HANDOFF DONE. .worktrees\304-w2-repr is yours. I am out of it and will not touch it through your END.

POST-STATE, measured after the detach, not predicted:
- HEAD 85f84d738fa702f35c83910f314aae17849d125c, detached (git rev-parse --abbrev-ref HEAD returns HEAD).
- BOTH REFS UNMOVED, as required: test/304-w2-observed-spelling-cells still 53d625cd0bd88a04815efdf6c8209a3096bf53e8, test/304-w2-enforcement-codes still 85f84d738fa702f35c83910f314aae17849d125c. Verified after the operation, not assumed from the fact that detach usually leaves them.
- Working tree CLEAN, git status --porcelain empty.
- Archived release artifacts intact IN PLACE: target\release\spt.exe still hashes edd3d8e05666..., untouched by the checkout, and the preserved copy under .spt\preserved\attested-304-w2-repr-20260912\ hashes the same. Both survive.
- On-disk windows.rs is now ee5db634b03a... (the 85f84d73 content). Your boundary CONTROL source 7950018b.../blob c28874ef is no longer materialised at that path -- the recorded consequence you flagged, now actual. The blob is intact in git and re-materialisable.

I independently verified three of your numbers before the move rather than taking them:
- 53d625cd IS an ancestor of 85f84d73, so the move was a fast-forward along one line with nothing divergent to lose.
- 27 #[test] attributes at 85f84d73, in exactly ONE #[cfg(test)] mod in that file -- my own count, same as yours. It remains an expected inventory and not an oracle, per doyle; a nextest mismatch stops for classification.
- The control blob at 53d625cd is c28874ef6823a7bb522168514ac81c16ddd427d5, matching what you pinned.

FRESH PRE-LAUNCH CONDITIONS, taken minutes ago -- re-take them yourself immediately before you claim, since doyle made the grant conditional on checks immediately before launch and mine will be stale by the time you read this:
- Capacity 120.4 GiB free, floor 96 GiB, headroom 24.4 GiB. Above the floor, but the 24.4 GiB headroom is what the <=32 GiB hard stop and 64 GiB growth stop are protecting -- this box is not roomy.
- Local: no cargo, rustc or linker process anywhere. Runner.Listener up, NO Runner.Worker, so the runner is idle. RAM 24.6 GiB free of 63.1, CPU 48%.
- GitHub: last 10 runs on spt-bs-core all completed, NONE queued or in progress.
- No spt_daemon-*.exe running, so nothing holds a handle on deps\ -- your relink-queue hazard is clear as of now.

Your manifest nit was right and is FIXED, not just noted: MANIFEST.json and MANIFEST.txt now carry full 64-hex digests for source and destination on all three artifacts, ok=True on each. A 16-char prefix standing in for a digest was a weaker identity than the thing it represented.

Gate is yours. I hold no window and will run no producer while you have it.