Sylpheed/docs/agents/CONSOLIDATION.md

Consolidating six repositories into two

Set by the human, 2026-09-13. The end state is two repositories:

repo	holds
Sylpheed	everything — the port, the RE corpus, every tool
Xenia-Canary (our fork)	the emulator and oracle, with our instrumentation

Everything else is retired. This page is the ordered how, with a gate after each phase, and a ledger of what was checked rather than assumed.

⚠️ Most paths on this page do not resolve in this repository, on purpose. They name files in xenia-rs, xex2tractor, xenia-canary and the project root — the things being moved in. A citation checker pointed at docs/agents/ will flag every one of them until the phases below are done, and that is the correct reading: each unresolved path here is work outstanding.

📌 Two decisions from the human frame the whole plan:

"The emulator is dead and can be dropped altogether. But the DuckDB generator should still be a useful tool… We probably can move it into a standalone tool under Sylpheed."

"xex2tractor was an early attempt and is long dead too… you may want to scan and compare the tools for anything that is in xex2tractor only."

---

Where things stand

repo	server	local tree	tracked	disposition
Sylpheed	598 MB	33 GB	131 MB / 1470 files	the target
Xenia-Canary	71 MB	20 GB	—	the keeper — ⚠️ dirty
Syplheed-Reborn	103 MB	1.4 GB	100 MB / 787 files	✅ absorbed → archive
Sylpheed-Godot	0.4 MB	99 MB	0.3 MB / 39 files	✅ absorbed → archive
xenia-rs	9 MB	1.2 GB	61 MB	harvest the DB tool → archive
xex2tractor	0.3 MB	no clone	—	harvest four things → archive

Sylpheed has two root commits — feat: initialise workspace (2026-03-25, Reborn's) and scaffold the Godot port (2026-08-28). Both histories were grafted in, not copied, which is why the absorption checks below pass.

The ledger — what was verified, and how

Nothing here is inference. Each row is a command that was run on 2026-09-13.

claim	method	result
Reborn fully absorbed	all 21 remote + 8 local branch tips vs Sylpheed/main	every one an ancestor
Sylpheed-Godot absorbed	both branch tips, + the root-commit graft	every one an ancestor
agent-backups/ redundant	git bundle list-heads, 23 refs across 4 bundles	all present in Sylpheed
3 unpushed pi/* branches redundant	git patch-id --stable, per commit	all 4 patches on main
nothing lives outside git	issues / PRs / releases / wikis, all six repos	0 everywhere but Sylpheed
the DB tool is separable	git grep "use xenia_cpu::" in xenia-analysis	three symbols

🔴 What the ledger found that a naive plan would have destroyed

1. The F1 instrument exists in one place, uncommitted. xenia-canary has two dirty files — hid/file/file_input_driver.h (--pad_file_repeat) and app/xenia_main.cc — plus 4 unpushed commits on sylpheed-re and a branch, audit-handle-lifecycle-probes, contained in no remote at all. That patch is the instrument behind REPEAT_DELAY/REPEAT_INTERVAL, now shipped in the port. Its own finding page says "not yet upstreamed."
2. xenia-rs has 4 unpushed branches — iterate-4B/ui-perf, 4C/jit, 4D/parallel, 4E/native-threads. Dropping the emulator is a decision; dropping work that was never backed up is a different act.
3. The tool is not on master. xenia-rs/master last moved 2026-06-15; the DuckDB and disassembler work is on origin/iterate-4A/apu-xma-stage1 (2026-09-10, +22 commits). Extract from the default branch and you get a three-month-old tool.

---

Phase 0 — Secure the single-copy work ✅ DONE

Before touching anything else. One git checkout in xenia-canary ends the reproducibility of a measurement that is already live in the port.

1. Commit file_input_driver.h + xenia_main.cc on sylpheed-re.
2. Push sylpheed-re (4 + 1 commits) and audit-handle-lifecycle-probes.
3. Resolve the third_party/DirectXShaderCompiler submodule drift. ✅ There is none. git submodule status shows no +: the checked-out commit matches the recorded one. The m in git status is nested submodule dirt — DXC's own external/SPIRV-Headers and SPIRV-Tools — which is upstream's business and must not be committed here.

