BFM-decomp/phase-ends/PhaseEnd_Phase33.md

PhaseEnd — Phase 33: 100% verification + the publishing surface + the releases — everything before the flip (re-chartered close)

Date: 2026-09-07 · Project Version: 1.32.0 · Phase Status: Complete on the re-chartered milestone (confirmed by Drew, gate 2 — 2026-09-07) · Generation: Gen2 (24th phase of the arc; the flip and the Gen2 exit are Phase 34)

Gen2 phase 24 (8→…→32→33). Sixth phase off docs/roadmap-to-100.md v2 (§3 P33, all six items — Drew: "everything as written"). Ran 4 sessions (S86–S89, 2026-09-06 → 2026-09-07) and 59 commits (b5b79b3ac..b46358f27, plus Drew's close commit). The granular trail — every control, every instrument defect and its repair, the four 🛑 checkpoints — is at phase-ends/logs/Phase33.md (R19, on demand). Owner decisions (Drew): gate 1 (2026-09-06, plan mode, Max; the eleven decisions in the plan — in-place flip with the full rewritten history, the purge set, AGPL for tools/docs with no license asserted over src/, the wiki, R65–R73 ratified); F3 and E5 "in FULL" (2026-09-07); the outward pushes of xsig and the permuter fork by Claude (R6 waived for those two); no AI acknowledgement in outward text — write it the way a developer writes (2026-09-07, after the permuter maintainer closed the PR); the re-charter (2026-09-07): close Phase 33 now on what is done, the flip-gated remainder becomes Phase 34 (the last Gen2 phase, v2.0.0), Gen3 opens at Phase 35.

Build Log

The phase in one line: the 100% contract was verified as one recorded run on a public-buildable tree; the repository's entire history was rewritten to remove every game-derived byte and pushed in place with a commit map; the publishing surface, the five releases, the story, the retrospective, the wiki and the how-to were written; the Gen3 charter was set — and the flip itself waits on GitHub Support purging the old objects, which is calendar time and is now Phase 34.

Files created/changed and complete — do not recreate (category summaries; phase-ends/logs/Phase33.md and git log b5b79b3ac..HEAD hold the per-commit list; 45 new tools, 49 new documents):

