc_links --strict rc 0; render selftest 32/0; every mined candidate banked or dispositioned — see Log 2026-09-07 Task 14.5
• 15 Close: real render; tools-health foreground; doc_links --strict; audit_public; PhaseEnd_Phase33.5.md + DIGEST §0/§2 + log archive; left for Drew — Max (Tier 1, WAIT for gate 2)

Blockers

• None at open. The GitHub Support purge (ticket #4736982) is Phase 34's blocker, not this phase's.

---

The execution ledgers (from the three gate-1 audits; the audit reports exist only in the planning session — R30)

L1. The docs/ classification (path → class · note). Classes: KEEP-LIVE · KEEP-GEN · FROZEN→SUNSET · INTEGRATE→SUNSET · SUNSET

Path (under docs/)	Class	Note
SETUP.md	KEEP-LIVE	§1a is false ("never committed") — rewrite (task 2)
accelerators.md	KEEP-LIVE	README-linked; its substance reaches the wiki via the kit corpus + the new-project page's table
actor-struct.md	KEEP-LIVE	Gen3 input (gen3-handoff §5)
automation-runbook.md	SUNSET	already bannered HISTORICAL/superseded by wave-playbook.md; economics in how-to ch.09
autopsy.md	FROZEN→SUNSET	generated by tools/autopsy.py from untracked .run/autopsy/; a P30 snapshot
backlog.md	KEEP-GEN	backlog.py prune in make report BINARY=main; 0 open
badges/*.json (4)	KEEP-GEN	progress.py write_badges(); README shields
burndown.jsonl	FROZEN→SUNSET	burndown.py, no target; superseded by story-timeline.md
calibration.md	SUNSET	P28/P29 close-rate sizing; economics in ch.09 + retrospective.md
canonical-decl-reconcile-design.md	SUNSET	"✅ RESOLVED — built as tools/reconcile_decls.py (cookbook §33)"
commit-map.tsv	KEEP-GEN	one-shot, public_rewrite/build_commit_map.py; README-linked
community-matching-model-plan.md	KEEP-LIVE	linked by the published matching-drafter-pipeline.md; parked (licensing decision)
concurrency-design.md	KEEP-LIVE	banner "not yet implemented" is FALSE — six tools implement it; fix (task 2)
cookbook-index.md	KEEP-GEN	cookbook_index.py --check in tools-health
debug-menu-list.txt	KEEP-LIVE	primary data (Drew's transcription, S45)
decision-log.md	KEEP-LIVE	README-linked; append-only (R31)
decompme-preset.md	KEEP-LIVE	README-linked; Phase 34 E1
difficulty.md	KEEP-GEN	make report
difficulty.{ov_SC01_005,ov_SC01_006,ov_SC01_077,ov_SC03_001,resident}.md (5)	SUNSET	orphaned per-binary snapshots (June/July; resident says "21 unmatched" — wrong now); regenerable by make report BINARY=<alias>
disc-completeness.md	KEEP-LIVE	wiki Overlays-and-modules links it
disc-ledger.md	KEEP-GEN	make audit-disc
distill/atbhbkbl.md, distill/axbm.md	SUNSET	banners say "PROPOSED ONLY" but both LANDED as cookbook §265–§269 (verified); kept as the A/B evidence ledger the cookbook cites — banner fix first (task 2)
doc_links_pending.txt	KEEP-LIVE	the pending-link mechanism
duplicates.md, duplicates.cross.md	KEEP-GEN	dup_report.py in make report BINARY=main
duplicates.{ov_SC01_077,resident}.md (2)	SUNSET	orphaned per-binary snapshots
effort-map.md	KEEP-LIVE	governs effort; how-to 01/08 link it
exemplar_curriculum.md	SUNSET	exemplar_miner.py snapshot of 2026-06-19; frontier empty
fable5-strategy-review-2026-07.md	SUNSET	dated standpoint; conclusions in retrospective.md + decision log
family-cousins.md, family-hseq.md	FROZEN→SUNSET	make atlas would rewrite them EMPTY (0 stubs); keep the last populated snapshot
family-endgame-megaplan.md	SUNSET	bannered SUPERSEDED by roadmap-to-100.md; PhaseEnd_Phase26 records the outcome
family-manifest.md	SUNSET	Phase-25 target manifest; superseded
formats.md	KEEP-LIVE	wiki Home + Overlays link it
frontier-atlas.md	FROZEN→SUNSET	make atlas (2,520 open at 252dc930c) → EMPTY at HEAD
frontier-p30.md, frontier-s39.md	SUNSET	point-in-time reports, self-declared not re-baselined
frontier-p32.md	SUNSET	PhaseEnd_Phase31 says "keep forever" → MOVE, never delete; outcomes in PhaseEnd_Phase32
gcc-2.7.2-map/* (7 incl. cite_overrides.tsv)	KEEP-LIVE	README-linked; gccmap_cites --check
gen2-mips-matching-model.md	KEEP-LIVE	status line "not started" is stale — it was built; fix (task 2)
gen2-roadmap.md	SUNSET	Gen2 closes at Phase 34; the record is DIGEST + PhaseEnds
gen3-handoff.md	KEEP-LIVE (+INTEGRATE)	summarised into Where-the-project-goes-next.md; doc_links DEFAULT
gen3-parking-lot.md	KEEP-LIVE	correctly parked; gen3-handoff §7 re-adopts it
gen3-standards.md	KEEP-LIVE (+INTEGRATE §3)	README-linked; §3 → Contributing "AI use — conduct" + the direction page
generic-decomp-package.md	KEEP-LIVE (+attribute)	how-to 00/06 link it; attribute in ch.03/06
hand-matching-process.md	INTEGRATE→SUNSET	→ The-matching-workflow (+ch.07)
hindsight-study.md	INTEGRATE→SUNSET	mid-P29 at ~78%; superseded part-by-part (ch.03, generic-decomp-package, retrospective, gen3-handoff §5; §7 realised as autopsy.py); referrers: how-to 00/02, retrospective.md:10–11, gen3-handoff §8, autopsy.py/residual_class.py docstrings, backticked cites
history/README.md, history/claude-bfm-starting-point.md, history/project_architect_v1.3.0.md, history/cheap-tier-ab-experiment.md	KEEP-LIVE	archive; story/retrospective cite them
history/Fable5-ultracode-Full research findings-w42n9cp9t.output	SUNSET	216 KB raw JSON, one reader (its README); folded into SETUP/formats long ago
history/{debug menu link,ghidra psx plugin link,sotn-decomp wiki link}.txt (3)	SUNSET	URLs already in SETUP
idxtab-map.md	KEEP-LIVE	Gen3 lever (gen3-handoff §5)
matching-cookbook.md	KEEP-LIVE	README-linked; grep by §
matching-drafter-pipeline.md	KEEP-LIVE	README-linked
memory-map.md	KEEP-LIVE	wiki links it
next-session-triage-ladder.md	SUNSET	"BUILT AND SHIPPED (P31 S69) — Status: DONE"; corrections = cookbook §376/§378
outreach/archipelago.md, outreach/tools-announcement.md	KEEP-LIVE	Phase 34 tasks
permuter-ils.md	KEEP-LIVE	README-linked; receives hindsight §7's pointer
phase25-frontier-map.md, phase25-t6-curriculum.md, phase25-t6-fable-brief.md, phase26-triage.md	SUNSET	closed phase artifacts; PhaseEnd_Phase25/26 carry the outcomes
phase34-seed.md	KEEP-LIVE	the live Phase 34 plan
portable-decomp-workflow.md	INTEGRATE→SUNSET	"the seed of the public deliverable" — the 13 chapters are its successor; diff, fold uniques
progress.md, progress.json, progress.fleet.md	KEEP-GEN	make report BINARY=main; CI reads progress.json
progress.{ov_SC01_005,ov_SC01_006,ov_SC01_077,ov_SC03_001,resident}.md (5)	SUNSET	orphaned per-binary snapshots (resident says 88.97% — wrong now)
proto-correspondence.md	SUNSET	verdict in PhaseEnd_Phase3.5 + config/ghidra/ROSTER.md
psyq-sample-idioms.md	KEEP-LIVE	small; SETUP §5.6 cites it
psyq-worklist.md	KEEP-LIVE	the only per-object record of the 1,256 LINKED functions; add COMPLETE banner (task 2)
public-flip-runbook.md	KEEP-LIVE	README-linked; §11 gains the residue note (task 8)
recovery-queue-s65.md	SUNSET	closed by outcome (backlog 0, 218/218)
resident-dossier.md	SUNSET	resident is 145/145; check each lever has a cookbook § before moving
retrospective.md	KEEP-LIVE	README-linked; links hindsight at :10–11 → re-point (task 5)
roadmap-to-100.md	SUNSET	goal met; successors phase34-seed.md + gen3-handoff.md
second-oracle.md	KEEP-LIVE	Makefile + four tools read it
sister-decomps.md	SUNSET	realised as xsig's cross-game check
story.md, story-timeline.md, story-timeline.svg	KEEP-LIVE / KEEP-GEN	README-linked; timeline.py to be wired (task 7)
struct-core-pivot.md	KEEP-LIVE	Gen3 input
tool-designs/* (11: aprop-lane-s59, frontier-analysis-s60, frontier-analysis-s61, jtbl-automation-s59, jtbl-island-split, jtbl-island-split-review, main-gate-defects, main-lane-fix-s59, o0-cc1-lanes, o0-path-s59, tells-lane-s59)	SUNSET	closed P31 session artifacts; outcomes in PhaseEnd_Phase31 + cookbook §8/§8a/§8b/§8e/§61b–d/§62/§154-A/§236/§260; re-point decl_from_use.py, integration_resolver.py, jtbl_carve.py first
tooling-audit.md	KEEP-LIVE	how-to ch.04's evidence base — cite it there (task 5)
verification.md	KEEP-LIVE	README-linked; CI header cites it
wall-taxonomy.md	KEEP-LIVE	Gen3 input
wave-metrics.md	INTEGRATE→SUNSET	→ ch.09/05; its .run/s37w.js execution block → wave-playbook.md first
wave-playbook.md	KEEP-LIVE	the procedure
worklist.md	FROZEN→SUNSET	July snapshot ("79 live stubs"); read by make audit-frontier (outside tools-health) — the audit reports absence
wiki/*, how-to-ai-decomp/*	the wiki	the source set of wiki_render.py

L2. Tracked .run/ (1,086 files / 15 MB at open)

• Untrack, self-sealing (172 tracked-but-ignored; verified this session): s6h 38 · s6f 25 · w6 16 · s37 16 · s6g 15 · s6i 8 · t4 6 · s6b_cracks 2 · fable_frontier 2 · audit 1 · 43 loose root files (alias_defs.py, attract_loadmap.jsonl, disc_code_sweep.txt, jr_family_sweep.py, maintenance.sh, resident_wave_verdicts.json, sc03_hunt_loadmap.jsonl, s3_sc07_whale.py, s6f_gate.py, s6g_*, s6h_*, s10*, s33–s36_capture.py, s36_wave.json, s36w.js, s37_wave.json, s37w.js, s7_*, s8_*, s9_*, w6*, w7_wave*.json). Prose citations to move first: wave-metrics.md:62 (.run/s37w.js) and :153 (.run/s6f_gate.py). Tool defaults/comments to annotate: wave_judge.py:8 (WAVE=.run/t4), claude_wave_packs.py:16, jtbl_family_bank.py:380, dedup_propagate.py:114.
• Untrack with .gitignore narrowing (enumerate at task 8): .run/P32/t3s3/ session logs (67 tracked: 44 .log, 15 bank_*) and the duplicate tools_health.log / report_close.log under t3s3/t4b/t4d/t4e — keep any copy a doc cites (verification.md cites .run/P32/t4e/r22_check.log).
• Untrack + audit rule (the firewall pair): .run/giants/fable_cd4/mine_full.txt (398 lines, 0: 27bdffc0 addiu sp,sp,-64) and target_full.txt (399 lines, 0: addiu $sp, $sp, -0x40) — allowlisted by !/.run/giants/fable_cd4/*.txt. purge_set.txt must NOT change (gate_scan.py reads it as the history census); the rule goes in tools/public_rewrite/untracked_after_rewrite.txt, read by audit_public.py.
• KEEP-TRACKED: P33/verify (15) · P33/e1 (1) · P32/** drafts/notes/banks/reproducers (482 minus the logs above) · giants (279, minus the pair) · S79w (33) · s43 (22) · fable_80178004 (22) · wave22 (19) + wave22_targets.json · near6 (16) + near6_measure.{py,json} + near6_ils_*.log · s45 (12) · s42 (3) · probe_jtbl (3) · backlog.jsonl · fuel_manifest.json.
• POST-FLIP (0 tracked files; on disk only): .run/decompme/, .run/public_rewrite/, .run/objdiff/ — Phase 34 tasks 2/7.
• Doc citations of .run/: 653 backticked (58 tracked / 581 untracked-on-disk / 14 dangling incl. cookbook :36918 .run/P32/t3/restored/sweep_func_80039308.py, .run/frontier_p32.json, .run/backlog_drafts/func_80011380.c, .run/banked_func_*.json, .run/distill_inflight/*.json, .run/rtu_shadow/X.jsonl, .run/s7/scratch/spliced.c, .run/diag_plumbing.py) — the checker is disk-independent (TRACKED/UNTRACKED by git ls-files), strict only for wiki/how-to pages.

L3. The memory store (81 files + MEMORY.md under ~/.claude/projects/-home-musashi-bfm-decomp/memory/)

• ALREADY-IN-REPO (52): autonomous-lane-architecture · breadth-isolated-agents-not-serial† · build-tasklist-after-plan-approval · capture-knowledge-before-fresh-session · carve-state-files-never-blanket-add · check-against-a-known-true-case · checkpoint-current-phase-before-pause · commit-banked-work-immediately · commit-message-from-tool-output · commit-per-task-after-phase-log · cross-project-idiom-discovery · decomp-accelerator-ledger† · decomp-community-ai-standards · derive-from-invariants-not-reparsing · dont-conclude-unsteerable-try-register-pins · effort-prompt-ultracode-on-breadth · exonerate-the-instrument · fable-agents-for-lane-tooling† · fleet-tool-parallelism-defaults · gating-speed-playbook† · journal-notes-are-pack-fuel · lane-blockers-are-harness-not-model · matching-cookbook · matching-is-solved-integration-is-the-bottleneck · mcp-reconnect-after-restart · mcp-renames-dont-persist-use-applysymbols · measure-the-steady-state-not-the-launch · no-commit-co-author · no-tmp-project-local-data · parallel-gate-via-worktrees · pass-j-to-every-build · phase-worklogs-reference-only · pkill-pattern-kills-own-shell · private-repo-backup-policy (partly stale) · quote-the-denominator · report-every-lane-not-the-loud-one · reprobe-exclude-lists-after-tool-fixes · rescan-twins-after-every-bank · session-start-list-rules-in-full · session-summary-plain-english · setup-md-keep-current · silently-narrowed-tool-scope · standalone-match-is-not-bankable · structural-family-mechanical-remap · subagent-model-ladder† · tool-must-refuse-unsupported-input · tools-folder-convention · ultracode-harvest-pattern · verdict-names-its-instrument · verify-blast-radius-not-just-defect · wave-harvest-is-a-pipeline-step · wave-prompt-seed-step0-and-gaps · web-research-compiler-quirks. († carries a portable rider: the quadratic main-loop-context arithmetic; the accelerator-ledger habit as a rule; the per-stuck-lane briefing recipe; "a slow gate is a BUG"; what the frontier model is NOT for.)
• PORTABLE-NOT-IN-REPO (14): in PA 2.0's seed already — repo-self-contained-claude-state, drew-working-preferences, dont-block-loop- with-askuserquestion, justify-new-tools-before-adopting, clarify-misconception-before-costly-action. New for the kit (9): no-sleep-polling-background-tasks, live-coop-answer-before-grinding, tools-health-foreground-not-background, resume-means-resumefromrunid, tool-change-updates-siblings-and-docs, bank-idioms-before-checkpoint, phaseend-verbosity-for-the-retrospective, offline-tooling-first, wave-playbook-is-the-procedure — plus the five riders and three decomp seeds (matching-flywheel, byte-gate-is-the-only-claim, mcp-reconnect-after-restart). upstream: PA tag on the four harness-generic ones.
• BFM-SPECIFIC (9, keep): bfm-decomp-context-system · continuous-gater-lane-plan · crack-wave-sweep-map-regen · dedup-backlog-leave-it · disc-dump-location · endgame-budget-unconstrained · gate-main-only-with-gate-main · roadmap-to-100 · wsl-disk-capped-75gb.
• STALE (7, update in place): cheap-tier-ab-validated · effort-doctrine-xhigh-default (the memory is right; CLAUDE.md is stale) · private-repo-backup-policy · project-endgame-deliverables · roadmap-to-100 · rom-content-git-policy · bfm-decomp-context-system's status block.
• OFF-PROJECT (1, move): vantage-edge-never-leaks → the Vantage project's memory directory under ~/.claude/projects/.

