Queen supervisor: two bees at once, and the two defects that only appear then

[gHashTag]

Two workers ran concurrently for the first time — browseros-ai#1093 from 10:21:16 to 10:23:18 and browseros-ai#1098 from 10:21:18 to 10:24:15, two minutes of genuine overlap. Each on its own branch, each with its own file boundary, each committing one file; both branches diffed against the base contain exactly their own file and nothing of the other's, and the working copy never left feat/queen-supervisor.

The green result is not the point. Running them in parallel exposed two defects that cannot occur any other way, which is why neither had ever been seen.

The ownership rule compared boundaries as strings

conflictingTasks used Set.isDisjoint, so docs and docs/live looked unrelated and both claims were admitted. A write, meanwhile, is judged by containment — so the bee owning docs was entitled to write docs/live/x.md, the file the other bee owned. Two writers, one file, and nothing complaining until the merge. The two rules had one notion of "path" each and disagreed.

They now share pathsOverlap, which compares by path component: nested boundaries clash, siblings do not, and a shared prefix (docs / docsite) is not containment. That last part matters — a rule that answered yes to everything would pass every clash test while quietly stopping all parallel work.

The Queen was losing bee reports

Six reports arriving at once persisted three. Not late — lost: the check now waits ten seconds and they never arrive. Every caller appends to the same main-actor buffer, so memory was always complete; then each passed its own whole-array snapshot to save, the saves raced, and the file kept whichever finished last rather than whichever held the most.

Writes to her chat are now queued, and each takes its snapshot when it runs rather than when it was queued.

The check that was hiding it

A fixed 300ms sleep. Enough on an idle machine, so the scenario passed every run I had ever done; it failed once in eight under a parallel build. The sleep is now a bounded wait, and the race has a deterministic check instead of a probabilistic one: the mock persister holds the first write to her chat open for two seconds while a second overtakes it. Verified from both sides — it fails with the fix removed and passes with it restored.

Also cleared: sixteen prototype exemptions naming files 939028c91 had already deleted. The gate had been failing on them and was right to.

Verification

• make check — dev app built, cassette suite passed (4 cassettes)
• SSE end-to-end suite: 455 checks, ratchet floor raised 437 → 455
• 20 consecutive runs under a parallel build, no failures

Closes gHashTag/trios#1099