• Block A — verification instruments + THE RECORDED RUN. make sig-main rewritten build-derived (tools/main_seed_ends.py; the Ghidra sig had left 3,628 words of real game code owned by no function — main game-code is 45,150 / 45,150, not 41,534); progress.py fail-loud + derived oracle line; backlog.py LINKED-aware; family_hseq.py carries its coverage; NO_SDK knob + make sdk-dual; tools/verify_contract.sh → .run/P33/verify/ (A5 on 3e0ecfacc, re-run C8 on the adopted tree 5bc4a5c0ab: every step EXIT=0, 218/218 clean fleet, sdk-dual both legs 143dbb89…, 0 PHANTOM 0 TRUNCATED 0 PAD-TAIL, UNCLAIMED 0, report 100.00 / 100.0 / 100.0, stubs 0, backlog 0); docs/verification.md.
• Block B — the public build path. make disc-extract (extract.py --expect-manifest / --allow-missing-audio; the four splat preset headers tracked; .gitignore re-tightened to H1); 34 absolute includes → ../shared/ (+ the portable-include audit class); tools/bootstrap.sh / make bootstrap + the fresh-clone proof (218/218 in 4 m 18 s, no SDK objects); tools/fetch_psyq.sh + tools/psyq_libs_from_disc.py + tools/psyq_CHECKSUMS.sha256 (20 LIBs, parser, RTL 4.2); Ghidra regenerability — ExportAnnotations.java / ImportAnnotations.java / tools/ghidra_rebuild.sh --proof, config/ghidra/*.jsonl + ROSTER.md, the six ghidra_*.sh repo-relative, ExportSymbols R15 fix (finished at medium, Drew's call); dumps/CHECKSUMS.sha1 (28/28) + INDEX.md; the ROM-free CI .github/workflows/no-rom.yml (tools/audit_public.py — the first-push gate reused as CI; tools/compile_only.py — fleet 4,170/4,170 TUs in 123 s); SETUP rows for every tool (R21).
• Block C — the history rewrite, pushed in place. tools/public_rewrite/ (purge_set.txt, gate_scan.py + fixture, hash_dict.py — 4,422 commit objects / 0 ambiguous / 0 collisions, scrub.py — 732 tokens == git's own lookup, run_filter.py, verify_rewrite.py, build_commit_map.py, resolve_tokens.py, absent_scan.py, probe_github.sh), git-filter-repo 2.47.0; two trial rewrites caught the empty-blob strip-list defect and the unchanged-commit case before the real run; the prep commit (git rm --cached ×251; audit_public first green); the bundle + Drew's mirror to the private archive Druthulu/BFM-decomp-archive; the rewrite: 4,032 → 4,031 commits, 1 pruned, 0 old hashes; verify_rewrite 4,031 pairs / 0 failures; absent_scan 0; gate_scan PASS; adoption; docs/commit-map.tsv
  • 1,238 tokens resolved to new hashes (residue 8 in 4 tokens, cited commits in no lineage); C8's R22 on the adopted tree; Drew's force-push; gc with two discoveries (the archive remote and 12 stale linked worktrees, 12 GB, pinning 3,729 old commits) → .git 1.5 GB → 93 MB, one pack. docs/public-flip-runbook.md.
• Block D — the publishing surface. README.md rewritten (generated numbers — docs/progress.json + progress.py --readme --check; Special thanks; the ProjectArchitect link; "About the history"; "Standards, and how the AI was used"; "Tools from this project"); LICENSE (AGPL-3.0 verbatim), src/NOTICE.md (no license asserted; "clean-room" appears nowhere), THIRD_PARTY.md (17 rows, each license read at its source), tools/README.md; docs/badges/*.json, tools/objdiff_report.py (validated with objdiff-cli 3.8.1), tools/frogress_upload.py, .github/workflows/progress.yml; SETUP public-clean; tools/doc_links.py (+ doc_links_pending.txt, EMPTY at close); DIGEST §1 corrections; docs/roadmap-to-100.md status block; docs/gen2-roadmap.md SUPERSEDED banner.
• Block E — the releases. docs/gcc-2.7.2-map/README.md + tools/gccmap_cites.py (135 cites tagged [2.7.2] 79 / [2.8.1 pm] 55 / [repo] 1, three instrument defects caught by its own controls); tools/xsig/ (library + CLI + MIT + a game-free two-address fixture, 8/8) and the public repo github.com/Druthulu/xsig; the permuter upstream PR (reloc-masked-scorer, the tracked patch, docs/permuter-ils.md; PR #213 + issue #214 CLOSED unmerged by the maintainer — see Deviations); docs/matching-drafter-pipeline.md; docs/decompme-preset.md + tools/decompme_replica.sh (decomp.me's exact toolchain — old-gcc 0.13 + maspsx 86ccd7d8 — rebuilt locally; func_80018F20 26/26 words BYTE-IDENTICAL; --upstream drift check); docs/outreach/archipelago.md (the note, plus the decomp's answers to the AP world's data: the day-of-week byte confirmed and our dayCounter row corrected, four patch sites + the Time-Sanity hook resolved to shared engine functions, memory-map Q#13 answered); docs/outreach/tools-announcement.md.
• Block F — the writing. tools/timeline.py → docs/story-timeline.md/.svg (72 dated rows); docs/story.md; tools/mine_hindsight.py → docs/retrospective.md; docs/wiki/ (13 pages) + docs/how-to-ai-decomp/ (13 chapters) + tools/wiki_render.py (12/12 selftest) + tools/wiki_sync.sh.
• Block G — Gen3. docs/gen3-handoff.md (the derived census: 143 raw casts · 61,898 D_ · 16,335 func_ · 44,243 pins · 1,232 struct defs · 2,220 dedup groups; the invariant; the levers; §2.1 labels vs derived structure; §2.2 the standard and the measured gap); docs/gen3-standards.md (the charter: accuracy met and stricter than the field; sotn STYLE.md adopted + four rules of ours; the use-of-AI conduct; the gap; a definition of done); docs/phase34-seed.md (the re-chartered remainder, every task with actor / effort / carrying document, the inventory, the milestone).
• Governance and memory. phase-ends/DIGEST.md §1/§3 (R65–R73 in full at A1; the H1 and roadmap bullets corrected at D5); CLAUDE.md fail-safe line "never git clean -x" (the hazard begins at the C3 commit); docs/decision-log.md P33 S86/S87/S89 entries (R31); docs/accelerators.md; the memory decomp-community-ai-standards; phase-ends/CURRENT_PHASE.md → phase-ends/logs/Phase33.md.