Gate — and the first version of it was wrong. It compared each local branch to a remote by SHA, and reported seven safe branches as unpushed: a branch that is an ancestor of a pushed branch is already preserved. A check that goes red for something unactionable is the failure this repo keeps naming, so the gate tests what it means — reachability:

git for-each-ref --format='%(refname:short)' refs/heads | while read b; do
  s=$(git rev-parse "$b"); found=""
  for r in $(git for-each-ref --format='%(refname:short)' refs/remotes); do
    git merge-base --is-ancestor "$s" "$r" 2>/dev/null && { found=$r; break; }
  done
  [ -n "$found" ] || echo "🔴 $b NOT REACHABLE from any remote"
done

✅ PASSED 2026-09-13 — 8e63a9542 commits the instrument; sylpheed-re and audit-handle-lifecycle-probes pushed; all 14 local branches reachable.

Phase 1 — Make the archive honest ✅ DONE

Push xenia-rs iterate-4B/ui-perf, 4C/jit, 4D/parallel, 4E/native-threads.

They are being retired, not rescued. But a retired branch that was never pushed is not retired — it is deleted, quietly, by someone who thought it was safe.

Gate: the same loop, run in every repo, prints nothing.

✅ PASSED 2026-09-13 — all four pushed (32 / 52 / 56 / 59 commits ahead of master). The loop is clean in all five repos except pi/clippy, pi/clippy-clean and pi/reauth3 in Sylpheed, which are unreachable by ancestry and redundant by patch-id — the ledger's reason, and Phase 7's disposal. The gate is right to flag them; they are the one case where the answer lives outside it.

Phase 2 — Harvest xex2tractor, then retire it ✅ DONE (PR #30)

Four things exist only there.

take	why
XEX2 devkit + XEX1 retail master keys	xenia-xex hardcodes retail only; devkit is present but #[allow(dead_code)] and XEX1 is absent. xex2tractor tries all three in a validating loop.
extract -r	🔴 RUN, AND THE INFERENCE WAS WRONG — see the gate below. Optional, and dangerous if it is ever mistaken for the .pe.
doc/xex2_format.md (39 KB) and doc/xbox360_exports.json (938 KB, 2,913 exports across xboxkrnl / xam / xbdm)	byte-identical to untracked loose copies in the project root. Tracked by no repo.
LICENSE (MIT)	Sylpheed has none.

Leave everything else: xenia-xex is ahead on resources.rs (XDBF/XACH), pdata.rs, tls.rs, basic zero-fill compression, and the analysis layer.

Gate — ✅ RUN 2026-09-13, and it corrected this page in two directions.

extract      (xex2tractor, no -r)    ✅ IDENTICAL to the project .pe
extract -r   (xex2tractor)           ❌ differs — 1,903 bytes over 273 runs
extract      (xenia-rs, iterate-4A)  ✅ IDENTICAL — and .xex.json identical too

Right: the .pe is an xex2tractor extract. Hash-for-hash.

Wrong: it was made without -r. The project's .pe is the plain decrypted-and-decompressed image, imports unresolved — and -r's 1,903 differing bytes sit in the import thunk region from VA 0x82000600.

🔴 So the dependency this gate existed to protect does not exist. This page claimed "drop this and the file every .pe-offset citation depends on can never be rebuilt." It can: xenia-rs extract, the tool Phase 3 keeps, reproduces it byte-for-byte — along with the loose .xex.json, which turns out to have come from xenia-rs, not xex2tractor.

⚠️ And -r must never become the .pe. It writes into the image, so a .pe built with it would silently invalidate every byte-offset citation in docs/re/. If resolved-import disassembly is ever wanted, it is a separate artifact under a different name, never a replacement.

That leaves xex2tractor's unique list at three items, none urgent — the two master keys, the two doc assets, and the licence. The doc assets are now the only reason to visit that repository.

📌 Learned in passing: iterate-4A builds clean (xenia-app, one unused variable warning), so Phase 3's source is not bit-rotted — and it needs libavutil-dev and friends via ffmpeg-sys-next, dragged in by the emulator's XMA audio. The standalone tool sheds a C library the CI image did not have.

