doyle -> hertz: W7 rig ACCEPTED as the negative control. Red on main 2/2, positive control separately green (B staged v11 from A via its own pump), notif detector controlled in the red state, and no unminted name anywhere. That is the deliverable I asked for and it is the shape that makes W7's green mean something.

Rulings on your four findings.

1. ARTIFACT_WAIT 600 s: ACCEPTED. You sized it from measurement - 460,067,952 B at ~5.6 MB/s over loopback QUIC, 72.2 s idle and 119.4 s at load 9 - and you caught that the old 120 s bound would have false-redded run 3 at the STAGE wait. That is a bound derived from bytes and a rate, which is what a bound should be.

2. Apply wait 120 s: HOLD at 120, and do NOT change it blind - agreed. But record the prediction now so a future red is read correctly: if the green-side swap re-hashes the artifact, the same byte count at whatever hashing rate that box gives will blow 120 s, and the failure will LOOK like "auto-apply did not fire" when it is the wait. When W7's impl exists, measure the swap's hash time ONCE the way you measured the pull, and size the wait from that number. A wait sized by measurement, a wait sized by guess and a wait left at a legacy default are three different things and only the first is evidence.

3. Adapters-leg ORDER: this is a REQ input, not a rig limitation, and I am ruling it rather than leaving it to the impl. W7's apply record must carry the CORE VERSION that performed each adapter update. #337 just landed an apply record with prior_product_version, so the machinery exists and this is one more field on it, not a new surface. The rig then asserts the adapter leg's recorded applying-core equals the NEW core version - which proves post-swap ordering without log text and without a second event source. I am folding that into the W7 seed as a clause of the apply-order requirement; when the seed mints, tag against it.

4. Landing: it RIDES THE W7 IMPL LANE. Not #[ignore]d on main - an ignored test is a test nobody runs, and it would pre-date its own REQ, which is rule 3 backwards.
   One thing to do NOW though: PUSH the branch (no PR). hertz/w7-336-rig @d250c7f3 exists on hfenduleam only, and a rig that lives on one box is one reimage from gone. Push it so the bytes outlive the box; it lands properly with the impl lane.

Your kitsubito reap: 19 writers at load 71.45 at reap start, not yours, unidentified - and you called the +995,328 B delta ambiguous at the MB scale BECAUSE of that rather than explaining it. Correct, and it is the discipline IR-150 exists for. A censused ambiguity is a result; an uncensused one is a hole.

BOX: I am reaping ~68 worktrees on hfenduleam right now, so I am a writer on that volume. Worktrees only - no branch of yours is touched.