Tools/packages installed: git-filter-repo==2.47.0 (venv, requirements-python.txt); the two cc1 tarballs stay tracked and sha256-verified; no apt. Downloads under .run/ (scratch, regenerable): old-gcc 0.13 gcc-2.7.2-psx.tar.gz (sha256 aca64479…, the decomp.me replica), the objdiff-cli 3.8.1 validator.

Verification results (literal, at close — S89, HEAD b46358f27):

• make clean && make extract-all JOBS=16 && make check-all JOBS=16 → clean EXIT=0 · extract-all: 217 extracted, 0 failed of 217 (+ main, serial) EXIT=0 · check-all: 218 passed, 0 failed of 218 EXIT=0, 3 min wall (.run/P33/verify/s89_r22_close.log).
• make tools-health (foreground) → OK (.run/P33/verify/s89_tools_health_close.log): sigs fresh; corpus(+resident) 0 PHANTOM + 0 TRUNCATED; cdecl; audit-binaries 218/218 (R36); report lint + dedup 2,220 / 0; cookbook-index; sig-main-oracle 0 PAD-TAIL; sdk-dual; doc_links --strict 0 broken / 0 pending; wiki_render --selftest 12/12; gccmap_cites --check; xsig tests 8/8; split_indicator 218 OK.
• The recorded contract run (docs/verification.md §2, C8 on 5bc4a5c0ab): 00–08 every EXIT=0, 218/218, sdk-dual both legs 143dbb89f34491258bbc27810d0a12ec8b43a8dd, UNCLAIMED 0, report 363214/363214 · 13492113/13492113 · 5820205/5820205 · main 45150/45150, stubs 0, backlog 0. The fresh-clone proof (B3): check-all: 218 passed with no SDK objects, 4 m 18 s.
• History: verify_rewrite 4,031/4,031 pairs; absent_scan PASS (0 offenders over 112,251 blobs / 16.2 GB text + 4,034 commits); gate_scan --all --worktree PASS; tools/audit_public.py OK (0 offenders); origin/main == main after Drew's force-push; both GitHub workflows GREEN on every push (no-rom 1 m 34 s; progress 17 s; the weekly full-fleet compile 18 m 42 s).
• Not passed, by design of the re-charter: tools/public_rewrite/probe_github.sh → 31 of 33 old hashes still ALIVE (S89); the repository is still PRIVATE (gh api … --jq .private → true); GitHub Support ticket #4736982 filed 2026-09-07. The flip is Phase 34.
• Tree: clean at HEAD (git status --short empty apart from nothing — ghidra/ is untracked since C3). Ghidra MCP stopped cleanly ("Save succeeded", R23); no DB in git to commit.