Phase 3 — Lift the DB tool out of the emulator ✅ DONE (PR #32)

From origin/iterate-4A/apu-xma-stage1, never from master.

xenia-analysis          10,658   db, vtables, jumptables, func, xref, rtti,
                                 strings, xdbf, static_init, demangle, sql_views
disasm slice of cpu      3,681   disasm.rs + decoder.rs + opcode.rs  (of 21,813)
xenia-xex                2,018   header, loader, lzx, pe, pdata, tls, resources
xenia-vfs                  444   so it reads an ISO, not just a loose .xex
xenia-types                361
                       ───────
                       ≈17,200   of ~71,000

The coupling is exactly three symbols — decoder::decode, disasm::DisasmItem, disasm::format. Interpreter, JIT, scheduler, VMX, block cache, reservations: none of it is reachable from the generator.

xenia-xex declares xenia-memory and xenia-types as dependencies and has zero use of either. Probably droppable — confirm by compiling, do not assume.

Travelling with it: zq.py, tests/db_schema_golden.rs, tests/disasm_goldens.rs, examples/decode_table_check.rs.

Two corrections to make during the move, not after:

• the CLI help says "SQLite database" in two places. It is DuckDB (duckdb = { workspace = true }).
• zq.py's documented escape hatch — "the engine vtable / rdata is NOT in the DB … read it from guest memory with xenia-rs exec --dump-addr" — dies with the emulator. Replace it with Canary's dump or a direct read of the .pe at VA − 0x82000000, and say which, in the file.

Gate — ✅ PASSED 2026-09-13. Schema a superset (22 new tables, none lost). instructions 1,865,751, pdata_entries, imports, sections and all three eh_* identical to the row; most others slightly higher. Two tables drop ~1.8 M and it is not a loss — the xref delta is entirely ind_call (everything else nets +1,512), and the new schema records why: 6,556 sites truncated=true carrying 1,801,075 candidates against 427 untruncated producing 4,199 rows. The old database materialised a cross product.

Phase 4 — Prune the captures, and give docs/re/ the check docs/port/ has ✅ DONE (PR #33)

docs/re/captures/ is 257 files / 117.6 MB, and it is not junk. Measured:

193  (81.2 MB)  cited by a docs/re page, or opened by a tool
 64  (36.5 MB)  referenced by nothing

🔴 Two are test fixtures and cargo test opens them:

crates/sylpheed-formats/tests/ui_paint_order_disc.rs  → captures/title-screen-oracle.png
crates/sylpheed-formats/tests/unit_layout_disc.rs     → captures/stage02-live-unit-definitions-deep.txt

Sixteen more tools under tools/port/ and tools/re-capture/ read from there.

Do:

1. ✅ DONE — tools/re/check-capture-citations, the mirror of tools/port/check-citations, which has only ever checked docs/port/*.md. It reports two classes — a capture nothing cites, and a page citing a capture that was never committed. It needs a --selftest that plants one of each, or it is not a check.
2. ✅ DONE — and there was 10 ONE. Eight of the ten were directory references that resolve; one was sentence punctuation in the regex. The real one, live-submenu-unidentified.png, had been committed and then deleted on 2026-08-30 while the page kept citing it and kept making the claim it backs. Restored, not de-linked.
3. ✅ DONE — 46, not 64. Two are opened by filename from screen_match.py, and sixteen sit inside directories that pages cite as directories. 30.7 MB out of the checkout.

⚠️ An orphan is not automatically waste. It may be evidence a page should have cited, and the page is the thing that is wrong. Fix the citations before deleting, or step 3 silently ratifies every omission in step 2.

📌 Going forward the class that matters is not captures-vs-code:

evidence — a screenshot a finding points at	belongs beside the finding, as the convention says
run output — .jsonl, .log, .csv	the same class as xenia-rs's audit-runs/, which this plan drops

The two .jsonl mission-state dumps alone are 4.3 MB, and one is an orphan.

🔴 A gate that walks refs/heads cannot see a working tree

Found 2026-09-14, on the other machine, while pulling the repos. xenia-rs had 468 uncommitted lines plus an untracked 338-line imports.rs sitting in crates/xenia-analysis — the very crate Phase 3 lifted. Written 2026-09-10 21:01–21:23; found four days later.

