Dogfood edit signal follow-up: verdict parking, outcome-gated arming, inline quit flush (codex review of #202)

[T0mSIlver]

What

Follow-up to #202 (merged before its codex/GPT-5.6 review landed); fixes all six review findings on main. The feature is dogfood-only instrumentation, but every one of these corrupted the measurement it exists to produce.

1. High — a superseded watch lost its verdict when its record write was still pending. Arming B replaced the single watch slot, so A's late attachRecord was rejected and A read as "never observed". Closed-but-unattached watches now park keyed by generation (cap 8, oldest evicted) until their attach arrives. The old test at the previous line 491 pinned this bug — it was rewritten to assert the survival, with a comment saying exactly that (this is correcting a bug-pinning expectation, not weakening).
2. High — the watcher armed on .failed and .copiedToClipboard commits, so Backspace in the target app recorded erasing an insertion that never happened, and phantom cleans inflated the denominator. The commit outcome is now passed through; the watch arms only on .succeeded (the record still writes, honestly behavior-free).
3. High — the quit flush ran inside a Task spawned at terminate — the exact non-guarantee its own comment warned about. flushForTermination() now runs inline in the main-queue notification callback (MainActor.assumeIsolated); only backend shutdown stays async. New seam (debugRegisterLifecycleObservers(on:)) lets the test post the real notification and assert the patch is on disk when the observer returns, with no await after post.
4. Medium — clipboard payloads were sized by the placeholder, giving a 100-word paste a 2s window instead of 15s. The substituted text now feeds the window ladder for measurement only; the persisted record keeps the placeholder (asserted — clipboard content must not leak into dogfood records through this fix).
5. Medium — the window timer promoted weak self to strong before its sleep, retaining the watcher for the whole window and defeating the isolated deinit cleanup. Promotion now happens after the sleep; a release test nils the last owner mid-window and asserts dealloc + monitor teardown.
6. Low — ⇧⌘A was classified as select-all. .shift excluded; the two-gestures test extended.

Proof

• Regression tests added, failing before / passing after (both runs via remote-build.sh dogfood, the instrumented lane — these suites are #if LOCALVOXTRAL_DOGFOOD and compile out of the plain test lane):

  before: Executed 59 tests, with 15 failures (0 unexpected)   # every finding reproduced
  after:  Executed 59 tests, with 0 failures (0 unexpected)
  

  Representative red lines per finding are in the commit message; e.g. finding 3: the patch must be on disk when the observer returns, finding 5: the window timer must not retain the watcher across its sleep.
• Full non-instrumented unit suite green on the pre-rebase base: Executed 2015 tests, with 2 tests skipped and 0 failures; instrumented suite re-run green after rebasing onto current main (post-Remote listener: say WHY a connection was rejected, in the log and in Settings #200): Executed 59 tests, with 0 failures.
• LLM lanes: DictationViewModel+Session.swift is on the filter path list, so the lane will run; the change passes a new parameter on the capture call and does not alter what reaches the model.
• Not machine-verifiable: OS-level quit ordering on a real GUI session — the notification-path test pins the wiring; a hand-quit during a dogfood window on the next dogfood build is the remaining eyeball.