Milestone achieved (confirmed by Drew, gate 2 — 2026-09-07, on the re-chartered reading): the phase's original milestone had four items; item 1 is met and re-proven today (the recorded run, sdk-dual, the oracles, the disc residue, the three 100% lines, and the C8 re-run on the adopted tree); item 3 is met — every deliverable file exists (verification.md, commit-map.tsv, config/ghidra/*.jsonl

• ROSTER.md, dumps/CHECKSUMS.sha1, decompme-preset.md, outreach/archipelago.md, gcc-2.7.2-map/README.md, tools/xsig/, the permuter patch + permuter-ils.md, matching-drafter-pipeline.md, story.md + timeline, retrospective.md, docs/wiki/ + how-to-ai-decomp/, gen3-handoff.md, and beyond the plan gen3-standards.md, phase34-seed.md, tools-announcement.md), with every outward action done (xsig public; the permuter PR/issue submitted and closed) or explicitly recorded as pending (P9: the decomp.me preset request, the Archipelago note, decomp.dev/frogress, the wiki push, the xsig topics, the fork deletion, the tools post); items 2 and 4 — the public flip and the v2.0.0 tag — are NOT met and are not redefined: they depend on GitHub Support purging the old objects, and by the owner's decision they are Phase 34's milestone, word for word. 36 of 41 plan tasks done; C10 and C11 move; G2 is this file.

Next: Phase 34 — the flip and the Gen2 exit (v1.32.0 → v2.0.0). Open a fresh session in plan mode, Max (Tier 1) from docs/phase34-seed.md — the daily probe until PASS (or the delete-and-recreate fallback past ~a week), C10, the outward halves of E1/E2/D3/F3 in Drew's own words, tool discoverability, C11, the PhaseEnd. Then Gen3 at Phase 35 from docs/gen3-handoff.md and docs/gen3-standards.md.

Roadmap delta (roadmap v2 §0 — what Phase 33 changed about the road ahead)

• P33 as written in v2 (verify + flip + Gen2 exit) is split: the verification, the rewrite, the publishing surface, the releases and the writing are DONE here; the flip and the exit are Phase 34, gated on a third party. The version plan follows: 1.32.0 here, 2.0.0 at the Phase-34 close. The two-repo mirror of gen2-roadmap.md Phase 14 stays SUPERSEDED (in-place flip, full history).
• Gen3 is chartered before Gen2 closes. Its seed is two documents, not one: gen3-handoff.md (where Gen2 ends, the census, the levers) and gen3-standards.md (the bar: sotn's style guide adopted, names only with evidence, every pin off or // !FAKE:, the 5,147 macro bodies as C, types from the code and names from observation, a definition of done). Order of work: pins → macro bodies → struct unification → names → formatting, each gated by the 218 hashes.
• Scale, honestly (R41): the gate-1 plan priced Phase 33 at 17–21 sessions; it took 4 (S86–S89) plus Phase 34's ~1–2. The writing blocks were cheap because the record they distil (PhaseEnds, the decision log, the cookbook, the logs) was complete; the one slow item is the calendar-gated purge, which no session count buys.
• Doctrine that outlives the phase: a repository's own metadata (the Activity view) publishes what its tree no longer contains — gate a flip on the objects being gone, never on a clean tree; a "read-only" probe that fetches is a writer (R57); outward text is a person's, in a developer's voice (candidate (j)); published numbers are generated, never typed.