It existed in exactly one place, and that place was not git:

	imports.rs	import_address	zq.py impcalls
PR #32, the lifted crate	✗	✗	✗
iterate-4A (committed)	✗	✗	✗
xenia-rs working tree	✓	✓ (4 files)	✓

Phase 0 is titled "Secure the single-copy work" and its gate passed anyway, because the gate tests branch reachability — it iterates refs/heads. A dirty working tree has no ref. Canary's two dirty files were secured only because a human already knew they were there; nothing found them. Phase 1 then pushed xenia-rs's four branches and called the archive honest while this sat beside them.

Phase 5 ends with "delete the local clone" and Phase 7 drops the emulator, so this had a deletion scheduled against it.

The gate needs a second half. Before archiving or deleting anything:

for d in Sylpheed sylpheed-reborn xenia-rs xenia-canary; do
  n=$(git -C "$d" status --porcelain | wc -l)
  [ "$n" = 0 ] || echo "🔴 $d has $n uncommitted path(s)"
done

Run 2026-09-14 across all four: only xenia-canary is dirty, and both entries (build-cross/, vkd3d-proton.cache) are already on Phase 7's drop list.

✅ Secured — xenia-rs branch harvest/import-thunk-naming @ b4f19f1, pushed. Committed exactly as found, not cleaned up: editing it first would have destroyed the thing being preserved. cargo check -p xenia-analysis = 0.

✅ Ported — PR (new) feat/xexdb-import-naming, stacked on #32. The lift's base is exactly the harvest's parent, so every file merged three-way: 10 conflicts, all of them caused by the lift's reformatting pass — 9 are the harvest's additions landing inside blocks rustfmt had rewrapped (so the content is kept and re-indented to the lift's style), and 1 is genuinely semantic: insertion order, because xdbf_achievements.image_id is now an FK onto xdbf_images(id), so the image rows must be inserted first. The conflict there was the stale copy of the block left at its old position. cargo test -p sylpheed-xexdb 10/0, clippy and fmt clean.

📌 Two things it carries, both absent from #32 as lifted: import-thunk recognition (disassembly says xboxkrnl.exe::RtlEnterCriticalSection instead of .long 0x01010194), and 16 schema-wide foreign keys — with functions deliberately excluded and the golden test now asserting zero inbound FKs onto it, since DuckDB's UPDATE is delete+insert and one inbound FK would make functions.name un-updatable and break apply_re_symbols.sql.

Phase 3 residue found while porting

• ✅ tools/zq.py hardcoded /home/fabi/…/xenia-rs/sylpheed.db — a default path inside the repository Phase 5 archives and Phase 7 drops, resolving on exactly one machine and days from becoming a duckdb exception with nothing saying why. Fixed in PR #35 (38cc170): $SYLPH_XEXDB → <repo root>/sylpheed.db → a refusal naming both, with the repo-root path taken relative to the script rather than the caller's cwd. No fallback beyond that, deliberately — the database is an untracked build artefact of a few hundred MB, so there is nothing to fall back to, and a default that silently resolves to the wrong database is worse than no default. Same shape as #16. The stale xenia-rs dis regeneration hint is now sylph-xexdb dis.
• SCHEMA.md is still headed "xenia-analysis schema reference" and cites xenia-rs dis --db — the retired repo and binary. Not fixed.
• Phase 3's "says SQLite, is DuckDB" correction was made where it counts (README and the binary's module doc both say DuckDB), but two inline comments at sylph-xexdb.rs:499 and :874 still say SQLite.

▶️ The Phase 5 and Phase 7 checks, re-run 2026-09-14 on the second machine

This page says "Do not trust this page's ledger when you get there." So the checks were re-run rather than read. No repository was archived and nothing was deleted — both are the human's, and both stay open. What follows is only what the checks now say.

Phase 5 · containment — and the check the page asks for is wrong for half its list

repo	disposition	tips	contained in Sylpheed
Syplheed-Reborn	absorbed	21	21 ✅
Sylpheed-Godot	absorbed	2	2 ✅

🔴 Phase 5 names four repos, but the containment check only means something for the two that were absorbed. xenia-rs and xex2tractor were never absorbed — their disposition is harvest, then archive. Run the check literally against xenia-rs and it reports 42 of 42 refs "NOT PRESENT in Sylpheed", every line true and every line unactionable. That is the failure this page already names once, about Phase 0's first gate: a check that goes red for something unactionable.

