Astra: L18 joint round, round 3

Your round 2 (CODEX-REVIEW-L18-JOINT-ROUND-FINAL-2.md) left three items open. This round reviews the model (region-ha-two-box-build at 372f7a4b) against the build (region-ha-1b-next-build at 6a4eadde).

1. P2, cold first serving wake: model fixed (917771bb). The first serving wake is now cold. It takes a fresh generation and a new installation stamp, records writes_lost for the stopped probe session, and retires P afterwards. The N6 witness asserts each of these.

Fixing the model surfaced a built-code data-loss finding: the probe stop did not flush before stopping, so writes the probe had acknowledged but not yet shipped were declared lost. Ruling 87 fixes it (6a4eadde; model 372f7a4b, rule reloc-probe-stop-flush). EndProbe now runs in this order: 1. stop the writers; 2. flush the database and files, through the release's shared flushStopped; 3. record the positions; 4. install the activator fence and stop root.

The serving wake stays cold. Its writes_lost is bounded at the last shipped transaction, so it names nothing acknowledged.

Ruling 75 is amended: a foreign fence refuses the stop after the flush, without asking root.

1. P3, session close and admission fence: model fixed (917771bb). The model now keeps these apart. The session's launches close at the stop, and root's admission fence follows at Collect. A new rule, reloc-admits-probe-only, mirrors relocatingAdmission, and no wake admits a writer in the interval.
2. P2, finding 1 (moves' restoring loses source files protection). Still open, and asked of moves on #236.

Judge:

1. Are items 1 and 2 closed?
2. Is ruling 87 sound? Check: - the flush order; - the failed-flush refusal; - the deadline; - whether a restart before the record goes stuck rather than losing writes; - the amended ruling 75.
3. Anything new?
4. Verdict: is finding 1 the only item left before relocation may merge dormant?

Give findings P1/P2/P3 with file:line.