L4. The wiki mechanism, the link policy and the kit design — see the Approved plan (§"What exists", D1–D6) below.

---

Log (newest last; one entry per task; the 🛑 block at the end is rewritten after every task)

2026-09-07 — Task 0 — Gate 1 close-out (Max)

Plan approved in plan mode after three exploration audits (docs/ tree, tracked .run/ + firewall, memory store + ProjectArchitect) and one adversarial Plan-agent review (12 amendments folded in: forward links via the pending list; referrer census by command; frozen snapshots instead of empty regens; the audit-only rules file because purge_set.txt is the history gate's census; a content check whose shapes come from the offenders; disk-independent citation checks; the kit as a Phase-0.5 overlay with an intake document for PA's own interview; dry-run write guardrails; a cut line). Spot-checked the claims the plan leans on (172 tracked-but- ignored; the two disassembly files; §265–§269 landed; the tools-health chain). R74–R83 written into DIGEST §3 from the Phase-33 candidates' text; DIGEST §0 notes the sub-phase; phase34-seed.md amended. Harness task list built (16 rows). Effort for task 1 is xHigh — the transition is prompted (R27). Commit 39d524991.

2026-09-07 — Task 1 — Unblock, code side (xHigh; Drew set /effort xhigh)

Seven docstring/comment cites re-pointed to durable records: decl_from_use.py + integration_resolver.py → the decision log's 2026-08-25 (P31 S61) "the wall is an integration wall" entry (the S60 frontier analysis's 178/292 = 61% and ~571 finished-draft measurements live there); jtbl_carve.py ×3 → cookbook §260 (the leading-island split; the review's correction #7); autopsy.py + residual_class.py → docs/how-to-ai-decomp/03-bootstrap-order.md Phase 2b (the permuter track — task 5 adds the "mine the permuter's failures" paragraph there; "originally hindsight-study §7, archived" kept as provenance). CLAUDE.md: R-range R1–R83; the H1 fail-safe dates the re-tightening to Phase 33 C3 (2026-09-06, the history rewrite) — the flip has not happened; the Reasoning section's "default working level is Max" replaced by the DIGEST §1 doctrine (xHigh for most tasks, Max for deep tasks, Ultracode for breadth; plan mode always Max). Verify: py_compile on the five tools OK; git grep -E 'tool-designs|hindsight-study' tools/ Makefile = 0; doc_links OK. Deviation from the plan's verify column: make tools-health was NOT run for docstring-only edits (proportionality — nothing it measures changed); the first tools-health of the phase runs at task 7, where the Makefile and the checkers change. Commit a0cf302e5.

2026-09-07 — Task 2 — Unblock, doc side (xHigh)

Six banners/status lines corrected in place (H5 — nothing dropped): concurrency-design.md "Not yet implemented" → implemented in shared_lock.py + the gate snapshot layer, five citing tools named; gen2-mips-matching-model.md "design / not started" → built in P22–P23, distilled by the published matching-drafter-pipeline.md; psyq-worklist.md COMPLETE at P31 S79 #5 with the residue named; distill/atbhbkbl.md + axbm.md "PROPOSED ONLY" → LANDED as cookbook §265–§267 / §268–§269 (the banners had been false since S59b/c); hindsight-study.md HISTORICAL banner + the seven-row supersession table (its §7 → autopsy.py/residual_class.py, live pointer in how-to ch.03). SETUP.md §1a rewritten: .run/ is ignored by contents (/.run/*) with 22 dated ! allowlist blocks; the R20 test ("commit what a rerun cannot reproduce"); what is tracked at the prune; untracked citations are breadcrumbs; the wiki page named in prose (no link yet — task 3 creates it). The pin count (R75): two published numbers came from two greps — the broad register\s+[^;]*__asm__ (44,243) also matched ~290 comment lines (two comment sentences repeated across 141 files, plus a "$N" placeholder), the strict \("\$[0-9a-z]+"\) form (43,857) missed 68 pins written without $ ("a0", "v0", "v1", "6"…). Canonical command grep -rhoE 'register [^;/]*__asm__\("\$?[a-z0-9]+"\)' src --include=*.c --include=*.h | wc -l = 43,925 (42,985 .c + 940 .h), written with its command into gen3-standards.md §4 and gen3-handoff.md §3 (both rows + the command block). Verify: doc_links OK; the command reproduces 43,925. Commit d06923a06.

2026-09-07 — Task 3 — Wiki: the conventions page + the ROM-firewall page (Max)

Two new pages under "Working conventions" in the sidebar. Docs-and-scratch-conventions.md: the kind-of-knowledge → home table (cookbook/decision-log/accelerators/runbooks/SETUP/memory-map/formats/phase-ends/how-to/wiki; a session note is never a reference), authored vs generated with the generator table (R75 — a prose number is a dated snapshot with its command; frozen snapshots are not regenerated), the link policy (wiki page not docs file; wiki→docs only via the Reference index; nothing links into the archive; the pending list R80; the render refuses dead links; backticked citations classified TRACKED/UNTRACKED by git, strict only in wiki pages), the archive procedure (git mv, index row, the referrer-census command, deletion is the owner's), and .run/ (R12; ignored by contents with the three-line ! idiom quoted from the P33 verify block; the R20 test; what is tracked at this phase; the per-session layout from the playbook's S80 addendum; never git clean -x; a tracked .run/ file is published). The-ROM-firewall.md: the rule with the exemption never to grant (the R1 relaxation and its bill; R74), the nine classes with the tracked form each takes instead, the copyable .gitignore template in a gitignore fence (generic paths; the class-9 tarball note; the scratch idiom with an example block — this block is the kit's templates/gitignore.decomp, byte for byte, task 11), the three enforcement layers (the audit's derived hash set, coverage assertions, R43 refusal, the content check "since Phase 33.5", the expected-offenders control; CI) and the one-list rule with the untracked_after_rewrite.txt sibling, the rewrite lessons as R76/R77/R81/R82/R78, the one recorded residue (the two fable_cd4 listings; the owner's decision and its reasoning), and the day-one pointer to the kit. Cross-links: Repository- layout (docs/ row, .run/ row, the not-in-repo section, R1…R83), Contributing (the firewall paragraph; "83 numbered rules"), how-to ch.11 (the day-one section + the proposed-rules heading now says "ratified as R74–R80"). Forward links registered as PENDING: Reference-index.md (task 4), Archive-index.md (task 6). Verify: wiki_render --selftest OK; doc_links 0 broken / 2 pending. Commit 5d10a0d12.

2026-09-07 — Task 4 — Wiki: the Reference index, the direction page, Contributing's conduct section, Home + sidebar (xHigh)

Reference-index.md: 47 rows in seven groups (environment and verification; formats and addresses; the knowledge base; verification and progress (generated); releases and outreach; the record; Gen3 inputs) + four directory rows (gcc-2.7.2-map/, badges/, outreach/, history/) — each with what it is, how to read it (the 3.5 MB / 560 KB warnings), who writes it; the page states it is the only sanctioned wiki→docs/ link site and that the two index pages are exempt. Where-the-project-goes-next.md: Phase 34 (the probe-gated flip, the outward actions in the maintainer's words, aftercare, v2.0.0), Gen3 (the floor, the bar — sotn's guide + the four rules —, the census table with the ONE pin count and its command, the invariant, the order of work, the levers, the conduct rules, shiftability scoped, the parked ideas), and the day-one-kit pointer in prose. Contributing gained "AI use — conduct" (the five rules of gen3-standards §3, linking the direction page). Home.md rewritten: the source-of-truth statement + sections Using / Working conventions / Reference / How to AI-decomp / Where next / History; _Sidebar.md gained Reference, Where-next and History (Archive index, pending). gen3-handoff.md §8: "R1–R73 plus candidates" → R1–R83; the last 44,243 → 43,925. doc_links_pending.txt: the Reference-index line removed (1 entry left: Archive-index, task 6). Coverage check (a Python one-liner over doc_links' LINK_RE — the task-7 checker's prototype): 132 docs/ files outside wiki/how-to; index ∪ README covers 66; the 66 uncovered are EXACTLY the task-6 leavers (33 root sunset + 4 integrate + 6 frozen + 12 per-binary snapshots + 11 tool-designs + 2 distill + 4 history files with spaces in their names) plus one KEEP file the first draft omitted — generic-decomp-package.md, added as a row (how-to 00/06 link it). Verify: wiki_render --selftest 12/12; doc_links 0 broken / 1 pending.

P6 rules check — re-read complete (CLAUDE.md's Mandatory Behavior + fail-safes, DIGEST §3 R1–R83): one task at a time, one commit per task after this log; never overwrite blind (H5 — every banner was added, none replaced); never commit ROM content (the two fable_cd4 listings leave at task 8, before any push Drew makes of this phase would matter — they are already in the pushed history); never git clean -x; a match is byte-for-byte (no build input has changed). Continuing with task 5.

2026-09-07 — Task 5 — Integrate the six documents; re-point every referrer (xHigh; three read-only diff agents)

Three Explore agents diffed the three big INTEGRATE documents against their successors; I folded, re-pointed and verified. portable-decomp-workflow.md vs the 13 chapters: a near-superset — 6 residual items. Folded: the per-binary gate LOCK (ch.02, the parallel bullet: worktree = isolation, lock = mutual exclusion); the one-sentence portability TEST ("still true for a PS2 game with a different compiler → transfers"; ch.00 "What transfers"); the three-part lane-restart trap (code / arguments / per-invocation defaults; ch.05 "Lanes"). Reconciled, not folded: the ≤30-instruction cheap-tier boundary (the chapters' ~50 is corroborated by the A/B and ch.09 — the ≤30 was the mid-campaign figure); "~780 sections" (measured: 490 ## § headings over 442 numbers — the chapters' "some 500" stands). Dropped: the automation-runbook sibling mention (it sunsets), the "any language" scope claim. wave-metrics.md vs ch.05/ch.09 + playbook §5: the chapters priced drafting by token but never per wave, never the gate, never in instructions, and carried two claims later waves refuted. Folded into ch.09: the sourcing rule (derived from logs, the gate is the arbiter); a "A wave, priced" table (waves 1–6, T/U/V: agents / tokens / wall / parallelism / banked) + the units (170k–300k tokens per banked head early, ~124k mature; harvest ≈ a quarter of a wave; recovery ≈0; the seed-twin payoff 15% tokens / 64% wall; outages as casualty counts); "Instruction weight, not head count" (the wave-5/6 table, pool realisation 43% vs 21%, wave T's 11,511); "The gate's cost is paid in wall-clock" (gate groups 5→7→9→3, drafts per rebuild 7.8→24.3, concentration is a draw-time CHOICE, the slowest-chain floor); Drafting-yield rows (the size curve's denominator caveat — a fleet-wide curve over 37 mostly pre-tooling waves; the prompt as the lever, mid-70s→100% with its honesty caveat; 98–100% standalone is noise, never a new floor; claimed-vs-banked per wave, 26 of 27 rejections byte-perfect, the reconcile lane 21/22); free siblings per wave (65·18·21·26·26·25 / 49·34·25·0·1·4; the 299→4,485 case; bimodal sweep, the family owns the rate); "budget recovery against gate drops, near-misses are permuter fuel". Folded into ch.05: the banked twin beats the open one (0/34 vs 4/4; the card builder discarded it), STEP 0 = the magic-literal grep, verify each claim separately + the stale prefetch; the dedup numbers (22 of 44 skeletons drawn twice; 19,266 potential = 6,557 drafted + 12,709 staged; 4,954 banked free); draw by open templatable instructions, FACE vs DELIVERED mass (the "nearly free mass" claim corrected: unknown-lever cards draft at 96–100% but yield zero siblings and a harvest of 4 vs 14), the UNKNOWN default-flag trap (607 of 1,179); step 5: pipeline vs batched 136→82 min / 2.5×→3.8× / 37–50-min dead gaps, the six 6,500-ins waves, the difficulty knob; step 6 REWRITTEN: "gate the directory, never the verdict list" replaces "read the verdicts" (the 362-ins NO-DRAFT and the 579-ins silent zero), the coverage assertion, gate-group concentration; NEW step 6b Reconcile before any rebuild + "when the reconcile lane goes quiet the prompt is doing its job"; step 8: index_hit/index_gap harvest, the yield series 9/14/14/4/7/5 vs already-covered 61/44/76/41/56/53 — "the knowledge base knew, the agents did not find it"; "Then recover": the gate number is not the close-rate (5/8 gate vs 8/8 oracle → 8/8 after recovery), the three plumbing levers. wave-playbook.md §5: the S37 execution block (the barrier-free pipeline() + the reduction, 46 lines, from .run/s37w.js — which is untracked scratch and stays so; the ladder inside it noted as superseded by the routing line) and the coverage law (the s6f gate script's assertion — NOT claimed to live in gate_wave.py: a grep found no such assertion there; worded as "whichever driver batches must assert it"). Internal inconsistencies in wave-metrics itself (its Finding-1 series mixes numerators; "405 drafted, 405 banked" is unsupported by its own table; wave W's draft count 69/70/71; "one more card" is two; ~45% is 47%) were NOT propagated — the table values were used, as the agent showed the tables are the derivable ones. hand-matching-process.md vs The-matching-workflow / ch.07 / ch.02: 48 unique items, 12 classes of STALE guidance excluded (the 55% / 136-binary numbers; pins as the lever and func_8012B8E4 as their proof — contradicted by R73; the pre-masked-scorer "permuter refuted" verdicts; harvest_verify as THE gate; the deleted signature tools' method; session scratch paths; forward plans; historical close-rates; the "only genuine stub" narrow-param wall — dissolved by cookbook §43; the private-era NON_MATCHING line; Ghidra program ids). Folded: a new "Matching by hand" section in The-matching-workflow.md (two decompiles + the asm as arbiter, the MCP/overlay mechanics, the m2c direct invocation; the loop + the two-command red-gate diagnosis; THE FIVE SIGNATURE MOVES with their byte-neutrality proofs and the surgical-not-blanket warning, the layer's placement, types byte-neutral vs local widths not; choosing what to hand-match — the ascending-mismatch anti-signal, the tractable shape, the count-weighted metric, no magic five, callees before callers, the circular callee; running it as a wave — conflicts are manufactured by parallelism (60%→33% = 100% compile errors), draft → sig_unify → gate, chunk 1, the recovery pass, the headless pre-decompile + the FUN_ rename + opaque ids, the Workflow args gotcha, scratch filtering, "no draft" as a valid result 12/30, propagation fail-closed + --auto-from + the signature blast radius); a "Levers from the hand-matching era" list in ch.07 (the scaffold ×4 arithmetic tell, branch polarity, rematerialize-vs-spill, the cross-jump count law, dual bases in first-use order, the mask-local idiom, goto ret0 clustering, v0/v1 coalescing's failed steers, the walking pointer, two phantom-frame variants, the permuter's two no-op modes, the triage routing with the pin branch re-routed to shape per R73); two bullets in ch.02 (parallelism manufactures conflicts → normalise before the gate, chunk 1, the recovery pass; a red gate is a hash — the two-command diagnosis; propagation as a fail-closed gate with the blast radius). The other three: generic-decomp-package.md attributed in ch.03 (the band's numbers 22/352 → 75/352, rank by work) — ch.06 already cited it; the hindsight §7 track → ch.03 Phase 2b (the autopsy paragraph) and permuter-ils.md §5; the Gen3 pair are the direction page (task 4). ch.04 now cites tooling-audit.md as the evidence base. Referrers re-pointed (11 sites): ch.00 :59–60 (both bullets → the archived-synthesis line), ch.02 :16 (→ ch.03), retrospective :10 (link → archived citation), :96, :130, :143, gen3-handoff :210 (→ ch.03 Phase 1), permuter-ils :118, wave-playbook :708 (+ the kit pointer), accelerators :7, SETUP :1312 (the S88 record line notes the archive). Census after: 0 LINKS to any of the four leavers outside records; 6 backticked prose mentions remain, each naming the archive (docs/sunset/… or "archived") — the sanctioned citation form. Verify: doc_links OK (1 pending), wiki_render --selftest 12/12.