For the two harvest repos the invariant is different — the harvested artefacts are in Sylpheed, and the repository is archived, not deleted:

artefact	on main	on its PR branch
docs/reference/xex2-format.md	✗	✅ #30
docs/reference/xbox360-exports.json	✗	✅ #30
LICENSE	✗	✅ #30
crates/sylpheed-xexdb (32 files)	✗	✅ #32
tools/zq.py	✅	✅

⚠️ So the harvest is real but is not on main yet. Archiving is still safe — the branches are pushed, and archiving is reversible — but Phase 5 should follow the harvest PRs, not precede them, or main alone does not carry what was harvested.

Phase 5 · nothing lives outside git — re-verified, still true

	issues	PRs	releases	wiki
the four retiring repos	0	0	0	none
Sylpheed	36	19	0	none

📌 has_wiki: true is the repository feature flag, not content — /wiki/pages returns 404 on all six, i.e. no wiki was ever initialised. Read the flag as content and you would report six wikis that do not exist.

📌 All six repos are still archived: false, private: false. The retention question this page leaves open is therefore unchanged and now has a number beside it: docs/re/captures/ is ~87 MB of game screenshots in a public repository.

Phase 7 · most of the drop list is not on this machine

The list was measured on the agent box. Here:

item	page	this machine
agent-backups/	687 MB	absent
xenia-rs/audit-runs/	57 MB	724 KB
local pi/clippy, pi/clippy-clean, pi/reauth3	3 branches	0 — not on this box
texcompare/, ship_render/	14 MB + 64 KB	both absent
root canary_*.log / .stdout / .stderr	~15 MB	1 file
stock-oracle/	34 MB	34 MB ✅
Sylpheed/target/	32 GB	34 GB ✅

✅ Phase 7's one open question is answered. It asks to "confirm no docs/re/ page cites" texcompare/ or ship_render/ before dropping them. Four citations of ship_render exist — and all four name the example program, not the output directory: cargo run --example ship_render in xbg7-mesh.md, two in BACKLOG.md, and the tracked examples/ship_render.rs itself. texcompare is cited nowhere at all. The directories are droppable; the example is tracked code and stays. (Same directory-vs-file distinction that made Phase 4's "10 dangling citations" turn out to be one.)

⚠️ Disk is tighter than this page records: 90 % used, 96 GB free (the page says 83 %). Sylpheed/target/ is 34 GB of that and is the only large item here — but it is a deletion on the human's disk, and rebuilding it is hours, so it stays until asked for.

Phase 5 — Verify the invariant, then archive ⬜ NOT STARTED

For each of Syplheed-Reborn, Sylpheed-Godot, xenia-rs, xex2tractor: re-run the containment check at that moment — not from this page's ledger, which will be days old — then archive read-only on Gitea, and only then delete the local clone.

Archive before delete, and leave them archived. The check proves the commits are reachable. It does not prove nobody will want the repository boundary back.

⚠️ The fork: history rewrite, or not

Dropping files in Phase 4 reclaims no space. The 118 MB is in history — 512 distinct blobs, 120.5 MB — and git rm leaves every byte in .git (590 MB local, 598 MB on the server). Only a rewrite reclaims it, and it has a hard blocker:

# crates/sylpheed-export/Cargo.toml:84
sylpheed-formats = { git = "…/Sylpheed.git", tag = "formats-pin-2026-09-01" }

The build depends on this repo by tag, pinned in Cargo.lock to #1cd5b8b1. A rewrite invalidates all 8 formats-pin-* tags, every commit SHA, PR #23's merge, both agents' clones, and needs a force-push through branch protection.

• Don't rewrite (recommended): 118 MB on a host with 154 GB free, where Sylpheed/target alone is 32 GB. The cost is not worth the blast radius.
• Do rewrite: then it happens here, in Phase 5 — while tags are being re-cut and agents re-cloned anyway, so the disruption is paid once instead of twice. Re-point the formats-pin dependency first.

This is a decision, not a default. Whichever is chosen, record it here.