Deviations

Item	Plan (gate-1 plan)	Actual	Reason
Phase scope at close	one phase through the flip and the v2.0.0 tag	re-chartered 2026-09-07: closes at v1.32.0 on everything before the flip; C10/C11 + the outward actions + the tag → Phase 34	the flip waits on GitHub Support (ticket #4736982; probe 31/33 ALIVE); Drew chose not to idle a phase on a third party (P5d)
Sessions	17–21	4 (S86–S89), 59 commits	the writing distilled an already-complete record; F3/E5 "in FULL" still fit
C10's gate	"old hashes 404" via the Support purge	the same gate, with a NEW reason measured in S89: GitHub's Activity view lists every pre-force-push tip (157 rows to 2026-06-11), and the old tip still served the EXE, the dumps and the archive	the "no one has the old hashes" premise was false — the host publishes them; runbook §11, rule candidate (i)
The purge probe	a read-only check	its git fetch origin <sha> imported 5.97 GiB of the purged history into the working repo (30 packs)	R57 — fixed to a throwaway scratch repo with a self-check; Drew's gc restored one 80 MB pack
B5 (Ghidra regenerability)	Max, proof on SLUS + resident	finished at medium (Drew's call): export/import/rebuild shipped, the OSGi one-bad-file lesson recorded	proportionate — the text export is the backup; the proof is regenerable
G1 order	after C11	pulled ahead (S88)	depended on nothing gated
E1 mechanics	"presets are created in-browser by a logged-in user"; decomp.me "bundles maspsx at our pin"	no create-preset UI exists — maintainers create presets from a GitHub issue template; decomp.me runs old-gcc 0.13 + maspsx 86ccd7d8 (not our 0.17 + 874855c5)	both measured from decomp.me's own repos; the toolchain delta proven text-identical on the probe by tools/decompme_replica.sh; three stale SETUP rows corrected (R14)
E5 outcome	a PR + an issue upstream	both CLOSED unmerged the same day; the maintainer: the wildcard is intentional, rename symbols in the asm, nonzero floors are fine interactively — and the issue text read as LLM-generated ("kinda similar to getting a robocall")	recorded fairly in permuter-ils.md, the decision log, failure-museum row 37; Drew's apology was well received; the scorer stays a local tool
Outward text	"plain contributor style"	that pass filtered AI tells out of model prose instead of writing like a person — Drew's decision: no AI acknowledgement; everything outward written the way a developer writes; both pending messages rewritten	rule candidate (j)
E2	a note asking for their research	the note plus the decomp's answers to their data (dow byte confirmed, patch sites → functions, Q#13 answered) and four memory-map corrections	reading their addresses against our source resolved our own open notes
Wiki timing	after the flip	briefly "can be done now" (S89) → wrong: free-plan GitHub offers wikis only on public repos	corrected within minutes from the settings page text (R14)
A2's denominator	"41,534 / 41,534" as the control value	the build-derived sig tiles 45,150 words — the Ghidra sig had left 3,628 words of game code owned by no function	a measured correction to the P31/P32 numbers; all metrics still 100%

What we believed, what failed, and what we would do sooner (for the retrospective; detail in docs/decision-log.md P33 S86–S89)

• Believed: once every document and every commit message cited new hashes, going public early was harmless. False: GitHub's Activity tab publishes every pre-force-push tip SHA, and each still resolved to the full old tree with the purged binaries one click deep. Sooner: enumerate the HOST's metadata surfaces (activity, events, PR refs, caches) in the rewrite design, not only the repository's content.
• Believed: the probe was read-only. It was a writer — a successful git fetch of an old SHA re-imported the purged history into the very repository it guarded (R57). Sooner: any check that fetches, clones or builds does so in a throwaway.
• Believed: removing AI tells from model prose made an upstream PR acceptable. It made it worse — the maintainer recognised it anyway and said why it cost him. Drew's diagnosis, shared: the dense code-span formatting and the exhaustive structure were the giveaway; the fix is writing the way a person writes, not filtering. Sooner: the owner writes every outward word from a facts list, and the target project's AI policy is read first (sotn-decomp forbids autonomous-tool PRs outright).
• Believed: decomp.me ran our toolchain and let a user create a preset. Both wrong, measured from its repositories; the preset is requested from maintainers, and their compiler build differs from ours in version — proven equivalent on the probe only by rebuilding their exact toolchain locally. Sooner: read the service's build recipe before writing the doc that depends on it.
• Believed: the phase would take 17–21 sessions. Four. The record was complete enough that every writing task was distillation. Cost: the phase's real expense was the two outward missteps (a maintainer's time and a closed PR; a 6 GB re-import) — both cheap to repair, both avoidable by the two rules above.