2026-09-07 — Task 6 — The sunset moves + the Archive index (xHigh)

The move list was built from ledger L1 (.run/P33.5_sunset_list.txt, scratch) and checked against the task-4 coverage one-liner. One correction to the plan (R14): docs/family-hseq.md is NOT a frozen snapshot — tools/verify_contract.sh step 02 regenerates it as part of the recorded contract run ("family_hseq regenerated: 217 binaries scanned, 0 open instances" is the contract's own statement) and tools/burndown.py reads it; it stays as KEEP-GENERATED and got a Reference-index row. The other five frozen snapshots (frontier-atlas, family-cousins, worklist, autopsy, burndown.jsonl) moved. The referrer census (the D3 command over all 59 basenames, excluding records/self/indexes): every hit was either a substring false positive (calibration, worklist, autopsy are ordinary words and tool names), an "archived"-style prose mention written in task 5, a provenance label in a tool comment naming an S60/S61 analysis by its short name (left as provenance — not a path), or a digest/SETUP line, which were refreshed: DIGEST §4 rewritten around the wiki (the index pages first; roadmap-to-100 / frontier-p32 / automation-runbook named as archived, phase34-seed + the Gen3 pair added), DIGEST §1's gen2-roadmap line marked archived, SETUP.md's integration_resolver row and docs/history/README.md updated. resident-dossier.md's eight cookbook § cites all resolve to headings (§17, §37, §41, §43, §47, §52a, §52b, §53) → safe to move. Moved: 59 files by git mv into docs/sunset/ — 25 root documents, distill/ 2, history/ 4, tool-designs/ 11, generated-snapshots/ 17 (the 5 frozen + the 12 per-binary reports); docs/tool-designs/ and docs/distill/ are gone. Indexes: docs/wiki/Archive-index.md (six groups; every file as a backticked path + what it was / what came of it / where it lives now, with wiki-page links only), docs/sunset/README.md (the ☐ review table, 59 rows; deletion is the owner's); the last pending entry cleared. Verify: README rows == git ls-files docs/sunset (59 = 59, comm -3 empty); every sunset basename appears in the Archive index (0 missing); the coverage check: 64 docs/ files outside wiki/how-to/sunset, 0 uncovered; doc_links: 56 documents, 429 links, 0 pending, 0 broken; render selftest 12/12; git log --follow on three samples confirmed AFTER the commit (a staged rename shows no history until committed).

2026-09-07 — Task 7 — The checkers (xHigh)

tools/doc_links.py rewritten around six checks, every count with its denominator: (1) links as before + the pending list; (2) the archive refusal — any link into docs/sunset/ is BROKEN; (3) wiki-first — a wiki/how-to link into docs/ must be a target of Reference-index.md or README.md (the allow-list is DERIVED from those two pages; the two index pages exempt); (4) coverage — every tracked docs/ file outside wiki/how-to/sunset is covered the same way (64 of 64); (5) citations — every backticked docs/… / .run/… path is TRACKED or UNTRACKED by git ls-files (disk-independent; <…>/globs/ellipses skipped; path:12-19 forms look up the file; `path` (not kept) is declared gone and skipped); a wiki page citing an UNTRACKED path fails; --disk adds the PRIVATE/DANGLING split; (6) wiki-first WARNINGS (exit 0) from a small topic→page map — today 2 (gen3-handoff and story link verification.md; left, they are records/narrative). tools/wiki_render.py --selftest gained the reachability assertion (every wiki page in the sidebar; every chapter in the sidebar AND the how-to index): 31 pages, 0 unlisted. tools/timeline.py wired: regenerated in report BINARY=main after progress.py --json, --check in audit-digest — it WAS stale at HEAD (72 → 73 rows: the Phase-33 close day); the two files regenerated and committed. New tools/gitignore_template_check.py (the page's ```gitignore fence == the kit's template; rc 1 drift, rc 2 no template yet — never a pass; refuses ≠1 fence) behind an existence test in tools-health that skips loudly until task 11. What the new checks found immediately: the conventions page cited two files only the future kit stamps (docs/README.md, .run/README.md) → reworded; the classifier's first cut read path:line cites as unknown paths → fixed. The R39 control on the cookbook (--disk docs/matching-cookbook.md, not in the default set): BEFORE = 13 DANGLING of 276 untracked citations (the audit's "14" — its regex differed by one); of the 13, FIVE were documents this phase archived (tool-designs/jtbl-automation-s59.md, tells-lane-s59.md, distill/atbhbkbl.md, distill/axbm.md, frontier-p32.md) → re-pointed to their tracked docs/sunset/… paths (my first pass had wrongly marked them "(not kept)" — caught by reading the list; plus 5 more archived-doc cites re-pointed earlier in the task: hand-matching-process ×2, exemplar_curriculum, phase25-frontier-map, calibration); the other EIGHT are .run/ scratch that exists nowhere (.run/_bak.c, .run/diag_plumbing.py, .run/s7/scratch/spliced.c, .run/wave3/func_8018FF98/vA,…c, .run/match/func_80187130.246967/, .run/harvest_x/vsweep.py|vsweep.json, .run/.../sweep2.py, .run/P32/t3/restored/sweep_func_80039308.py) → annotated (not kept) in place. AFTER = 0 DANGLING, 263 PRIVATE (breadcrumbs, allowed in a record). cookbook_index.py --check OK (1,169 sections; no heading changed). Default set: 56 documents, 429 links, 0 pending, 0 broken, 551 citations (430 tracked / 121 untracked, none in a wiki page), coverage 64/64; --strict OK. make audit-digest rc 0 with the timeline check. SETUP rows for doc_links, wiki_render, timeline and the new tool (R21). The conventions page's description of the checks matches the implementation.

2026-09-07 — Task 8 — The tracked-.run/ prune, the audit-only rules file, the content check (xHigh)

Removal set = what .gitignore says. After narrowing the re-includes (!/.run/P32/t3s3/*.log → the three cited R22 logs by name; explicit re-excludes for the uncited duplicate close logs under t4b ×2, t4d ×2, t4e ×1; /.run/giants/fable_cd4/*_full.txt after the *.txt allow line), git ls-files .run | git check-ignore --stdin --no-index yielded 218 paths — the 172 inertia files (43 loose root files; s6h 38, s6f 25, w6 16, s37 16, s6g 15, s6i 8, t4 6, s6b_cracks 2, fable_frontier 2, audit 1) + 39 uncited t3s3 root logs (the gate/ verdict ledger and the 3 cited R22 logs stay) + 5 duplicate close logs + the 2 listings — and that list IS the removal set (git rm -r --cached), so the ignore file and the index agree by construction. Kept: everything a document cites (12 cited logs under .run/P32/, found by grepping each tracked .log name over docs + phase-ends + README). Invariants after: tracked .run/ 1,086 → 868; git ls-files -ci --exclude-standard .run = 0; git status --porcelain .run | grep '^??' = 0; all 218 removed paths still on disk; staged deletions == the list (cmp YES). The firewall pair: untracked; tools/public_rewrite/untracked_after_rewrite.txt created (audit-only; the header says why it is not the purge set); purge_set.txt byte-unchanged; gate_scan.py --all --worktree PASS (112,518 reachable blobs / 16.25 GB, 0 offenders, 73 s). audit_public.py check 4 (CONTENT): longest contiguous run of disassembly-shaped lines per text file, cap 64. R39 controls, both recorded: the first cut of the asm-differ shape required an operand, and nop lines broke the target listing's run to 60 < 64 — the control caught it (only mine_full.txt at 398 failed); the shape now accepts operand-less mnemonics and .L label operands, and the control by --paths over the pair + the two xsig fixtures reads 398 and 179 → FAIL, 14 and 14 → pass; the full audit AFTER the untrack: OK — 0 offenders among 6,443 tracked paths (6,661 before), longest runs 14/14/6. runbook §11 gained the residue paragraph; the four tool comments annotate the untracked defaults (wave_judge.py WAVE=.run/t4, claude_wave_packs.py, jtbl_family_bank.py, dedup_propagate.py); SETUP row for the audit (R21). Deviation from the plan's numbers (R41): the plan said "~50 allowlisted logs"; the enumerated count was 44 (39 + 5) because the t3s3 gate/ logs and every rtu/variant log under t4b/t4d/t4e are evidence and stay — and the audit's "≈860" became 868 for the same reason.

P6 rules check — re-read complete (CLAUDE.md's Mandatory Behavior + fail-safes; DIGEST §3 R1–R83): one task per commit; never overwrite blind (the .gitignore edits are additive re-excludes under dated comments — R60's carve-state rule does not apply, no splat yaml or overlays.mk was touched); never commit ROM-derived content (the pair is out of the index; check 4 now guards the class); never git clean -x; purge_set.txt never edited in this phase. Continuing with task 9.

2026-09-07 — Task 9 — Memory reconciliation + the kit's memory seed (xHigh)

The project store (~/.claude/projects/-home-musashi-bfm-decomp/memory/, outside git by Drew's decision): the off-project vantage-edge-never-leaks.md was MOVED to ~/.claude/memory-parked/ (no Vantage project store exists under ~/.claude/projects/ — only this project's and -mnt-c-Users-user, which has no memory directory; never deleted) and its index row removed. The seven stale memories got a dated "Updated 2026-09-07 (Phase 33.5)" block APPENDED (the originals stay as records): the cheap-tier A/B (→ the record; routing superseded twice), the effort doctrine (CLAUDE.md now agrees), the private-repo backup policy (HISTORICAL; R78 is the backup), the endgame deliverables (all three shipped), roadmap-to-100 (archived; plan from phase34-seed + gen3-handoff), the ROM-content policy (INVERTED — H1 in force, R74, the relaxation's bill), the context-system status block (P33.5 open / P34 next / Gen3 at 35, the R64 load order). Their MEMORY.md hooks rewritten; two more index defects fixed while there — continuous-gater- lane-plan.md had NO index row (added) and the model-ladder hook still described the dead Haiku→Sonnet→Opus ladder (now the 2026-09-01 two-tier rule). Index rows == files: 80 == 80. The kit seed decomp-architect/memory-seed/: 16 files + MEMORY.md (PA 2.0's own sixteen seeds cited by name, not copied) — the three decomp seeds (the-byte-gate-is-the-only-claim, the-matching-flywheel, mcp-reconnect-after-restart), the nine portable working agreements (offline-tooling-first, one-runbook-is-the-procedure, keep-an-accelerator-ledger, checkpoint-means-everything-is-already-in-a-file, phaseend-carries-the-narrative-axis, tool-change-ships-with-its-consumers-and-docs, no-sleep-polling-background-tasks, long-checks-run-in-the-foreground, resume-means-resume-the-run, answer-before-grinding-in-live-coop) and the riders folded into three more (a-slow-gate-is-a-bug, breadth-is-isolated-agents, route-by-measured-difficulty — the measured cliff, escalate sooner, the frontier-agent briefing recipe, what the frontier tier is NOT for). Five carry upstream: PA (the harness-generic ones: the four planned + the live co-op one). Every body de-BFM'd: no project name, owner name, paths, sessions, addresses, rule numbers or section numbers (the kit_lint pattern + Drew → 0 hits); numbers appear only as "measured on the source project". Verify: 16 files == 16 index rows; frontmatter parses (name == stem, type ∈ the four, optional upstream: PA); every seed has Why and How to apply; every index link resolves.

2026-09-07 — Task 10 — Kit part 1: README, intake, methodology, registry-E, corpus, PLACEHOLDERS (Max; S91)

Six files under decomp-architect/ (sizes at commit): README.md 8.7 KB (the three install steps — PA 2.0 first, handing it intake.decomp.md; copy the folder; "Read decomp-architect/SETUP.md and do it" — what it installs, what it does NOT (no tools; the MANIFEST is Phase-1 work until the split), the PS1/PsyQ-complete + TODO(platform) scope, the PA-2.0 pin, new-projects-only, the package tree); intake.decomp.md 20 KB (Part A: PA Mode-1's twelve items pre-answered with FILL markers; Part B: the canonical phase ladder, 11 rows (0 governance+firewall → 1 extraction+manifest → 2 oracles+load map+segmentation evidence → 3 the all-assembly baseline → 4 the compiler pinned → 5 the census + the differential harness at 0% → 6 the multipliers → 7 the codegen map + the permuter → 8 first cracks, the manual wave, the campaign → 9 publish → 10 readability), each with a machine-checkable milestone, the tools built, the G-rules in force and the DK-kernels that earned its position; Part C: the six readability inversions; Part D: what to tell PA's generation step about §E and the fail-safes); decomp-architect.md 27 KB (the methodology, 13 sections, copied to docs/ at install); templates/registry-E.decomp.md 31 KB (G1–G65 in seven groups — the oracles and the gate, the ROM firewall, instruments, the campaign, compiler walls, publishing and the record, the use of AI — each with a provenance: line); corpus/decomp-kernels.md 62 KB (DK-1…DK-64 in seven groups + the failure museum's 37 exhibits condensed and de-numbered, PA's "growing it" rule at the top, one calibration fence per kernel that carries a number); templates/PLACEHOLDERS.md 6 KB (the contract: 22 placeholders — 4 shared with PA incl. DOMAIN_FAILSAFES as referenced-only, 13 copy-time from the game interview, 5 generation-time — plus what is deliberately NOT a placeholder). Design decisions that bind tasks 11–13 (X1): (a) the lint vocabulary — §+digit and R+digits are banned outside fences/provenance, so the kit refers to PA's sections as "step N" and the kit's own SETUP.md (task 12) uses Step 0 … Step 10 headings, not the plan's literal §0–§10 (same content, a heading form the lint allows); (b) registry numbering by construction — the intake's Part D asks PA's generation step to leave §E empty except a one-line pointer, SETUP Step 7 asserts that before appending G1–G65 and documents a renumber-by-offset fallback (no {{N}} placeholder); (c) the placeholder contract — tasks 11/12 may add a placeholder only by editing PLACEHOLDERS.md in the same change; kit_lint diffs the set found by \{\{[A-Z_0-9]+\}\} over the package against the backticked set in that file; (d) the firewall config carries literal hash sources tagged required:/pending: (a missing required source fails; a missing pending one warns loudly; zero resolvable sources refuse — R32/R43) so the audit runs at Phase 0.5 before a manifest exists; (e) kit_lint must be fence-aware — the checkpoint's hand grep grep -v calibration only drops lines containing that word, not the fence CONTENTS; the honest check (used here; the spec for task 11) strips ```calibration blocks and provenance: lines, then greps. Verify: lint 0 hits over 23 files (the 6 + the 17 seed files); placeholders used 22 == listed 22; G ids 65 (1..65), DK ids 64, 0 dangling cites of either anywhere in the kit; ladder rows 11/11 each cite ≥1 resolving DK (row 10 first cited only Part C → DK-64 "readability at draft and bank time" added, the museum renumbered to part 8); doc_links --strict rc 0; wiki_render --selftest 12/12, reachability 31 pages / 0 unlisted; TODO(platform) 5, TODO(phase-N) 7. Finding, fixed here (R58/R56 class): at HEAD e1463d430 tools/doc_links.py exited 1 in BOTH modes — one BROKEN: docs/wiki/The-ROM-firewall.md line 163 cited .run/giants/fable_cd4/mine_full.txt, UNTRACKED since task 8's git rm --cached (the classifier is git-based by design, so the untrack flipped the verdict) — i.e. make tools-health was red on the docs side and the S90 checkpoint's "doc_links --strict green" was stale (task 8's verify list did not include a doc_links run; the claim was carried from task 7). Fix: the residue paragraph now names the two basenames and cites the tracked tools/public_rewrite/untracked_after_rewrite.txt; strict rc 0. Lesson for task 15's checklist: re-run doc_links --strict after ANY untrack, not only after a docs edit. Deviation from the plan's verify column (R41): the hand grep prescribed in the checkpoint is not fence-aware; the fence-aware script above was used instead and is what kit_lint.py implements. The kit's docs/ cross-references are by chapter title and kernel/rule id, never by this repository's paths, so the split can lift the folder unchanged.