Phase 6 — Adopt the orphan assets ✅ DONE (PR #34)

Tracked by nobody today, and tools or references by the two-repo rule:

asset	→
ppc-manual/ — 397 files, including generator/generate_manual.py	Sylpheed
docs/ppc_instructions.{json,md}, docs/xbox360_exports.*, docs/xex2_format.md	Sylpheed docs/
docs/CROSS_BUILD_SETUP.md (58 KB)	Canary — it builds Canary
run-canary.sh	Sylpheed tools/
ai-agent-*.md ×4 (56 KB)	Sylpheed docs/, or discard — human's call
scratch/xbg7/extract.py — clean-room XBG7 scanner	superseded by sylpheed_formats::mesh; keep as history or drop
Sylpheed-try/…/examples/_voice_span.rs — 46-line untracked probe	keep or drop

Phase 7 — Intended drops ⬜ NOT STARTED

Stated so that nothing is dropped by omission.

drop	size	why it is safe
agent-backups/	687 MB	all 23 refs verified present in Sylpheed
stock-oracle/	34 MB	two Canary binaries, rebuildable from source
root canary_*.log / .stdout / .stderr	~15 MB	run output
vkd3d-proton.cache	15 KB	runtime cache
xenia-rs audit-runs/	57 MB, 41 files	already .gitignored; iterate-4A deleted 40 of 41
local pi/clippy, pi/clippy-clean, pi/reauth3	—	all 4 patches on main, by patch-id
Sylpheed/target/	32 GB	build output; the host is at 83 %
texcompare/ (14 MB), ship_render/ (64 KB)	—	tool output — confirm no docs/re/ page cites them first
the xenia-rs emulator	~54k LOC	the human's decision: superseded by Canary

---

▶️ STATUS 2026-09-13 — paused here, for the other machine to finish

Six of eight phases are done and pushed. Everything below is committed; nothing is left in a working tree. Five pull requests are open and none has been merged — merging is the human's, and #32 contains #30 and #31.

phase	state	where
0 secure single-copy work	✅ done	Canary 8e63a9542, sylpheed-re + audit-handle-lifecycle-probes pushed
1 make the archive honest	✅ done	xenia-rs iterate-4B/4C/4D/4E pushed (32/52/56/59 commits)
0b the dirt the gate could not see	✅ done 2026-09-14	xenia-rs harvest/import-thunk-naming b4f19f1, ported in feat/xexdb-import-naming
2 harvest xex2tractor	✅ done	PR #30 harvest/xex2tractor-assets
3 lift the DB tool	✅ done	PR #32 feat/xexdb-tool
4 captures + the missing check	✅ done	PR #33 fix/capture-citations
6 adopt the orphan assets	✅ done	PR #34 chore/adopt-orphan-assets, + Canary 99cc72662
— containers and agents	✅ done	PR #31 docs/containers-setup
5 archive the four repos	⬜ open — needs the human
7 intended drops	⬜ open — partly needs the human

🔴 Five claims on this page were wrong. Read these before trusting the rest.

1. "10 dangling citations" was ONE. Eight were directory references, which resolve and are simply absent from git ls-tree; one was a path at the end of a sentence with the full stop pulled into the match. The figure is corrected in Phase 4 below and in PR #33.
2. extract -r is not how the .pe was made, and nothing depends on it — Phase 2's gate, run.
3. The Phase 0 gate compared branches by SHA and called seven safe branches unpushed. It tests reachability now.
4. There is no DXC submodule drift. The m is nested submodule dirt.
5. Phase 0 did not secure all the single-copy work, and its gate could not have: it walks refs/heads, and 468 uncommitted lines in xenia-rs have no ref. Secured 2026-09-14 — see the section above Phase 5.

What the other machine should do next

Before anything: merge or decide on the open PRs

#30  harvest/xex2tractor-assets     reference files + LICENSE
#31  docs/containers-setup          container and agent setup
#32  feat/xexdb-tool                the DuckDB tool  ⚠️ CONTAINS #30 and #31
#33  fix/capture-citations          the docs/re check + prune
#34  chore/adopt-orphan-assets      PPC manual + canary launcher

⚠️ #32 is stacked. Merge #30 and #31 first, or merge #32 and close them — its build.rs needs #30's export table, so they cannot be independent.