Commit Message

(Phase 33 landed as 59 commits b5b79b3ac..b46358f27, 2026-09-06 → 2026-09-07 (S86–S89). This close = the archived log +
this PhaseEnd + the DIGEST update + the two close logs, Drew's milestone-close commit.)

chore(phase-33): CLOSE (re-chartered) — verified, rewritten, published-ready: the contract run, the in-place history rewrite,
LICENSE/NOTICE/README, the releases, story/retrospective/wiki/how-to, the Gen3 charter; the flip is Phase 34 (v1.32.0)

- PhaseEnd_Phase33.md written; CURRENT_PHASE.md archived to phase-ends/logs/Phase33.md (R19); DIGEST §0/§2/§3 appended (step 3b)
- 4 sessions (S86–S89): A1–A5 (the recorded run; main game-code 45,150/45,150), B1–B9 (public build path, Ghidra text export, CI),
  C1–C9 (filter-repo rewrite: 4,031 pairs / 0 failures, commit-map.tsv, force-pushed in place, .git 93 MB), D1–D5, E1–E6, F1–F3, G1
- close R22 on b46358f27: extract-all 217/217 (+main), check-all 218 passed / 0 failed (exits 0/0/0); tools-health OK
- re-charter (Drew): the flip-gated remainder (probe → C10 → outward E1/E2/D3/F3 → discoverability → C11 → v2.0.0) is Phase 34
  (docs/phase34-seed.md); Gen3 = Phase 35 from gen3-handoff.md + gen3-standards.md
- rules: R65–R73 ratified at gate 1; candidates (a)–(j) proposed for Phase 34 gate 1
- v1.31.0 -> v1.32.0

Rules Added This Phase

Ratified at gate 1 (2026-09-06): R65–R73 (the PhaseEnd_Phase32 candidates (i)–(ix)); full text in phase-ends/DIGEST.md §3.

Candidates recorded in the log, for Drew to accept, modify or reject at Phase 34 gate 1 (P10):

Rule	Reason
(a) No ROM-derived bytes in ANY published artifact — test fixtures, JSON, badges, reports, wiki pages included.	The xsig fixture was built from our own C at two link addresses for exactly this reason; the regenerated target .s stays under the ignored .run/.
(b) Published numbers are generated, never typed (R51 applied to documents).	docs/progress.json, the README block, the badges, the timeline: every hand-typed number in the old README was stale.
(c) Rehearse every irreversible repository operation on a scratch copy and prove it pair by pair with POSITIVE assertions.	Trial rewrite #1 exposed the empty-blob strip-list defect and the unchanged-commit case; the real run had 0 failures.
(d) A linked worktree's HEAD is a ref — audit git worktree list before any gc or purge.	12 stale worktrees (12 GB) pinned 3,729 old commits through the gc.
(e) R20 amendment: the text export + checksums + the archive repository are the backup; never git clean -x.	The purged paths are ignored-but-present; a -x clean deletes the RE database (CLAUDE.md fail-safe since C3).
(f) pkill -f never with a literal the calling shell's own command line contains.	S87 killed its own shell twice (exit 144).
(g) A checker that widens its document set must classify a missing promised page as PENDING, never BROKEN.	doc_links + doc_links_pending.txt: 10 planned pages mid-F3, 0 false failures.
(h) A probe or guard must never write into the repository it guards — a "read-only" check that fetches, clones or builds does so in a throwaway (R57 exemplar).	The purge probe re-imported 5.97 GiB of the purged history on every run.
(i) A rewritten history is not private until the objects are gone from the HOST — its Activity view publishes every pre-force-push tip; gate a flip on the probe (or recreate the repository), never on a clean tree.	157 activity rows to 2026-06-11; the old tip still served the EXE, the dumps and the archive.
(j) Outward text to third parties is written by Drew the way a developer writes — short, plain, few or no code spans, no bullet walls, no evidence scaffolding — rewritten from the facts, never a model draft with the tells removed; no AI acknowledgement; the target project's AI-contribution policy is read first.	The permuter maintainer closed #213/#214 and named the LLM-looking text as the reason he minded; sotn-decomp forbids autonomous-tool PRs outright.