2026-09-07 — Task 11 — Kit part 2: the firewall pack, READMEs, overlays, skeletons, MANIFEST, kit_lint (xHigh; S91)

Written under decomp-architect/templates/ (18 new files): gitignore.decomp (EXTRACTED by script from the wiki page's ```gitignore fence, 71 lines — gitignore_template_check now runs in tools-health and reads OK); firewall.txt (the audit's ONE config: purge:/glob: rules for the nine classes with {{TARGET_BINARY}} as the first purge path, hash sources tagged required: / pending: / fixture: — the manifest, the medium hash and the contracts are pending: with TODO(phase-1/3) promotion notes); audit_public.template.py (a de-BFM'd generalisation of tools/audit_public.py: sources from the config, four checks incl. the disassembly-shaped-content run ≥ 64, refuses zero resolvable sources / a missing required source / an unknown rule kind, warns loudly on a missing pending source, --paths for the control; compiles); firewall-fixture/ (blob.bin = the 16 synthetic bytes DECOMP-FIXTURE!!, blob.sha1 = d4bc7b5d…, README: only the .sha1 is installed into the new repo as config/firewall-fixture.sha1; the control PLANTS the blob under scratch, asserts FAIL by --paths, removes it, asserts PASS); no-rom.template.yml (the audit job; compile-only as TODO(phase-3)); docs-README.md + run-README.md (the conventions page's two tables); ops-setup.decomp.md (the decomp rows: pins with {{TOOLCHAIN_TRIPLE}}/{{COMPILER_FAMILY}}, the game/medium/dump lines, {{EXTRACT_CMD}}/{{FLEET_CHECK_CMD}}, the two oracles, the git posture, the tooling rows); bootstrap.template.sh (skeleton, every phase-gated step a TODO(phase-N); bash -n clean); CLAUDE.decomp-overlay.md (the four fail-safes; the session-start extras: the digest replaces the full PhaseEnd read, the checkpoint replayed verbatim, the oracle ping only for RE tasks, the flywheel via {{COOKBOOK_NAME}}, the effort line); pa-overlays.md (SEVEN fenced blocks, each headed by target + marker + step: the DIGEST template (CREATE), the 🛑 SESSION CHECKPOINT block for CURRENT_PHASE (APPEND, sections 0–3), the PhaseEnd narrative axis (APPEND), the decomp per-phase effort rows (APPEND), the cookbook entry shape + the symptom-keyed triage table (APPEND), the wave-playbook skeleton with an empty measurement column (CREATE), settings.json SessionStart slot + mcp.json (MERGE/CREATE)); the skeletons LICENSE.skeleton.md (replaced by the license's verbatim text), NOTICE.src.md (no license over src/; "clean-room" appears nowhere), README.skeleton.md, CONTRIBUTING.skeleton.md (the five conduct rules); .clang-format (the community style per gen3-standards: 4 spaces, 80 cols, Attach, pointer left, InsertBraces; provenance comment, not fetched) + make-format.snippet.mk (format / format-check, dotfiles excluded). tools/MANIFEST.md (45 KB): 325 tool files by ladder phase from ONE read-only Explore agent's survey (coverage asserted 325/325 of the find; P1 2 · P2 26 · P3 17 · P4 9 · P5 32 · P6 59 · P7 25 · P8 102 · P9 25 · P10 12 · project-only 16), each row = what it does + what it hard-codes, framed as Phase-N TASKS until the split; the agent's ten "unsure" placements kept as given (the agent could not write the deliverable file — read-only mode — so its inline result was transcribed; the seven TODO(platform) markers now include the manifest's). In the repo: tools/kit_lint.py (six checks with denominators: fence-aware LEAK; the PLACEHOLDERS set-diff; SYNTAX by bash -n / an in-memory compile() (the first cut used py_compile, which wrote a __pycache__ INTO the kit, then refused os.devnull) / JSON+YAML parse; the gitignore diff delegated; TODO counts; coverage — zero files refuses; --selftest = the R39 control, a planted leak + a planted unlisted placeholder must be caught while a fenced line and a provenance line must not; rc 1/2), wired into tools-health after the gitignore block; decomp-architect/README.md added to doc_links DEFAULT; the SETUP row (R21). The PLACEHOLDERS "Used in" cells reconciled against a script listing where each placeholder actually landed (6 cells corrected; no placeholder added — the 22-set held). Verify: kit_lint --selftest OK; kit_lint OK over 43 files (leak 0; placeholders used 22 == listed 22; syntax 0 failures over 3 scripts; gitignore 71 lines identical; TODO(platform) 7, TODO(phase-N) 31); the audit template py_compiles; the CI template parses; bash -n clean; doc_links --strict rc 0 (with the kit README in the set); make tools-health OK (run DETACHED via setsid nohup into .run/P33.5/tools_health_t11.log, 5,903 lines — the foreground cap is 10 min and the chain runs longer — with a background pid waiter; the first full run of the phase: sigs fresh, sdk-dual both legs, corpus/cdecl/binaries/text audits, report + audit-digest, cookbook/gccmap/roster checks, doc_links, wiki_render, gitignore check, kit_lint selftest + run, xsig 8, work_evidence, split_indicator 218/218). The report step regenerated docs/story-timeline.md (the 2026-09-07 row's commit count 43 → 48; a generated file, committed — R75). Deviations: none from the plan's file list; the checkpoint's "audit template + config/firewall.txt" landed as templates/firewall.txt + templates/audit_public.template.py + firewall-fixture/ (three files, the config separate from the code so a project edits the config only).

2026-09-07 — Plan amendment (Drew, S91): task 13.5 added

Asked whether tools/ had been audited the way docs/, .run/ and the memory store were, the measured answer was no: of 326 tool files (submodules/vendored excluded) 224 have a SETUP row, 54 a Makefile consumer, 14 CI, 34 the playbook, 4 the how-to chapters, 58 a wiki page, 210 another tool; 91 have no runtime consumer at all; tools/README.md is 1 KB; the kit's MANIFEST is an agent-transcribed table. Drew: add task 13.5 (the audit + a derived tool index + the MANIFEST re-derived), and retire only on his criterion — a tool that was required to finish the decomp is judged on its value to a future decomp: still needed, superseded (name the successor), or one-off (name what it produced). Inserted after the dry-run and before the wiki page so task 14 can link the index.

2026-09-07 — Task 12 — Kit part 3: decomp-architect/SETUP.md (Max; S91)

Written: SETUP.md (25 KB): the human header (the exact instruction; answers: <path> for unattended runs); the honesty section near the top (documents, configuration and one script; no gate/splitter/compiler/disassembler/emulator/permuter/context/ harness — those are docs/tools-manifest.md's tasks; the generation-time values as TODO(phase-N) lines; the visibility decision is the developer's); Step 0 the execution contract (KIT/ROOT, git -C "$ROOT"; plan mode = "execute Steps 1–10"; copy verbatim + Edit-tool fills, the two placeholder classes with the five generation-time literals; marked sections never edits; per-step ✓; idempotent via the Phase-0.5 CURRENT_PHASE.md; explicit-path commits, no trailers, never push, never git clean, no relative rm -rf, never under the home directory; the interview once or the answers file — a missing key STOPs, never defaults); Step 1 the PA-2.0 pin (twelve assertions incl. PhaseEnd_Phase0.md closed and exactly one cookbook), the ladder-in-constitution check with the two owner options if absent (the constitution is never edited), the §E precondition (record k), Path A only (an existing decomp repo is a STOP), the Phase-0.5 phase file from PA's template; Step 2 the interview: 3 read-backs (PROJECT_NAME from CLAUDE.md's title, COOKBOOK_NAME from the one docs/*cookbook*.md, INSTALL_DATE) + the 13 copy-time answers in one round (DUMP_PATH absolute and outside the repo, never copied) + the KEY: value answers-file contract; Step 3 the firewall (the ignore block APPENDED under a dated line, never replacing PA's; config/firewall.txt, config/firewall-fixture.sha1, tools/audit_public.py, the CI workflow; the negative control — plant the blob under scratch, the audit MUST exit 1 naming it, remove the one explicit file, the audit MUST exit 0; eight positive and five negative check-ignore probes); Step 4 layout + the two READMEs; Step 5 the bootstrap skeleton + the checkpoint commit (PA's own mid-install precedent); Step 6 the flywheel documents that outlive the package (the wave playbook from Block 6, docs/tools-manifest.md, docs/decomp-kernels.md, optionally docs/decomp-ladder.md); Step 7 the registry: the seed body from its ## §E — line INSERTED before ## §E′ (PA's heading + pointer stay above), the renumber-by-offset fallback; Step 8 the overlays by marker (CLAUDE, DIGEST create, the CURRENT_PHASE template append, the PhaseEnd template INSERT before "PhaseEnd Changelog", effort map, cookbook, ops-setup), config/decomp-hooks.snippet.json + config/mcp.json.template (Block 7 — not a live merge), the methodology copy, the skeletons with the present/absent rules (*.decomp-skeleton.md beside an existing file; LICENSE flagged for the verbatim text), .clang-format, the Makefile snippet; Step 9 the memory seed (16 files + the index rows under a marker; the memory dir asserted to be PA's .claude-state/memory; nothing under ~/.claude); Step 10 the leftover-placeholder audit (allowed hits = PA's three), the audit/bash -n/py_compile checks, G-count, gitignore the package, the sorted CREATED/APPENDED/SKIPPED manifest to .run/decomp-architect-install-manifest.txt, PhaseEnd_Phase0.5 from the template with the narrative axis, the state sweep, the git mv to logs/PhaseLog_0.5.md, the close commit by explicit path, the five-part final message, 🛑; the install manifest; troubleshooting (seven rows). Three supporting changes (X1, each a correctness fix): (1) class 7 of the firewall — ProjectArchitect keeps the agent's transcripts INSIDE the repository (.claude-state/transcripts/, committed while private under its H8) and a decomp's transcripts quote the target's disassembly (the source project purged ~260,000 lines of .s from its session archive): /.claude-state/transcripts/ added under class 7 in the wiki page's fence AND the template (re-extracted, 71 → 75 lines, gitignore_template_check OK), with !/.run/README.md under the scratch block so Step 4's README is trackable; (2) Block 7 is no longer a live merge — a SessionStart hook running a stub every session and an MCP entry pointing at nothing would only add noise/prompts, and _comment keys inside the hook/MCP objects risk the harness's schema; the block now installs config/decomp-hooks.snippet.json + config/mcp.json.template, activated at Phase 2 (the JSON bodies carry no comment keys); (3) PLACEHOLDERS.md's "Filled" wording aligned with the steps (answered at Step 2, filled at Steps 3–9; generation-time as TODO(phase-N) at Step 8). Verify: kit_lint OK over 44 files (leak 0; placeholders 22 == 22; syntax 0/3; gitignore 75 identical; TODO(platform) 7, TODO(phase-N) 38); doc_links --strict rc 0; the structural read-through as a script — 19/19 PA installer mechanisms have a counterpart (contract, plan-mode line, verbatim copy, per-step ✓, idempotent resume, explicit-path commits, no trailers, never push, one interview round, the phase file, the placeholder audit, the manifest, troubleshooting, hard stop, the delete note, the fresh session, the version pin, the write-scope guardrails, the negative control); 12/12 markers SETUP greps for exist verbatim in their templates; SETUP names all 21 non-PA placeholders and no unlisted one; every Step N reference resolves (0–10); PLACEHOLDERS' step refs (2, 8, 10) exist; 0 §+digit. Deviation from the plan's wording: §0–§10 → Step 0 … Step 10 (the lint), --answers → the answers: <path> clause of the instruction (a Claude Code prompt has no flags). No build input changed.

2026-09-07 — Task 13 — The dry-run install under D6's guardrails (xHigh; S91) — three runs, PASS

Shape: a throwaway git init under .run/P33.5/kit-dryrun/repo/ with PA 2.0 (/mnt/z/…/project-architect-2.0/) and the committed kit copied in; ONE general-purpose subagent (Opus) per run under the D6 brief (writes only under the throwaway + three deliverables; never under ~/.claude; every git call git -C "$R"; no git clean, no relative rm -rf; every question answered from answers.md, a missing key STOPs; a failed ✓ line is recorded, never patched; deliverables early — install.log as it goes, manifest.txt, verdict.json last, the final message one JSON line); the judgment mine (judge.py, tracked as evidence: the manifest SET vs expected-manifest.txt, the Step-10.1 placeholder audit, 9 + 5 check-ignore probes, G == 65, seeds == 32, MEMORY.md links, trailer-free commits, the audit in the throwaway, PhaseEnd/log/tree state, and the BEFORE/AFTER guardrails — settings and statusline sha1s, the 81 real memory files, the 25 ~/.claude entries, ghidra/ 4, dumps/*.bin 28, the real tree's dirty PATHS). Run 1 — FAIL at kit Step 3.7 (PA install clean: two trailer-free commits, the constitution with the eleven-phase ladder, §E the pointer line). The negative probe for .run/README.md succeeded (ignored): PA's own ignore block writes the directory form .run/; git never descends into an excluded directory, so the kit's /.run/* + !/.run/README.md beneath it could never re-include — and Step 3.1 forbade editing above the marker. The source project's own conventions page documents exactly this trap. Fix (SETUP): Step 3.0 rewrites that one line to /.run/* (the single named edit above a marker; Step 0.4 names it; idempotent, runs even when 3.1's marker is present); a troubleshooting row. Also from run 1: PA's placeholder audit is unscoped (it lists the package folders' own template bodies; only a gitignore-aware grep wrapper made it look clean) → Step 10.1 scoped past BOTH package folders with command grep (a wrapper that honours .gitignore silently narrows). Run 2 — resumed from Step 3 in the same throwaway (SendMessage; the fixed kit copied in): PASS 10/10, four commits; the resume path and Steps 3.1–3.5's idempotence exercised (marker SKIP; templates verified identical). Two more findings → fixes: Step 10.2's PYTHONDONTWRITEBYTECODE branch does nothing (py_compile always writes) → an in-memory compile() one-liner; Step 10.1 trips on prose that QUOTES the token (the agent's own phase file and PhaseEnd did) → the rule "a tracked document never pastes a literal double-brace; name the placeholder in prose" (SETUP + templates/docs-README.md); Step 10.6 now asks for the Phase-0.5 digest synopsis (the agent did it unasked, per the digest's own rule). decomp-architect/README.md gained "Findings for ProjectArchitect" (the .run/ form; the unscoped audit; .claude/settings.local.json hidden by a host's global git ignore) — carried upstream at the split. The judge itself had two false FAILs (R40/R57 in miniature): its placeholder grep was not scoped past the still-present PA package (the same gap it was judging), and its dirty-tree COUNT caught my own uncommitted kit fix (the agent had written nothing in the real tree) → scoped, and the guardrail compares the real tree's dirty PATH LIST before/after. Re-judged run 2: PASS, 46 == 46. Run 3 — a fresh throwaway, a fresh agent, the final kit: PASS 10/10, 0 defects, four commits (980dabd, a751a36, cead72f, 9f62ccc in the throwaway); Step 3.0 rewrote PA's line; the control FAILED on the planted blob (OFFENDER .run/firewall-control/planted.bin) then PASSED; 13/13 probes. One judge finding: the manifest had 57 entries against my expected 46 (run 2's agent had produced 46 from the same words) — a SPEC gap, not an install defect: every extra was legitimate (directories, the manifest itself, the PhaseEnd and archived log, five SKIPPED optional artifacts). Step 10.5 tightened to an exact rule (one line per FILE, sorted; CREATED incl. the manifest/PhaseEnd/log; APPENDED with markers; SKIPPED for the five optional artifacts with reasons; directories never; a moved file once under its final path — the agent's better reading); the run-3 agent regenerated under it: 53 == 53; judge PASS 20/20. Guardrails: every BEFORE/AFTER value equal across all runs (the real tree's only dirty paths were my own kit edits; the agent reported it never wrote there, and the diff agrees). Cost (R41): three agent runs ≈ 209k + 300k + 257k tokens, ≈ 19 + 12.5 + 23 min wall; the coordinator's cost was the fixture, the judge and the four SETUP fixes. Evidence tracked under .run/P33.5/kit-dryrun/ (the .gitignore block, *.py added for the judge): answers.md, expected-manifest.txt, before.txt/after.txt (run 3) + before.run2.txt/after.run2.txt, install.run1-2.log + install.log (run 3), manifest.run2.txt, manifest.run3-first.txt, manifest.txt, verdict.run2.json, verdict.json, judge.py; the throwaways repo/ (run 3) and repo-run2/ stay on disk, ignored. audit_public --paths over the 14 fixture files: OK. Deviation from the plan: run 2 was a RESUME rather than a fresh re-run (cheaper, and it proved the resume path); the fresh run 3 is the recorded proof. No KNOWN-DEFECTS.md needed. Findings for the record: (1) a governance system that keeps transcripts in the repo needs the class-7 line the kit now carries; (2) the by-contents /.run/* form must be the FIRST thing under .run/ in any ignore file — a later block cannot undo a directory rule; (3) an unattended installer's manifest must be specified as a derivable SET or two honest agents will produce two honest sizes.