⏳ One result outstanding: the workspace cargo test count for #32. check, clippy, fmt and wasm are all 0; the test leg was interrupted twice by branch switches and never reported cleanly. Run it and post the number:

docker/ci/run cargo test --workspace

Phase 5 — archive, and the history fork

Do not trust this page's ledger when you get there. Re-run the containment check at that moment — days will have passed and branches may have moved:

for repo in Syplheed-Reborn Sylpheed-Godot xenia-rs xex2tractor; do
  # every ref of $repo must be an ancestor of a Sylpheed ref
done

Then archive read-only on Gitea. 🔴 Do not delete any repository. Archiving is reversible; deletion is not, and it is the human's call, separately, later. Nothing outside git needs migrating — issues, PRs, releases and wikis are zero in all four (verified).

The history-rewrite fork is still open and still recommended AGAINST: the 118 MB is in history, and crates/sylpheed-export/Cargo.toml depends on this repo by tag (formats-pin-2026-09-01, pinned in Cargo.lock). A rewrite invalidates all 8 tags, every SHA, PR #23's merge and both agents' clones.

Phase 7 — the drops

All verified, none done — they are deletions on the human's disk:

agent-backups/          687 MB   23 refs verified present in Sylpheed
Sylpheed/target/         32 GB   build output; the host is at 83 %
stock-oracle/            34 MB   two Canary binaries, rebuildable
root canary_*.log        15 MB   run output
xenia-rs audit-runs/     57 MB   already .gitignore'd; iterate-4A deleted 40 of 41
pi/clippy, pi/clippy-clean, pi/reauth3   all 4 patches on main, by patch-id
texcompare/ (14 MB), ship_render/        confirm no docs/re/ page cites them first

Still the human's, untouched

ai-agent-*.md ×4 (56 KB), scratch/xbg7/extract.py, _voice_span.rs, and the retention question: all six repos are private=False, and docs/re/captures/ is ~87 MB of game screenshots in a public repository.

What each finished phase actually produced

• Phase 2 — docs/reference/{xex2-format.md, xbox360-exports.*, ppc-instructions.*} and an MIT LICENSE; this repo had none. The first two are byte-identical to xex2tractor's, verified with cmp.
• Phase 3 — sylpheed-xex + sylpheed-ppc + sylpheed-xexdb, ≈17,200 lines of ~71,000, coupled to the emulator by three symbols. The gate: a regenerated database whose schema is a superset, with instructions, pdata_entries, imports, sections and eh_* identical to the row. The two −1.8 M deltas are the max_indirect_candidates ceiling, recorded by the tool as 6,556 truncated sites standing for 1,801,075 candidates — checked, not assumed: the xref delta is entirely ind_call. build.rs no longer reads a sibling checkout, and a missing export table now fails instead of silently producing a database with no import names.
• Phase 4 — tools/re/check-capture-citations, 212 captures all cited, 46 orphans dropped. It caught its own damage twice: two "orphans" are opened by filename from screen_match.py, and the first prune emptied two directories that pages cite as directories.
• Phase 6 — tools/ppc-manual/ (393 files) and a rewritten tools/run-canary.sh; the old one pointed at a symlink Wine cannot resolve and hardcoded one machine's paths. The full CROSS_BUILD_SETUP.md went to Canary, where the tracked copy was a 6.8 KB stub of a 59 KB guide.

The method, if you only read one thing

Every claim on this page that was reasoned turned out wrong, and every one that was run held. The gates are the point: Phase 2's gate refuted its own premise, Phase 3's proved the two alarming row deltas were an improvement, and Phase 4's check caught a deletion I had just made. Run them again rather than trusting the results recorded here.

Still open, for the human

• Retention. The .iso (7.6 GB), .pe, .xex.json and sylph_extract/ (6.2 GB) stay on disk and out of git. But docs/re/captures/ is 118 MB of game screenshots inside a public repository, about to be joined by a 938 KB export database, and all six repos are private=False. Worth choosing rather than inheriting.
• The history fork in Phase 5.
• ai-agent-*.md, scratch/, _voice_span.rs — keep or drop.