PhaseEnd Changelog

v1.31.0 → v1.32.0 — Phase 33 complete on its re-chartered milestone (Gen2 phase 24). The 100% contract was verified as one recorded run (218 of 218 from a clean rebuild, both SDK legs, every oracle green, the three metrics at 100% with main's game-code denominator corrected to 45,150) on a tree a stranger can build from their own disc (the fresh-clone proof, make disc-extract, make bootstrap, the optional SDK fetch, the Ghidra database regenerable from text). The repository's entire history was rewritten with git filter-repo to remove every game-derived byte — 4,031 commits verified pair by pair, one pruned, every old hash replaced by an inert token and mapped in docs/commit-map.tsv — force-pushed in place, and shrunk from 1.5 GB to 93 MB. The publishing surface (README with generated numbers, AGPL for tools and docs with no license asserted over src/, NOTICE, THIRD_PARTY, badges, objdiff/frogress publishing, a ROM-free CI), the five releases (the gcc 2.7.2 codegen map, xsig, the permuter driver, the drafter write-up, the decomp.me preset proven through decomp.me's own toolchain), the outreach notes, the story, the retrospective, the wiki and the thirteen-chapter how-to were written. The Gen3 charter was set: accuracy already stricter than the field, readability to sotn's standard with evidence-based names and no match-forcing tricks, and conduct rules for using AI that keep the project apart from low-effort machine decomps. The flip itself waits on GitHub Support and is Phase 34, which closes Gen2 at v2.0.0; Gen3 opens at Phase 35.

Plain-English Recap

This phase took a finished-but-private translation and made it ready to publish. First we proved the whole claim one more time, end to end, on a computer set up the way a stranger would set it up: every one of the 218 files rebuilds byte for byte from our source and a copy of the game disc. Then we cleaned the repository's history: the old commits contained the game's executable, memory dumps, Sony's development kit and our own transcripts, none of which may be published, so we rewrote all four thousand commits to remove them, checked every rewritten commit against its original, and replaced the repository on GitHub with the clean version. We wrote everything a reader needs: a README with numbers generated by the tools rather than typed, the licenses, a thirteen-chapter guide to how a project like this is run with an AI agent, the story of the project and an honest retrospective, and a wiki. We packaged the tools worth sharing, and one of them was offered to the maintainer of a widely used tool, who declined it and, fairly, objected to the machine-written wording of the request; that taught us a rule we now hold. Finally we wrote down the standard the next generation must meet: the code matches perfectly, but it still reads like machine output, and turning it into code a human maintainer would call finished is the work ahead. The one thing this phase could not do is flip the switch to public: GitHub has to purge the old data first, and that is in their hands, so it becomes its own short phase.

🛑 Stop Here

PhaseEnd written; CURRENT_PHASE.md archived → phase-ends/logs/Phase33.md (R19, via git mv); phase-ends/DIGEST.md §0/§2/§3 appended (step 3b); the two close logs moved into .run/P33/verify/ — all left uncommitted for Drew's milestone-close commit (R6 gate-2 sign-off; the message is above). The Phase-33 work is committed across the sessions (b5b79b3ac..b46358f27); Drew commits AND pushes this close. The headless Ghidra MCP was stopped cleanly ("Save succeeded", R23); nothing under ghidra/ is tracked. Gen2 continues — do NOT start Phase 34 here. Start a fresh session (effort Max, plan mode) for Phase 34 — the flip and the Gen2 exit, opening from docs/phase34-seed.md. Keep this file forever.