2026-09-07 — Task 13.5 — The tools audit, the need-keyed index, the two dictionaries (xHigh; S91)

The census instrument: tools/tool_census.py — two enumerations of every tool file under tools/ (find vs git ls-files, the submodules/vendored/downloaded dirs excluded; they must agree — it refused once, correctly, while its own file was untracked); per tool the docstring line, its SETUP row, its CONSUMERS (Makefile/.mk, CI, the playbook, other tools by import or by name) and hence its class; the AUTHORED half in config/tool_dictionary.tsv (8 columns: path · phase · portability · need · what · adapts · status · successor_or_product; LF endings — the first write was CRLF from Python's csv default, caught by git's warning; the seed rows parsed from the transcribed MANIFEST, incl. its two-column project-only table). Coverage both ways (a new tool without a row fails --check). Generates docs/tool-index.md (KEEP-GEN, keyed "when you need to…"; Reference-index row; linked from The-matching-workflow + Toolchain-setup; one pointer line in how-to ch.04/05/10), the kit's tools/MANIFEST.md (--manifest; header states live rows, superseded pointer rows AND their total) and the two corpora (--corpus): decomp-architect/corpus/tools/<phase>/ (302 verbatim copies + 28 *.superseded.md pointers + INDEX.md) and corpus/cookbook/ (the cookbook, its symptom index, the codegen map, behind an authored front page saying what transfers per compiler). make kit-corpus regenerates; tool_census --check in tools-health asserts equality by sha1 (it caught my own edit of the census tool once — the copy was stale). kit_lint exempts the two corpus dirs from the leak/placeholder checks (verbatim evidence; 332 files counted) but syntax-checks them (318 scripts). The review (Drew's criterion): one Opus agent drafted the 93 no-consumer verdicts (55 still-needed / 28 superseded / 10 one-off) and 326 need-keys (none leaky); my verification: 0 verdict defects, every successor exists and stays live, 0 retirement candidates with a live consumer outside the set, 0 basename collisions; I read all 38 and overturned four one-offs to STILL-NEEDED — every future project needs them (the hindsight miner; the decomp.me toolchain replica; the two history-rewrite executables, the package being one unit). Final: 234 LIVE + 59 STILL-NEEDED (293 live files) + 34 retired (28 superseded, 6 one-offs) → git mv → tools/sunset/ (R100 renames, history kept); tools/sunset/README.md (the ☐ review table); the 4 SETUP rows that named only retired tools moved to a "Retired tools" section (3 rows shared with a live tool left in place, noted); docs/wiki/Archive-index.md gained the tools group (paths cited, never linked). Census after: 293 live (LIVE 232 · REFERENCED 30 · ORPHAN 31) + 34 retired. The kit: G66 (consult the tool dictionary first) + G67 (translate an inherited idiom through its pass) in registry-E.decomp.md (G1–G67 everywhere; the judge and SETUP counts updated); two memory seeds (18 at install; SETUP Step 9 = 34); SETUP Step 6 installs docs/knowledge-corpus.md and checks the manifest's row count against the manifest's own stated TOTAL; the ops-setup overlay's "two dictionaries" rows; the intake's Phase 7 cites G66/G67 and Phase 10 + Part C now name the raw-cast → declared-symbol step explicitly (Drew's question, S91); the methodology's flywheel section names the dictionaries; templates/layout-contract.md — the five-tool probe (the corpus oracle, match_one, rtu_match, harvest_verify, dedup_propagate): six assumption classes, no assumption outside them; a DRAFT for the split. README "Findings for ProjectArchitect" gained the close-step ordering nit. Run 4 (a fresh throwaway, a fresh Opus agent, the final kit): FAIL at kit Step 6 on MY check (321 != 293 — the whole-file row count against the header's LIVE figure while the superseded table exists; the file was right) → both sides derived (the header prints the total; Step 6 compares against it) → resumed from Step 6 → PASS 10/10, 4 commits, 321 == 321; judged: manifest 56 == 56, 19/20 checks OK; the one guardrail flag was docs/story-timeline.md, regenerated by my detached tools-health at its report step (commit count 48 → 53), not the agent — the dirty-path set difference over the run held exactly that one path. Evidence: .run/P33.5/kit-dryrun/{install.run3.log,manifest.run3.txt,verdict.run3.json,before/after.run3.txt} (run 3 archived) + {install.log,manifest.txt,verdict.json,before/after.txt} (run 4) + repo-run3/, repo/ on disk; .run/P33.5/tools_audit/ (all_tools.txt, noconsumer.txt, review.tsv, need.tsv, layout_probe.json, judge_run4.txt) tracked by a dated block. Verify: tool_census --check OK; kit_lint OK; doc_links --strict rc 0; wiki_render --selftest 12/12 + 31/0; audit_public OK over 6,842 tracked paths (the corpus included); make tools-health OK (detached, .run/P33.5/tools_health_t135.log, 5,911 lines — kit_lint 332 corpus files exempt; census 302 + 28 checked). Also this session: the purge probe PASS (every sampled old hash gone; the live control resolves; .run/P33.5/probe_github_s91.log) — Phase 34's gate is OPEN. Knowledge banked (R30/R31): docs/decision-log.md "P33.5 S91" (the dictionaries pivot, Drew's criterion, the measured why, the hindsight path); docs/accelerators.md "P33.5 S91" (four entries: the ignore-form wall; a typed figure vs a derived one; the judge as an instrument; a tool without a dictionary row is invisible). Cost (R41): the review agent 187k tokens; run 4 220k + 276k (the resume); the coordinator's cost was the census tool, the dictionary seed, two derived-check repairs and the retirement review. Deviations: four verdict overrides (recorded); run 4 needed a resume for a check of my own making; the ORPHAN count moved 91 → 94 → 31 as the instrument replaced the ad-hoc scan and the retirements landed.

2026-09-07 — Task 14 — The S91-b kit edits, the wiki page Start-a-new-decomp-project.md, the rows (xHigh; S92)

The kit first (decision-log "P33.5 S91-b" items (a)–(e), implemented, not re-recorded): (a) corpus/decomp-kernels.md DK-65 "Types are a banking lever and a width lever, not a byte lever" — the kernel (a type name never moves a byte, measured early and correctly; banking was the bottleneck and declaration conflicts a large share of it; width/signedness the one place a type does move bytes and the permuter cannot reach it; the discipline = a canonical type file from the first bank, one proven field at a time, widths fixed by the bytes at bank time, no duplicate definition or raw address cast banked; it shortens no crack), when it applies (Phase 6), the cost, the calibration fence (0 better / 10 same / 2 worse on the 12-function sample; ≈92% byte-correct / ≈27% banked; 1,232 definitions; 143 raw casts; the width class of the last hand cracks), provenance; the Coverage section reads "DK-64 … DK-65 — 2 kernels" and gained an "In all: DK-1 … DK-65 — 65 kernels" line. (b) intake.decomp.md row 6's milestone gains the canonical type layer (one definition per shape; widths and signedness proven by the bytes at bank time; a bank refused for a duplicate definition or a raw address cast), its tools column the type file + the bank-time check ("the type lifter and the signature reconciler start here, not at Phase 10"), its rules G62 (its bank-time clause), its earned-by DK-65; row 10's title now says it is short when Part C AND Phase 6's type layer held, earned-by gains DK-65. (c) templates/registry-E.decomp.md G62 extended (kept G1–G67, no renumber) with the bank-time paragraph and a provenance addendum. (d) config/tool_dictionary.tsv: the phase column is ONE token (tool_census.PHASES; the corpus directory derives from it), so lift_types.py and canon_sig_reconcile.py moved P10 → P6 and all five need-keys (lift_types, canon_sig_reconcile, sync_tu_decls, decl_prior, conform_decls) name the canonical type layer at Phase 6 ("again at readability (Phase 10)" where true); corpus/cookbook/README.md's "In both cases" bullet gained the type verdict sentence (a NAME never moves a byte; a WIDTH or SIGNEDNESS is the one place a type does; the permuter cannot change a type, so a width near-miss is fixed at the declaration, never by a dial). (e) decomp-architect.md part 6 gained the paragraph "Types — the two-sided verdict" (+ two fence lines); its Coverage line and the kit README read DK-1 … DK-65. Two findings fixed on the way (R75/accelerator S91-(2) class): SETUP.md Step 6 asserted grep -c '^### DK-' … = 64 — a TYPED figure — now "equals the 'In all' figure of that file's own Coverage section (both derived)"; and the kit README's install table and package tree did not name the two dictionaries or templates/layout-contract.md (added — the page describes the final kit, so the README must too). make kit-corpus → 293 tool files, 302 copies + 28 pointers, INDEX.md, MANIFEST + tool-index regenerated (git sees the two moves as renames under corpus/tools/P6/). The wiki page docs/wiki/Start-a-new-decomp-project.md (≈230 lines): the three steps from the kit README (+ the answers: clause); what it installs (a 7-row table) and what it does NOT — the two dictionaries in their place (the tool corpus keyed by need, linking the live tool index; the inherited knowledge base behind its front page; G66 named); the ladder as an 11-row table (phase · milestone in one line · earned by, with row 6 carrying the type layer); the five AI-use rules (G61–G65); the six inversions (item 3 carries DK-65 and the raw-cast → declared-symbol step); the compiler question (same family = direct use + re-prove; other = translate through the pass, G67); the accelerators as a one-line-per-entry table — 58 rows (A1–A9, the seven harness wounds, #12–#19, the journals, #20, the three S77, the two S78, the two S79, the five S80, the sixteen P32 T3/T4, the four P33.5 S91), each with the phase it was found in; the four dry-runs as a table (run · outcome · what it found); where it lives and how it is kept honest (kit_lint, tool_census --check, kit-corpus, tools-health, the split). Links: wiki pages relative; docs/ only at Reference-index targets (../tool-index.md, ../accelerators.md, ../retrospective.md, the how-to chapter); the kit by backticked path; ProjectArchitect by the URL the README already uses; .run/P33.5/kit-dryrun/judge.py cited (TRACKED). The citation trap, re-hit and fixed: the first draft cited the NEW project's docs/decomp-architect.md and doc_links read it as a path in THIS repo → BROKEN (the task-7 lesson: name a foreign path in prose, never as a backticked docs/… path). The rows: _Sidebar.md (under "Using the project") + Home.md (table row); Tools-from-this-project.md two rows (the kit; the tool index → docs/tool-index.md); README.md two bullets (the kit → decomp-architect/README.md + the wiki page; the tool index); docs/SETUP.md one row for decomp-architect/ after the tool_census row (what it is, the checks, the dry-run harness, R21); Where-the-project-goes-next.md (the probe PASSED 2026-09-07 — the gate is open, re-run once before the flip; the kit pointer → a link to the new page); docs/phase34-seed.md (task 0 PASSED with the log path — a DEFAULT doc, an untracked citation is allowed there; the 33.5 note: the kit carried WHOLE through the flip, split later as xsig was). No decision-log/accelerators entry written (both already banked in S91; nothing new was earned that the S91 entries do not carry). Verify (literal): wiki_render --selftest: 12 cases, 0 failed; reachability 32 pages, 0 unlisted · doc_links --strict: 58 documents, 451 links, 0 pending, 0 broken; citations 564 (439 tracked / 125 untracked); coverage 65 of 65; OK (2 wiki-first warnings, unchanged) · kit_lint: OK (leak 0 over 379 files; placeholders 22 == 22; syntax 0/284; gitignore 75 identical; TODO(platform) 8, TODO(phase-N) 38) · tool_census --check: 302 copies + 28 pointers, 0 gaps, OK · every DK id cited in the kit resolves (65 defined, 65 cited, 0 dangling) · audit_public --paths over the four largest touched files: OK · gitignore_template_check OK. make tools-health NOT run for this task (docs + kit only; task 15 runs it as the close's gate). No build input changed. Session note (S92): the log entry was dropped once by a stale slice index while rewriting the checkpoint and re-inserted — the grep for the entry's heading before the commit is what caught it (R66's shape: verify the artefact, not the intent).

2026-09-07 — Plan amendment 3 (Drew, S92): task 14.5 added — the record, the coverage checks, the log mining

Drew asked whether the kit holds "the whole of our experience — the tools, the cookbook, the hindsight, the PhaseEnds and logs, the rules, the memories". Measured from the kit's own provenance lines: the tools (293) and the cookbook/index/codegen map are in VERBATIM; the rules are a distillation citing 57 of the 83 R-rules (the rest mostly environment rules or ProjectArchitect's own — unchecked); the kernels cite 40 of the 53 accelerator entries, the retrospective (itself mined from the decision log's hindsight bullets, every PhaseEnd's Deviations table and the three believed/failed/sooner sections — P31–P33 only), the failure museum, the how-to and seven memories; no PhaseEnd and no phase log is cited anywhere in the kit — they reached it only through what was banked at the time (R30/R31), and S91-b is the proof that some hindsight never was; the record is absent verbatim (the 13 how-to chapters, the decision log's 79 pivots, the accelerators, the retrospective, the playbook — the kit has a skeleton with an empty measurement column —, the digest); memories 18 of 81 seeded (52 already in the docs, 9 BFM-only, 5 in PA's seed, 1 off-project). Recommended three items; Drew: "do all three, add task 14.5". Inserted after task 14 (the wiki page will describe three dictionaries), before the close. Harness task list 17 → 18 (R28).

2026-09-07 — Task 14.5 — The record as the third dictionary, the coverage check, the phase-log mining pass, dry-run 5 (xHigh; S92; three commits 0c29c9e16 / 8a698e1b7 / 8594081d8 + this one)

(1) The record — decomp-architect/corpus/record/: the 13 how-to chapters, decision-log.md, accelerators.md, retrospective.md, story.md, wave-playbook.md, effort-map.md, gen3-standards.md, gen3-handoff.md, phase-ends/DIGEST.md and every PhaseEnd_*.md (34) — 56 files VERBATIM behind an authored front page (what each is and how to read it; what is NOT there — the worklogs, R19 — and that the mining pass is their distillation; the vocabulary caveat). tool_census.py: RECORD_SOURCES + record_dest + the third corpus in plan/write/check → 358 copies + 28 pointers, --check OK; kit_lint exempts it; the kit's SETUP Step 6 gains 2c (docs/inherited-record.md, verify line extended); ops-setup "The three dictionaries" + the record row; the kit README table/tree, the methodology, the wiki page, Home, the Tools page, README, the SETUP row: two → three. (2) The coverage check — tools/kit_coverage.py (+ config/kit_coverage_map.tsv, the authored residue; a dictionary row; SETUP row; tools-health after tool_census --check): derives R1..R83 from DIGEST §3 (asserted contiguous) and the accelerator entries at numbered-item granularity (58 then, 59 now), and refuses an entry no registry/kernel provenance: line cites unless the map dispositions it (G<id> / DK-<n> / FOLDED:G<id> / ENV / PA / SEED:<file> / KIT:<path> / RECORD / COOKBOOK / NOT-PORTABLE; unknown ids refused; an un-numbered entry must be cited by its FULL group text). First run: 26 uncited rules + 21 uncited entries. Three were genuine gaps → DK-66 (a ledger's tie-break, a checker's widening and a blanket commit are part of the instrument — R70/R80/R52), DK-67 (the ignore file's directory-form wall — S91 (1)), DK-68 (a summarised signal is a claim, not ground truth — R14/R66) in a new kernels section 8 (the museum → 9; the intake's "part 8" → 9); 41 dispositions authored (rules: 15 PA, 3 ENV, folds into G6/G18, DK-22/66/68, 1 KIT template; accelerators: folds into DK-12/19/20/25/26/31/35/44/45/46/57/61, G38/G66, 1 COOKBOOK). Now 0 UNCOVERED on both populations (rules 57 cited + 26 dispositioned / 83; accelerators 42 cited + 15 dispositioned / 59). (3) The mining pass — .run/P33.5/log-mining/BRIEF.md (read-only; the "already banked?" grep protocol over every distilled record with the greps recorded; deliverable early; one JSON line last); 21 Opus slices over the 26 worklogs (batch 1 eleven small slices, batch 2 the giants split by line range: P29 ×4, P30 ×2, P31 ×3, P33) — 777 candidates considered, 634 ALREADY-BANKED (82%), 143 NEW, 30,540 lines read (of 30,510 + overlaps); harvest.py printed every candidate's cited log line VERBATIM (R14: 0 unparsable, 0 out of range; every quote read and on-topic); bank.py clustered them by theme into DK-69 … DK-80 (instrument blind spots; verdict staleness + the health suite; what earns belief; denominators/units/labels; leverage vs tractability + campaign scoping; models and prompts; the unattended run; agents and the tree; edits that keep their proofs; the search harness + the compiler as evidence; maintaining the knowledge base; hosts and services) with a GENERATED provenance: line per kernel naming every worklog line (142 banked; 1 dropped on the miner's own verdict — HARVEST_TABLE.md has all 143 with verdict and home); Coverage "In all: DK-1 … DK-80"; every DK cite resolves (80/80). docs/accelerators.md "P33.5 S92" (two accelerators: a distillation ships with a coverage check; the end-of-project worklog pass recovers the fixed-but-never-generalised — one in five here) — cited by the twelve kernels so the coverage check stays green; docs/decision-log.md "P33.5 S92" (R31: the question, the measurement, the pivot, the 82/18 why, the hindsight path). Cost (R41): 21 agents ≈ 4.3M tokens, 8–15 min each, two batches; the coordinator's cost was the brief, the two scripts, the clustering prose (~20 KB) and the provenance generation. Evidence tracked under .run/P33.5/log-mining/ (dated .gitignore block): BRIEF, 21 slice reports, HARVEST.md, HARVEST_TABLE.md, harvest.py, bank.py. (4) Dry-run 5 — Step 6 changed, so a FRESH full run (rule from task 13), not a resume: repo/ re-initialised (run 4's throwaway → repo-run4/, its evidence → *.run4.*), PA 2.0 + the final kit copied in, BRIEF-run5.md, one Opus agent: PA §1–§9 + kit Steps 0–10, 4 trailer-free commits, stopped_at: null, failed_verify_lines: []; judge PASS 20/20 (manifest 57 == 57 incl. docs/inherited-record.md; G 67; seeds 34/34; probes 9/9 + 5/5; audit OK in the throwaway; tree clean; every guardrail unchanged). Acted on from its notes: SETUP Steps 6/8 named pa-overlays.md without templates/ → $KIT/templates/pa-overlays.md; the agent's own write-scope slip (two staging files under /tmp, deleted before the verdict — outside the measured guardrails; recorded); the manifest rule's silence on a file the installer creates and deletes itself (noted, no change). Two instrument findings of my own (R40/R57): my BEFORE snapshot .strip()-ed the porcelain output and ate the first line's status space, so a deletion under the dry-run's own directory (after.txt, renamed for run 4) read as run/… and escaped the exclusion → the judge's first FAIL was the snapshot's; re-FILTERED the recorded value (not re-snapshotted); and the judge's reader crashed on an empty dirty set (a one-token line, never seen in runs 1–4) → tolerated. The second judge pass flagged exactly my six post-run edits (the wording fix + the run-5 rows) — attributed, committed as part 3a, judged again on the tree the agent left: PASS. Verify (literal): tool_census --check OK (358 copies + 28 pointers, 0 gaps) · kit_coverage OK (0 UNCOVERED ×2) · kit_lint OK (0 leaks over 436 files; 389 corpus files exempt) · doc_links --strict OK · wiki_render --selftest 12/12, 32 pages / 0 unlisted · audit_public --paths OK on the new files · judge PASS 20/20. make tools-health NOT run (task 15's gate; it now carries kit_coverage). No build input changed. Deviations: three commits for one task (parts 1/2/3a) + this close commit — each a bank, each before the next command that could touch the tree (R42); the giants' slices read 30 lines of overlap for context (counted in "lines read", hence 30,540 > 30,510); the first BEFORE snapshot's parse defect and the judge's empty-value crash are recorded above rather than hidden.

🛑 SESSION CHECKPOINT — Tasks 0–14 and 14.5 ✓; NEXT = task 15 (the close: real render, tools-health, doc_links --strict, audit_public, the P7 walk, PhaseEnd_Phase33.5.md + DIGEST §0/§2 + the log archive; Max — Tier 1, prompt Drew (R27) and WAIT for gate 2)

0. How to use this block

You are a FRESH SESSION (or this one, continuing) that has read PROJECT_CONTEXT.md, phase-ends/DIGEST.md, PhaseEnd_Phase31/32/33.md and this file, and nothing else (R64). Replay this block verbatim, state phase / done / NEXT / effort, list the rules from the digest (R1–R83), then WAIT for Drew. Rebuild the harness task list (18 rows incl. tasks 13.5 and 14.5, R28) marking tasks 0–14 and 14.5 completed and task 15 in progress. Task 15 is Max — state the Tier 1 prompt and wait for /effort max before starting it.

1. Where we are

Phase 33.5 (sub-phase; v1.32.0 → v1.32.1), gate 1 approved 2026-09-07 by Drew in plan mode at Max. Baseline HEAD 80d45b29b; task 0 = 39d524991; task 1 = a0cf302e5; task 2 = d06923a06; task 3 = 5d10a0d12; task 4 = 9970f1e62; task 5 = a0d4ae836; task 6 = 6ec4786bd; task 7 = 21c98ed5a; task 8 = ae71efe56; task 9 = caefe9872; S90 checkpoint = e1463d430; task 10 = 83de29c02; task 11 = 9235800fb; amendment 1 = 6fb2bc7f8; task 12 = 4c1d5d9f7; task 13 = a291c2b47; amendment 2 = 3307775e5; task 13.5 = 827295e24; S91-b decision-log entry = 3c4e185fe; task 14 = 9937fd7e3; amendment 3 = ecd1c2f51; task 14.5 = 0c29c9e16 (the record + kit_coverage) / 8a698e1b7 (the mining pass banked) / 8594081d8 (dry-run 5) / the commit after this block (S92, 2026-09-07). No build input has changed in this phase; the fleet is 218/218 at the Phase-33 close; the last make tools-health OK on this tree was at task 13.5 (.run/P33.5/tools_health_t135.log) — tasks 14 and 14.5 changed docs + the kit + config/ (tool_dictionary, kit_coverage_map) + tools/ (tool_census, kit_lint, the new kit_coverage) + the Makefile (kit_coverage wired into tools-health after tool_census --check), every per-tool check green (the task-14 and task-14.5 logs); task 15's tools-health is the first run that exercises kit_coverage in the chain. doc_links --strict rc 0; pending list EMPTY; kit_lint OK; tool_census --check OK; render selftest 32 pages / 0 unlisted. The kit now carries THREE dictionaries (tools 302+28, cookbook, the record 56 files — tool_census --check 358 copies), 80 kernels (DK-66–68 from the coverage check, DK-69–80 from the worklog mining pass: 777 candidates / 634 already banked / 143 new), tools/kit_coverage.py green (83 rules, 59 accelerator entries, 0 uncovered), and dry-run 5 PASS 20/20 on the final kit. Every plan task except the close is done: the wiki is the source of truth (five new pages: conventions, firewall, reference index, direction, start-a-new-project; the archive index), 59 docs + 34 tools sunset with indexes, tracked .run/ 1,086 → 868, the memory store reconciled, the kit complete with the three dictionaries, DK-1–DK-80 and the type layer (S91-b) — dry-run-proven five times. Phase 34's gate is OPEN (the purge probe PASSED 2026-09-07, .run/P33.5/probe_github_s91.log, untracked). Nothing runs in the background; the headless Ghidra MCP was never used (no RE work; nothing under ghidra/ is tracked). The plan file ~/.claude/plans/max-effort-set-plan-encapsulated-muffin.md is a copy of the "Approved plan" section below — this file is the one that counts.

2. What NEXT does (task 15, Max, Tier 1 — the plan's row 15 + the Phase Boundary Protocol) — exact steps

Prompt first (R27): "🟡 PhaseEnd creation is a Tier 1 task — confirm effort: Max is set." WAIT for /effort max. Then, in order:

1. A real render: .venv/bin/python tools/wiki_render.py --out .run/P33.5/wiki-render/ (check --help for the exact flag; the selftest is not the real run) → exit 0, every page rendered, no RenderError; record the page count (32 + 13 chapters).
2. make tools-health — it runs LONGER than the 10-minute foreground cap: launch DETACHED (setsid nohup make tools-health > .run/P33.5/tools_health_t15.log 2>&1 < /dev/null &), then wait on the pid from a background Bash (never a harness background task, never sleep-poll). Its report step regenerates docs/story-timeline.md (+ .svg) — a generated file, commit it (R75). Read the whole log: sigs fresh · sdk-dual both legs 143dbb89… · corpus 0 PHANTOM / 0 TRUNCATED · cdecl · audit-binaries 218/218 · report + audit-digest (timeline --check) · cookbook-index · gccmap_cites · roster · doc_links · wiki_render --selftest · gitignore_template_check · kit_lint --selftest + run · tool_census --check · xsig 8/8 · work_evidence · split_indicator 218/218 · the final OK line. (The plan said "foreground, ~15-min timeout"; task 11's log records why it is detached.)
3. doc_links --strict rc 0 (0 pending — the list must be EMPTY at a PhaseEnd, R80); tools/audit_public.py OK over every tracked path (the corpus included; ~6,845); git status --short clean apart from the deliverables of this task.
4. The P7 walk of the six milestone items (top of this file), each with the command and its literal output: (1) coverage ⊆ + --strict + the real render; (2) comm -3 of the Archive-index paths / docs/sunset/README.md / git ls-files docs/sunset (59 = 59) and tools/sunset/README.md vs git ls-files tools/sunset (34); git log --follow on three samples; (3) git ls-files -ci --exclude-standard .run | wc -l = 0; git status --porcelain .run | grep -c '^??' = 0; the two listings named in tools/public_rewrite/untracked_after_rewrite.txt; purge_set.txt unchanged since 80d45b29b (git diff 80d45b29b -- tools/public_rewrite/purge_set.txt empty); gate_scan.py --all --worktree PASS (73 s at task 8 — re-run or cite); (4) kit_lint OK + tool_census --check OK + kit_coverage OK + the run-5 judge verdict (python3 .run/P33.5/kit-dryrun/judge.py → PASS 20/20 on the tree the agent left; verdict.json run 5, 0 failed verify lines) — the final kit is dry-run-proven; no further run is owed unless a SETUP step changes; (5) ls ~/.claude/projects/-home-musashi-bfm-decomp/memory/*.md | wc -l == the MEMORY.md row count (80 == 80 at task 9; re-count) and decomp-architect/memory-seed/ 18 files == its index rows; (6) this item is the PhaseEnd itself.
5. phase-ends/PhaseEnd_Phase33.5.md — the constitution's format + the narrative axis (Build Log with every file/tool, the literal verification lines, the milestone; Deviations table; "What we believed, what failed, what we would do sooner"; Commit Message; Rules Added — R74–R83 ratified at gate 1, then the candidates this phase produced: (i) an ignore file's directory-form rule is a wall no later block can climb — write the by-contents form first; (ii) a check that compares a typed figure against a derived one is wrong the day the derived side moves — derive both sides (SETUP Step 6, twice: the manifest total and the DK total); (iii) a dry-run judge is an instrument — negative-control it before believing its FAILs; (iv) a tool without a dictionary row is invisible to the project that owns it — a new tool ships with its row (tool_census --check); (v) re-run doc_links --strict after ANY untrack or archive move, not only after a docs edit; (vi) an unattended installer's manifest is specified as a derivable SET, or two honest agents produce two honest sizes; (vii) a governance system that keeps transcripts in the repository needs the firewall's class-7 line (a decomp's transcripts quote the target); (viii) a foreign project's path is named in prose, never as a backticked docs/… citation (the checker reads every such citation as this repo's); (ix) a DISTILLATION ships with a coverage check against the populations it claims to distil (kit_coverage: 26 + 21 uncited on its first run, three genuine gaps); (x) the worklogs get one final read at the end of the project with an "already banked?" grep per candidate (one in five candidates was banked nowhere); (xi) a snapshot instrument never .strip()s a porcelain listing (the run-5 false flag) — plus the S91-b doctrine as a candidate if Drew wants it as a rule: types are a banking and width lever, the canonical type file from the first bank; PhaseEnd Changelog v1.32.0 → v1.32.1; the Plain-English Recap (R18/R25); 🛑 Stop Here).
6. DIGEST §0 + §2 append (the P33.5 synopsis: what it delivered, the key findings, "R74–R83 ratified at gate 1", the candidates); §0's status paragraph updated (33.5 CLOSED at v1.32.1; Phase 34 next, its gate OPEN); §4 already reflects the wiki.
7. git mv phase-ends/CURRENT_PHASE.md phase-ends/logs/Phase33.5.md (R19). **Leave the PhaseEnd + the archived log + the DIGEST edit
   • the tools-health log UNCOMMITTED for Drew's milestone-close commit (R6);** commit nothing else after step 3's generated timeline (that one goes in the task-15 prep commit BEFORE the PhaseEnd is written, so Drew's close commit carries only the close).
8. Final message: the plain-English recap + "PhaseEnd file created. Commit the file and start a new Claude Code session for the next phase." HARD STOP — do not preview Phase 34 (its seed is docs/phase34-seed.md; plan mode, Max, fresh session). P6 count: rules checks were done after tasks 4, 8 and 12; tasks 13, 13.5, 14 and 14.5 are FOUR since — task 15 opens with the check ("Rules check — re-read complete. Continuing with task 15.").

3. Standing facts for every task of this phase

• One commit per task, after this file's log line (R8/R42 form); commit by explicit path; no trailers (R5); Drew pushes (R6).
• Forward links from any new wiki page go into docs/doc_links_pending.txt as <path>\t<task N> in the SAME commit; the render refuses dead relative links (exit 1); tools/doc_links.py non-strict is in tools-health; --strict at task 15 — and after any untrack or archive move (task 10's finding). A backticked docs/… or .run/… path in a wiki page is a citation of THIS repo and must be TRACKED — name a foreign project's path in prose (task 7 and task 14 both hit this).
• Nothing moves while a wiki page / how-to chapter / doc_links DEFAULT document / tool docstring cites it — the referrer census command in the plan's D3 runs before every git mv; for TOOLS the census is tool_census.py --consumers FILE.
• purge_set.txt is never edited in this phase. Never git clean -x. The untracked 29 GB of .run/ is out of scope.
• Effort transitions are prompted and waited for (R27): Max for task 15.
• The kit's vocabulary (task 10): rules by G-id, kernels by DK-id (DK-1 … DK-65 since task 14; the kernels file states its own total and SETUP Step 6 compares against it), PA's sections as "step N", the kit's SETUP as "Step N"; source numbers only in ```calibration fences; the placeholder set is PLACEHOLDERS.md's 22 and nothing else (kit_lint enforces); the two corpus directories are verbatim evidence (lint-exempt, syntax-checked, coverage by tool_census --check); the dictionary's phase column is ONE token of tool_census.PHASES (a tool lives under exactly one corpus/tools/<phase>/).
• Any edit to a tool under tools/, to the cookbook/codegen map, or to config/tool_dictionary.tsv makes a corpus copy stale: run make kit-corpus and re-stage decomp-architect/corpus + docs/tool-index.md + decomp-architect/tools/MANIFEST.md before committing (tool_census --check in tools-health refuses otherwise; write_corpus rmtree's corpus/tools/ and rewrites it).
• make tools-health runs longer than the 10-minute foreground cap: launch it detached (setsid nohup make tools-health > .run/P33.5/<log> 2>&1 < /dev/null &), then wait on the pid from a background Bash; never as a harness background task; never sleep-poll in the foreground. Its report step regenerates docs/story-timeline.md (a generated file — commit it).
• The wiki fence and templates/gitignore.decomp are ONE source in two copies: edit the page, re-extract the template by script, gitignore_template_check must read OK (75 lines since task 12).
• The dry-run harness is reusable: .run/P33.5/kit-dryrun/{answers.md,expected-manifest.txt,judge.py} + the D6 brief (task-13 log); a kit change is re-verified by a resume on the run-4 throwaway (repo/), a fresh full run only when SETUP's steps change; the judge compares the real tree's dirty PATH SETS — attribute every foreign path (a detached tools-health regenerates the timeline) before calling it a write by the agent.

---

Approved plan (verbatim, 2026-09-07)

Phase 33.5 — Docs consolidation, the wiki as source of truth, the tracked-.run/ prune, and the day-one decomp kit

Version: v1.32.0 → v1.32.1. Gate 1 ratifies the Phase 33 candidates (a)–(j) as R74–R83 (DIGEST §3). Effort: planning at Max (set); per-task effort annotated (R7). Phase number 33.5 per ProjectArchitect's sub-phase convention (PhaseEnd_Phase33.5.md sorts after 33 and before 34 under sort -V). Phase 34 (the flip) is untouched and follows this phase; docs/phase34-seed.md stays live. Reviewed adversarially by a Plan agent (12 amendments folded in).

Context

Phase 33 closed at v1.32.0 with everything before the public flip done; Phase 34 (flip + Gen2 exit) waits on GitHub Support (ticket #4736982). Reviewing docs/ ahead of the flip, Drew found the tree is not publishable as a documentation set: ~100 authored files from 33 phases — an out-of-date study (hindsight-study.md, written mid-P29 at ~78%), ledgers whose banners say "never applied" (docs/distill/), a 216 KB raw agent-research JSON (docs/history/…output), eleven per-session tool designs, frontier analyses and worklists for sessions long closed — beside the live references. The wiki (13 pages + 13 how-to chapters) was written in Phase 33 as a summary layer over those files and links back into them; it carries no page on how docs/ or .run/ are used, no .gitignore template, no AI-conduct rules, nothing on the Gen3 standards, and no "set up a new decomp project" installer. The tracked .run/ evidence (1,086 files) is visible on GitHub as-is, and 172 of those files are tracked only by index inertia. The working knowledge that ran the project — 81 memory files under ~/.claude — is two-thirds already written into the repo and one-third (Claude-Code harness and working-agreement wisdom) nowhere in it.

Drew's charter (2026-09-07): (1) the wiki is the single source of truth; every docs/ file is KEEP / INTEGRATE / SUNSET (docs/sunset/, indexed, for Drew's final review; no information lost); in-repo docs link wiki pages, not docs/ files (README's links may stay). (2) .run/ public readiness for what git tracks only — finished work leaves tracking, post-flip inputs stay; no scan of the untracked 29 GB. (3) The wiki documents the docs/ and .run/ conventions, a .gitignore template and the rules that keep ROM content out of history forever. (4) A Claude Code installer in the ProjectArchitect style that says "install ProjectArchitect 2.0 first", then sets up an entire new decomp project for the chosen game from this project's hindsight — the optimal starting order, the ROM firewall from commit one, the flywheel, the model ladder + effort doctrine, the strict AI-use rules, and the Gen3 readability items — built in-tree now, split out after the flip as xsig was. (5) The memory store reconciled; the portable memories become the kit's seed.

Owner decisions at this gate (AskUserQuestion, 2026-09-07): kit in-tree now (split later) · Phase 33.5 / v1.32.1 / ratify (a)–(j) now · sunset = git mv → docs/sunset/ + git rm --cached of finished tracked .run/ · the kit's SETUP.md opens with "install ProjectArchitect 2.0 first, then this" · BFM's own memories stay in ~/.claude (kit seed only; no .claude-state/ retrofit) · the two ROM-disassembly files: untrack + an audit rule + a content check in audit_public, no second history rewrite (the tracked C + the pinned compiler already reproduce those bytes) · link form: relative in-tree links to docs/wiki/<Page>.md.

What exists (the three audits, condensed — the execution ledger; copied into CURRENT_PHASE.md at task 0, R30)

The wiki mechanism (audit A §D)

tools/wiki_render.py: source set = docs/wiki/*.md + docs/how-to-ai-decomp/*.md only; how-to chapters render as How-to-AI-decomp-NN-name; wiki→wiki links become page names; any other in-repo relative link becomes https://github.com/Druthulu/BFM-decomp/blob/main/<path> (tree/ for dirs, raw.githubusercontent for images); a dead relative link = RenderError, exit 1. _Sidebar.md, Home.md's table and How-to-AI-decomp.md's table are hand-written; nothing checks a page is reachable from the sidebar. tools/wiki_sync.sh deletes every *.md in the wiki clone and copies the render in (--push is Drew's; the clone only exists after the flip). tools/doc_links.py: 25 DEFAULT documents + globs docs/wiki/*.md, docs/how-to-ai-decomp/*.md; checks every relative Markdown link; docs/doc_links_pending.txt (<path>\t<task> → PENDING instead of BROKEN; --strict refuses pending; must be empty at a PhaseEnd) is the mechanism for forward links; backticked .run/ / docs/ citations are NOT links and are not checked. make tools-health runs doc_links (non-strict) and only the render selftest (fixture), not a real render. Consequence: every new page's forward links are registered in doc_links_pending.txt in the same commit; nothing moves while any wiki page / how-to chapter / DEFAULT doc links it — re-point first, move second.

docs/ classification (audit A — the classes; per-file detail in the audit table, reproduced in CURRENT_PHASE.md)

• KEEP-LIVE (45): SETUP.md, accelerators.md, actor-struct.md, community-matching-model-plan.md, concurrency-design.md (banner "not yet implemented" is false — six tools implement it), debug-menu-list.txt, decision-log.md, decompme-preset.md, disc-completeness.md, doc_links_pending.txt, effort-map.md, formats.md, gen2-mips-matching-model.md (status line stale), gen3-handoff.md, gen3-parking-lot.md, gen3-standards.md, generic-decomp-package.md, idxtab-map.md, matching-cookbook.md, matching-drafter-pipeline.md, memory-map.md, permuter-ils.md, phase34-seed.md, psyq-sample-idioms.md, psyq-worklist.md (add "COMPLETE at P31 S79" banner), public-flip-runbook.md, retrospective.md, second-oracle.md, story.md, struct-core-pivot.md, tooling-audit.md (how-to ch.04's evidence base — cite it there), verification.md, wall-taxonomy.md, wave-playbook.md, gcc-2.7.2-map/* (7), outreach/* (2), history/README.md, history/claude-bfm-starting-point.md, history/project_architect_v1.3.0.md, history/cheap-tier-ab-experiment.md.
• KEEP-GENERATED (12, regenerated by a live target): progress.md/.json/.fleet.md, difficulty.md, duplicates.md/.cross.md, backlog.md, badges/* (4), cookbook-index.md, commit-map.tsv (one-shot), disc-ledger.md, story-timeline.md/.svg (timeline.py: today wired to nothing and STALE at HEAD — it derives from git log dates; regenerate inside make report BINARY=main and --check inside make audit-digest, the progress.py --check pattern).
• FROZEN SNAPSHOTS of the closed frontier → SUNSET, never regenerated (6): frontier-atlas.md, family-hseq.md, family-cousins.md (make atlas would rewrite them EMPTY — 0 stubs), worklist.md (a July snapshot read by make audit-frontier, itself deliberately outside tools-health), autopsy.md, burndown.jsonl. Regenerating replaces the last populated snapshot with an empty document — information loss. They move to docs/sunset/generated-snapshots/ with archive rows naming the generator, the fact that a regen at HEAD renders empty, and the commit of the last populated run (via docs/commit-map.tsv). No worklist.py wiring.
• INTEGRATE-INTO-WIKI (6), then sunset the residue: gen3-standards.md §3 (AI conduct) → Contributing + the direction page; gen3-handoff.md → the direction page; hand-matching-process.md → The-matching-workflow (+ch.07); wave-metrics.md → how-to ch.09/05 (+ the .run/s37w.js execution block → wave-playbook.md); portable-decomp-workflow.md → diff against the 13 chapters, fold uniques into ch.03/05/09; hindsight-study.md → ch.03 (its §2.1 + §7; §7 is realised as tools/autopsy.py). Known referrers of hindsight: how-to 00/02, gen3-handoff §8, retrospective.md:10–11 (a DEFAULT doc — real links), two tool docstrings, backticked cites in SETUP/accelerators/permuter-ils/wave-playbook/decision-log/cookbook. The hand list is not trusted: task 6 runs the referrer-census command per file.
• SUNSET (~55 + the 6 snapshots): the 12 orphaned per-binary snapshots (progress|difficulty|duplicates.{ov_SC01_005,ov_SC01_006, ov_SC01_077,ov_SC03_001,resident}.md); tool-designs/* (11; re-point decl_from_use.py, integration_resolver.py, jtbl_carve.py first); distill/* (2 — "PROPOSED ONLY" banners are stale: both landed as cookbook §265–§269, verified; keep as the A/B evidence ledger the cookbook cites); history/Fable5-ultracode-…output, history/*link.txt (3); automation-runbook.md, calibration.md, canonical-decl-reconcile-design.md, exemplar_curriculum.md, fable5-strategy-review-2026-07.md, family-endgame-megaplan.md, family-manifest.md, frontier-p30.md, frontier-p32.md (PhaseEnd_Phase31 says "keep forever" → move, never delete), frontier-s39.md, gen2-roadmap.md, next-session-triage-ladder.md, phase25-frontier-map.md, phase25-t6-curriculum.md, phase25-t6-fable-brief.md, phase26-triage.md, proto-correspondence.md, recovery-queue-s65.md, resident-dossier.md (check each lever has a cookbook §), roadmap-to-100.md, sister-decomps.md.
• README's own docs/ links (stay): verification, matching-cookbook, gcc-2.7.2-map/, decision-log, accelerators, story, retrospective, public-flip-runbook, commit-map.tsv, gen3-standards, permuter-ils, decompme-preset, matching-drafter-pipeline, how-to-ai-decomp/00-README (+ badges/progress.json).
• Wiki coverage gaps (audit A §E): .run/ = one table cell; docs/ layout = one prose row; gitignore rules = prose, no template; accelerators = links only; portable-decomp-workflow / generic-decomp-package = one link each; gen3-standards / gen3-handoff / AI conduct = absent; no "start a new decomp" document (ch.03's 12-item checklist is the closest). SETUP.md §1a still says .run/ is "never committed" — false since P25. CLAUDE.md line 11 says "R1–R64", line 61 "since the public flip (Phase 33)" (should be "since Phase 33 C3"), and its "default working level is Max" contradicts DIGEST §1's xHigh doctrine.

Tracked .run/ (audit B — 1,086 files / 15 MB; spot-checked this session)

• 172 tracked-but-ignored (verified: git ls-files .run | git check-ignore --stdin --no-index | wc -l = 172): s6h 38, s6f 25, w6 16, s37 16, s6g 15, s6i 8, t4 6, s6b_cracks 2, fable_frontier 2, audit 1, 43 loose root files. git rm --cached is self-sealing (no .gitignore edit). Move first: wave-metrics.md:62 (.run/s37w.js execution block → wave-playbook) and :153 (.run/s6f_gate.py). Tool defaults/comments citing pruned paths get a "(untracked since P33.5)" note: wave_judge.py:8, claude_wave_packs.py:16, jtbl_family_bank.py:380, dedup_propagate.py:114.
• Allowlisted FINISHED session logs (enumerated at task 8): .run/P32/t3s3/ (67 tracked; 44 .log, 15 of them bank_*); the duplicate tools_health.log / report_close.log copies under t3s3/t4b/t4d/t4e — keep any copy a doc cites (verification.md cites .run/P32/t4e/r22_check.log); narrow their .gitignore re-includes.
• KEEP-TRACKED (~860 after the prune): P33/verify (cited by verification.md, two wiki pages, the CI header, verify_contract.sh's cleanliness assertion), P33/e1 (E1 open in Phase 34), P32/** drafts/notes/banks (PhaseEnd_Phase32: "the only copy"), giants (cookbook ×12+), S79w, s42, s43, s45, wave22, near6, fable_80178004, probe_jtbl (a project-built TU, not ROM), 7 allowlisted root files.
• FIREWALL: .run/giants/fable_cd4/{mine_full.txt,target_full.txt} (verified: 398/399 lines of 0: 27bdffc0 addiu sp,sp,-64 / 0: addiu $sp, $sp, -0x40), allowlisted by !*.txt, invisible to audit_public (path/hash/size). purge_set.txt is also gate_scan.py's census over reachable HISTORY — adding a rule there with no rewrite turns the runbook's first-push gate red permanently. So: an audit-only sibling file (D4).
• POST-FLIP-PENDING: .run/decompme/, .run/public_rewrite/, .run/objdiff/ — 0 tracked files; on-disk only, governed by phase34-seed.md tasks 2 and 7. Nothing to do.
• Doc citations of .run/: 653 backticked, 58 tracked / 581 untracked-on-disk / 14 dangling (cookbook :36918 etc.). Any disk-dependent check would go red on a fresh clone — the checker must be disk-independent (D2).

Memory store (audit C — 81 files + MEMORY.md, all outside git)

• (1) ALREADY-IN-REPO — 52. Five carry portable riders: the quadratic main-loop-context arithmetic; the accelerator-ledger habit as a rule; the per-stuck-lane briefing recipe; "a slow gate is a BUG"; what the frontier model is NOT for.
• (2) PORTABLE-NOT-IN-REPO — 14: in PA 2.0's seed already (cite): repo-self-contained-claude-state, drew-working-preferences (→ who-is-dev), dont-block-loop-with-askuserquestion, justify-new-tools-before-adopting, clarify-misconception-before-costly-action. New for the kit: no-sleep-polling-background-tasks, live-coop-answer-before-grinding, tools-health-foreground-not-background, resume-means-resumefromrunid, tool-change-updates-siblings-and-docs (+ no end-of-session audits), bank-idioms-before-checkpoint, phaseend-verbosity-for-the-retrospective, offline-tooling-first, wave-playbook-is-the-procedure; + the five riders; + three decomp seeds (matching-flywheel R16, byte-gate-is-the-only-claim, mcp-reconnect-after-restart R29). The four harness-generic ones (no-sleep-polling, tools-health-foreground, resume-means, tool-change-updates-siblings) are tagged upstream: PA so the split phase sends them to ProjectArchitect instead of burying them in a decomp kit.
• (3) BFM-SPECIFIC — 9 (keep, not seeded). (4) STALE — 7 (update in place: cheap-tier-ab-validated, effort-doctrine-xhigh-default — the memory is right, CLAUDE.md is stale —, private-repo-backup-policy, project-endgame-deliverables, roadmap-to-100, rom-content-git-policy, bfm-decomp-context-system's status block). (5) OFF-PROJECT — 1: vantage-edge-never-leaks — move it to the Vantage project's memory directory under ~/.claude/projects/ (find it with ls), never delete from a store with no history; the kit's who-is-dev interview carries a generic "cross-project export policy" line with none of its content.
• Other defects: pin count 43,857 (gen3-standards.md §4) vs 44,243 (gen3-handoff.md §3) — re-derive once, publish one number as a dated snapshot WITH its command in both docs and the wiki (the "re-derive, do not trust" form gen3-handoff already uses).

ProjectArchitect 2.0 (the model for the kit — and the constraint on it)

Package at /mnt/z/Storage/git/ProjectArchitect/project-architect-2.0/: SETUP.md (§0 execution contract; ✓ Verify per step; idempotent via CURRENT_PHASE.md checkboxes; commit by explicit path; never push; §8 fills the generation-time placeholders {{DOMAIN_FAILSAFES}}/{{DOMAIN_SESSION_START_EXTRAS}}/effort-map/cookbook/ops-setup via the Mode-1 intake, then §9 hard-stops; PROJECT_CONTEXT.md is permanent from then on, PA P1). templates/ (CLAUDE, PROJECT_CONTEXT skeleton, RULES_REGISTRY seed §A–§M

• §E project rules with provenance tags + §F per-phase, CURRENT_PHASE with a per-task checkpoint log but no verbatim SESSION CHECKPOINT block, PhaseEnd without the believed/failed/sooner axis, cookbook, effort-map, ops-setup), corpus/, memory-seed/ (16 + MEMORY.md rows - [title](file) — line; §5 verifies the count 16), statusline/, tools/backup-claude-state.sh. BFM runs the older PA 1.3.0 layout (no RULES_REGISTRY.md); the kit targets PA 2.0 and is not testable against BFM — it must therefore pin the PA version it overlays and refuse otherwise (R43).

Design

D1. Wiki information architecture (the target _Sidebar.md; every new page's sidebar row ships in the page's own commit)

• Home (rewritten; "records behind the wiki" → the Reference index).
• Using the project — the existing 10 pages. Contributing gains "AI use — conduct" (gen3-standards §3 summarised, linking the direction page); Repository-layout's docs/ and .run/ rows link the conventions page.
• Working conventions (NEW) — Docs-and-scratch-conventions.md: where each kind of knowledge goes (cookbook / decision-log / accelerators / runbook / PhaseEnd / DIGEST); authored vs generated ("numbers are generated, never typed"); the doc_links contract + pending list; docs/sunset/ + the archive index; frontier snapshots are frozen at the closed frontier; .run/: R12, the contents-exclude + ! allowlist idiom with its dated provenance comment, "commit what a rerun cannot reproduce", the per-session .run/<session>/ layout, untracked-.run/ citations are private breadcrumbs, the citation checker. · The-ROM-firewall.md: H1 with no private exemption; the nine classes of ROM-derived content; the copyable .gitignore template (a fenced block that must diff clean against the kit's templates/gitignore.decomp — one tools-health line); audit_public + the no-ROM CI + one purge set consumed by audit AND rewrite; the content check; cite-by-date; the post-purge realities; the one recorded residue (the two fable_cd4 files, in history by owner decision).
• Reference (NEW) — Reference-index.md: one row per KEEP-LIVE / KEEP-GENERATED file, grouped (environment · formats and addresses · the knowledge base · verification and progress · releases · Gen3 inputs · history/), each with what it is, how to read it, who generates it. The only sanctioned place for wiki→docs/ links; a directory row (gcc-2.7.2-map/, badges/) is a prefix. The two index pages are exempt from the rules they define.
• How to AI-decomp — 13 chapters unchanged in number; ch.03 absorbs hindsight §2.1/§7; ch.05/09 absorb wave-metrics; ch.04 cites tooling-audit; ch.11 links The-ROM-firewall; ch.03/06 attribute generic-decomp-package.
• Where the project goes next (NEW) — Where-the-project-goes-next.md: Phase 34 (from phase34-seed) and Gen3 (standards §1–§3 in summary, the census with one dated pin count + its command, the invariant, the levers, the order of work, parked ideas).
• Start a new decomp project (NEW, task 14) — Start-a-new-decomp-project.md: (1) install ProjectArchitect 2.0, handing PA's interview the kit's intake.decomp.md; (2) "Read decomp-architect/SETUP.md and do it"; what it installs and what it does NOT; the hindsight phase ladder; the AI-use rules; the Gen3 items at day one; the memory seed; the tools manifest; the accelerators as a one-line-per-entry table.
• History (NEW section) — story, timeline, retrospective (via the index), Archive-index.md (NEW): every sunset file — what it was · its outcome · where the information lives now. Rows cite the file as a backticked path + the v1.32.1 tag, never as a link into docs/sunset/ — the index must outlive the folder's deletion without turning doc_links/wiki_render red.

D2. Link policy — tools/doc_links.py extensions (task 7), wired into tools-health

• Canonical target = the wiki page, as a relative in-tree link to docs/wiki/<Page>.md.
• Refuse any link into docs/sunset/ from any document (no permitted referrer — the archive index uses backticked paths).
• Refuse a wiki/how-to link into a docs/ file not listed in Reference-index.md (the allowlist is derived from the index page with LINK_RE, resolved paths, directory rows as prefixes — R33; introduced only after tasks 4/6 exist).
• Warn (list, exit 0) on non-README in-repo docs linking a docs/ file whose topic has a wiki page (a small topic→page map).
• Backticked citations, disk-independent: classify every `.run/…` and `docs/…` citation in the checked set as TRACKED / UNTRACKED (git ls-files, never the disk); fail only when a wiki/how-to page cites an UNTRACKED path (exempt <…> placeholders); print counts with denominators for the records (R41); a local-only --disk flag adds the PRIVATE/DANGLING split for the R39 control (the 14 dangling cookbook lines are fixed once by hand). PhaseEnds and phase-ends/logs/ are records — excluded from the citation rules.
• Reachability: wiki_render.py --selftest (or doc_links) asserts every docs/wiki/*.md except _Sidebar/_Footer/Home is linked from _Sidebar.md, and every how-to chapter appears in _Sidebar.md and How-to-AI-decomp.md (0 unlisted today).
• Coverage: from task 7 on, git ls-files docs ':!docs/sunset' ':!docs/wiki' ':!docs/how-to-ai-decomp' ⊆ (Reference-index targets ∪ README targets), computed with LINK_RE over both pages (resolved paths; prefixes) — not substring grep.

D3. Sunset mechanics

Per file, in this order: (1) fold unique content into its wiki page; (2) referrer census by command — git grep -n -F '<basename>' -- '*.md' '*.py' '*.sh' Makefile ':!phase-ends/PhaseEnd_*' ':!phase-ends/logs' ':!docs/sunset' must return only Archive-index.md / docs/sunset/README.md rows; (3) git mv docs/<path> docs/sunset/<same path> (subfolders generated-snapshots/, tool-designs/, distill/, history/); (4) the archive-index row + the docs/sunset/README.md review row (file · class · ☐ reviewed). Set-equality check: comm -3 of the paths in Archive-index.md and docs/sunset/README.md vs git ls-files docs/sunset ':!docs/sunset/README.md' is empty. DIGEST §4's doc map and docs/history/README.md are refreshed in the same task. Deletion is Drew's, at Phase 34 or later; nothing in this phase deletes a document.

D4. Tracked .run/ prune (task 8)

(1) git rm -r --cached the 172 inertia paths (after task 5 moved the two wave-metrics citations); (2) the finished session logs: git rm --cached + narrow their .gitignore re-includes; (3) the two firewall files: git rm --cached, a .gitignore exclusion, and a new audit-only rules file tools/public_rewrite/untracked_after_rewrite.txt (dated provenance comment: "remains in history by owner decision 2026-09-07 — the tracked C reproduces the bytes"), read by audit_public.py check 1 alongside purge_set.txt (refuse if empty, R43); purge_set.txt stays equal to the rewritten set so gate_scan.py --all stays green; runbook §11 + The-ROM-firewall cite the file; (4) audit_public.py check 4 — CONTENT: shapes derived from the offenders (asm-differ ^\s*\d+:\s+[a-z]{2,8}(\.[a-z]+)?\s+(\$|-?0x|-?\d|[a-z_]), objdump ^\s*[0-9a-f]+:\s+[0-9a-f]{8}\s+\w+, splat /* ADDR HEX hex */, glabel/dlabel blocks); criterion = longest CONTIGUOUS run per file, threshold 64 lines; prints each file's max run with denominators; R39 controls recorded in the log: before untracking exactly the 2 files fail; after, 0 — and tools/xsig/tests/fixture_{a,b}.txt (40 lines, the project's own function) pass. Verify: the removal list is enumerated in the log entry and git diff --name-only --diff-filter=D HEAD~1 | sort equals it; git ls-files -ci --exclude-standard .run | wc -l = 0 (the real invariant); git status --porcelain .run | grep -c '^??' = 0; every removed path still exists on disk; the four tool defaults carry the note; audit_public.py OK; CI no-rom green on Drew's next push.

D5. The kit — decomp-architect/ (in-tree; mirrors PA 2.0's package shape; runs AFTER PA 2.0 as the project's Phase 0.5)

The slot. PA's constitution is permanent after PA §9 and PA §8 fills the domain placeholders during the PA install, so the kit never writes PROJECT_CONTEXT.md. Instead it ships intake.decomp.md — PA Mode 1's 12 intake items pre-answered for a decomp plus the hindsight phase ladder with machine-checkable milestones — which the developer hands to PA §8 during the PA install. The kit's own SETUP.md then runs as Phase 0.5 and does marked-section APPENDS (PA §5.1's precedent) to CLAUDE.md (decomp fail-safes + session-start extras), RULES_REGISTRY.md §E (G-numbers, each with provenance: BFM … per PA's registry rule — never BFM R-ids), docs/effort-map.md, docs/<project>-cookbook.md, docs/ops-setup.md, and appends rows to .claude-state/memory/MEMORY.md (PA §5's count 16 becomes 16+N). §1 asserts grep -q '^> \*\*Version:\*\* 2\.0' docs/project-architect.md and RULES_REGISTRY.md + phase-ends/ exist, else STOP and send the user to PA (R43); an existing decomp repository (PA Path B/C) is an explicit STOP — out of scope for this version.

decomp-architect/
  README.md                 for humans: (1) install PA 2.0 with intake.decomp.md as the interview input; (2) copy this folder;
                            (3) "Read decomp-architect/SETUP.md and do it"; what it installs; what it does NOT (see below); scope
  intake.decomp.md          PA Mode-1 intake pre-answered + the phase ladder (milestone · tools built · rules ratified · the
                            accelerator/retrospective citation that earned its position)
  SETUP.md                  §0 contract · §1 PREREQUISITES (PA 2.0 pinned; Path A only) · §2 the game interview (title, platform,
                            region/serial, dump path, container layout, SDK/compiler-era evidence, community work, goals, the
                            AI-disclosure + license decisions) · §3 the ROM firewall from commit one (gitignore, audit + its
                            planted-fixture negative control, no-rom CI, config/firewall.txt) · §4 layout + conventions READMEs ·
                            §5 toolchain bootstrap skeleton (bootstrap.template.sh; the fingerprint ladder as a Phase task) · §6 the
                            flywheel SKELETON (cookbook seed, effort map, the MANIFEST as Phase-1 tasks — the tools are NOT installed)
                            · §7 registry §E appends (G-rules, distilled R-rules, AI conduct, Gen3-at-day-one) · §8 CLAUDE/effort/
                            cookbook/ops-setup overlays + LICENSE/NOTICE/README/CONTRIBUTING skeletons · §9 memory seed + corpus ·
                            §10 verify (placeholder diff against PLACEHOLDERS.md), manifest, gitignore the package, delete-the-package
                            note, hard stop. Every step `✓ Verify`; idempotent via CURRENT_PHASE.md; `--answers <file>` unattended
                            mode ("if the file exists every question is answered from it; an unanswered question STOPs, never defaults")
  decomp-architect.md       the methodology: the ladder; the byte gate; lanes/waves; flywheel laws; models/effort as METHOD (BFM
                            numbers only inside ```calibration fences); conduct; Gen3-at-day-one; what stays hard
  templates/                gitignore.decomp · audit_public.template.py (hash sources from config/firewall.txt, R32 refusal when
                            absent) · no-rom.template.yml · firewall-fixture/ (a 16-byte synthetic blob + its sha1) ·
                            docs-README.md · run-README.md · ops-setup.decomp.md · bootstrap.template.sh · CLAUDE.decomp-overlay.md ·
                            registry-E.decomp.md · pa-overlays.md (ONE file of fenced blocks: DIGEST synopses-only template, the
                            🛑 SESSION CHECKPOINT block for CURRENT_PHASE, the PhaseEnd narrative axis, effort-map, cookbook,
                            wave-playbook, settings.json, mcp.json) · LICENSE/NOTICE/README/CONTRIBUTING skeletons · .clang-format
                            (sotn's) + a `make format` snippet · PLACEHOLDERS.md (copy-time vs generation-time; the grep is a diff)
  corpus/decomp-kernels.md  accelerators A1–A9/#12–#20/S77–P32 + the failure museum + retrospective §4, de-BFM'd; PA's "growing it" rule
  memory-seed/              the 9 new + 5 riders + 3 decomp seeds; MEMORY.md rows in PA's format; PA's 7 cited, not duplicated
  tools/MANIFEST.md         portable BFM tools by stage: what each does, BFM hard-codes flagged, "copy after the split" — Phase-1 tasks

Honesty (R41/R43): the installer verifies skeleton documents and config; it does not install a byte gate, a permuter harness or an m2c context — those are Phase-1 tasks from the MANIFEST until the split phase de-BFMs the tools. SETUP.md says so. Leak instrument: tools/kit_lint.py in tools-health — grep SLUS|Musashi|BFM|Druthulu|func_80|ov_SC|/home/musashi|/mnt/z|172\.17\.| \bR[0-9]{1,2}\b|§[0-9]+ over decomp-architect/ outside ```calibration fences and provenance: lines; the placeholder-set diff against PLACEHOLDERS.md; bash -n / py_compile; the gitignore-template diff against the wiki's fenced block. Platform scope: generic in structure, PS1/PsyQ-complete in content; other platforms get TODO(platform) markers (R43).

D6. The dry-run (task 13) — write-scope guardrails

A subagent in .run/P33.5/kit-dryrun/repo/ (a throwaway git init), with PA 2.0 copied from /mnt/z/…/project-architect-2.0/ and the kit copied in, runs PA's install then the kit's, from answers.md: fictional title + serial SLUS-99999, a dump path that does not exist, container/SDK-evidence strings, goals, license + AI-disclosure choices, PA §1 = Path A, PA §5 answers, "effort Max already set" for PA Gate 2 (logged as a deviation), pinned INSTALL_DATE, "skip PA §6" (no writes under ~/.claude), the synthetic control blob + sha1, positive AND negative git check-ignore probes (disks/x.bin, asm/x.s, extracted/x vs src/main.c, config/x.yaml), and expected-manifest.txt so "manifest present" is a diff. The brief forbids git clean anywhere and relative rm -rf; every git call is git -C <abs>. Before/after assertions by the main session: sha1sum ~/.claude/settings.json ~/.claude/statusline.sh, ls ~/.claude/projects/-home-musashi-bfm-decomp/memory | wc -l, git -C ~/bfm-decomp status --porcelain, ls ghidra | wc -l, ls dumps/*.bin | wc -l — all unchanged (R57). .gitignore gains one block: !/.run/P33.5/ /.run/P33.5/* !/.run/P33.5/kit-dryrun/ /.run/P33.5/kit-dryrun/* !…/answers.md !…/expected-manifest.txt !…/*.log (the nested repo stays ignored). Cut line: if the dry-run is still red at the end of session 3, ship decomp-architect/KNOWN-DEFECTS.md, make the re-run the split phase's first task, and the PhaseEnd says so (R66).

D7. Governance for the phase

Gate 1 ratifies (a)–(j) → R74–R83 (full text from PhaseEnd_Phase33 → DIGEST §3, the gate-1 commit); phase34-seed.md's "candidates proposed" line amended. One commit per task after the CURRENT_PHASE log line (R8/R42 form). No src//config/ build inputs change → no R22 fleet rebuild; make tools-health (foreground, ~15-min timeout) is the gate, plus a real render at task 15. P6 rules check after tasks 4, 8, 12. Forward links always via doc_links_pending.txt, emptied by task 15.

Tasks (in order; one commit each; effort per R7)

#	Task	Effort	Verify
0	Gate 1 close-out: phase-ends/CURRENT_PHASE.md (this plan as the checklist + the ledgers + a 🛑 checkpoint); R74–R83 into DIGEST §3 (+ §0 "Phase 33.5 open"); amend phase34-seed.md; TaskCreate one task per row (R28)	Max	DIGEST diff shows R74–R83; TaskList = 16 rows
1	Unblock, code side: re-point tool docstrings (decl_from_use.py, integration_resolver.py, jtbl_carve.py → cookbook §; autopsy.py, residual_class.py → the ch.03 anchor); CLAUDE.md lines 11 ("R1–R83"), 61 ("since Phase 33 C3"), the effort-default sentence → DIGEST §1's doctrine	xHigh	git grep -n 'tool-designs|hindsight-study' tools/ = 0; tools-health OK
2	Unblock, doc side: banners/status lines (concurrency-design implemented, gen2-mips-matching-model built, psyq-worklist COMPLETE, distill/* LANDED as §265–§269, hindsight-study HISTORICAL + supersession table); SETUP §1a rewritten (the allowlist idiom, what IS tracked); the pin count re-derived (one grep, one dated number + command in both gen3 docs)	xHigh	doc_links 0 broken; the pin-count command reproduces the number
3	Wiki — conventions + firewall: Docs-and-scratch-conventions.md, The-ROM-firewall.md (the gitignore template block); sidebar rows; Repository-layout/Contributing/ch.11 links; forward links registered in doc_links_pending.txt	Max	wiki_render --selftest; doc_links 0 broken (pending allowed)
4	Wiki — reference + direction: Reference-index.md (list-driven: every KEEP-ledger path has a row), Where-the-project-goes-next.md, Contributing "AI use — conduct", Home rewritten, sidebar sections. P6 rules check	xHigh	every KEEP-LIVE/GENERATED ledger path appears (script over LINK_RE); render selftest
5	Integrate the six (hand-matching-process, wave-metrics + the s37w block → wave-playbook, portable-decomp-workflow diffed vs the 13 chapters, generic-decomp-package attributed, hindsight §2.1/§7 → ch.03 + permuter-ils, gen3 docs → the direction page); re-point every referrer incl. retrospective.md:10–11	xHigh (the three diffs may fan out to agents — breadth)	doc_links 0 broken; a per-file "unique claims folded" checklist in the log
6	Sunset moves per D3 (~61 files incl. the 6 frozen snapshots); referrer census by command per file; docs/sunset/README.md; Archive-index.md rows (backticked paths + tag); DIGEST §4 + docs/history/README.md refreshed	xHigh	comm -3 empty; git log --follow on 3 samples; doc_links 0 broken
7	Checkers: doc_links.py extensions per D2; reachability in wiki_render --selftest; timeline.py regen in report BINARY=main + --check in audit-digest; the 14 dangling cookbook cites fixed (+ cookbook-index regen); the gitignore-template diff line; tools-health wiring; SETUP rows (R21)	xHigh	doc_links --strict 0 broken/0 pending/0 untracked-cited; R39 control with --disk: the pre-fix tree fails on exactly the 14; coverage ⊆ check passes
8	Tracked .run/ prune per D4 (the 172 + the session logs + the 2 firewall files; the audit-only rules file; the content check with both controls; runbook §11; the four tool notes). P6 rules check	xHigh	the D4 verify list, all recorded in the log
9	Memory reconciliation: move the off-project file; update the 7 stale in place; refresh bfm-decomp-context-system; write the seed set under decomp-architect/memory-seed/ (+ upstream: PA tags)	xHigh	MEMORY.md rows == files; frontmatter valid; kit_lint grep clean on the seeds
10	Kit part 1: README.md, intake.decomp.md, decomp-architect.md, templates/registry-E.decomp.md, corpus/decomp-kernels.md, templates/PLACEHOLDERS.md	Max	every ladder phase cites an accelerator/retrospective item; kit_lint clean
11	Kit part 2: the firewall pack (gitignore, audit template + config/firewall.txt, fixture, no-rom yml), docs/run READMEs, ops-setup, bootstrap, CLAUDE overlay, pa-overlays.md, LICENSE/NOTICE/README/CONTRIBUTING skeletons, .clang-format + format snippet, tools/MANIFEST.md; tools/kit_lint.py in tools-health	xHigh	kit_lint clean (placeholder diff, bash -n, py_compile, gitignore diff vs the wiki block)
12	Kit part 3 — SETUP.md per D5 (§0–§10, --answers, the version pin, the honesty section, Path A only). P6 rules check	Max	placeholder diff clean; a structural read-through against PA's SETUP.md
13	Dry-run install per D6; fix defects; re-run to green; the answers fixture + expected manifest + logs tracked under .run/P33.5/kit-dryrun/	xHigh	manifest diff empty; placeholder grep clean; both check-ignore probe sets; trailer-free commits; the before/after assertions unchanged
14	Wiki Start-a-new-decomp-project.md + README/Tools-from-this-project.md rows + SETUP rows for the kit and the checkers (R21) + decision-log.md P33.5 entry (R31) + accelerators.md entry if the dry-run taught one	xHigh	render selftest; doc_links --strict
15	Close: a real render to .run/P33.5/wiki-render/; make tools-health foreground; doc_links --strict; audit_public; git status clean; PhaseEnd_Phase33.5.md (Build Log, Deviations, believed/failed/sooner, rules, changelog, recap), DIGEST §0/§2 append, CURRENT_PHASE.md → phase-ends/logs/Phase33.5.md; left uncommitted for Drew (R6)	Max	every checkbox walked (P7); pending list empty; the R18/R25 recap

Scale, with its denominator (R41): 16 tasks. Calibration: Phase 33 was priced at 17–21 sessions and took 4 (59 commits); this phase is roughly half of P33's writing plus one novel item (the dry-run). Estimate 2–3 sessions; the kit (tasks 10–13) is about half the work; the cut line is D6's.

Verification (end of phase)

• make tools-health OK (foreground): the existing chain + doc_links (with D2) + render selftest with reachability + timeline --check + kit_lint + the gitignore-template diff.
• A real wiki_render.py run exits 0; doc_links --strict 0 broken / 0 pending / 0 untracked-cited from wiki pages.
• tools/audit_public.py OK incl. check 4; both R39 controls recorded (2 before / 0 after; xsig fixtures pass).
• git ls-files -ci --exclude-standard .run | wc -l = 0; git status --porcelain .run | grep -c '^??' = 0; the removal list == the diff; purge_set.txt unchanged; gate_scan.py --all --worktree still PASS.
• comm -3 archive-index / sunset README / git ls-files docs/sunset empty; the coverage ⊆ check passes.
• Kit: the dry-run's manifest diff empty and the guardrail assertions unchanged; kit_lint clean.
• Memory: MEMORY.md rows == files in ~/.claude/…/memory/; the seed has no BFM literals.
• The fleet untouched: docs/progress.json unchanged (218/218, 100/100/100), asserted by make audit-digest.

Handoff rule for the executing session

Task 0 copies the ledgers above into CURRENT_PHASE.md — the audit reports exist only in this planning session's context (R30). The 🛑 checkpoint after every task carries: the task done, the exact commands, the counts, the pending-list state, and the next single step.