fix(mobile): return finished threads from the Working shelf#25398

Problem

A completed thread can remain in the mobile Working shelf after its two-second grace period. Reproduce by starting a run, keeping Home visible while it finishes, and leaving the list untouched: the row reads Done but stays under Working until a later list rebuild.

Change

Pass the inbox tracker version into the memoized observation so React Compiler includes it in the cache key. The timer notification now refreshes shelf placement. The tracker accepts an optional version argument; existing callers and grace/attention rules retain their behavior.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/24901

Triaged bug via the automated issue flow. GitCafe: #24901. This restores the existing completion behavior without adding settings, persisted state, or protocol changes.

Verification

• Red first: the new Metro compiler regression test failed on main because the observation cache guard omitted inboxVersion; it passes with the fix.
• vp run check:changed: passed formatting, lint, both knip checks, all seven affected-package typechecks, and 980 tests across compiler coverage and inbox tracking. The final rerun confirmed the cached passing result with exit 0.
• Independent GPT-6.1-Sol and Claude Opus 5.5 code reviews returned MERGE with no findings.
• Real Android development clients, React Compiler enabled, remote connection to an isolated server using explicit --home-dir. Native fingerprint 0328f4ec326e7e4bf1bf87625118f3ff2f7a18d7; no release-build claim.

Evidence

Proof audit and independent rebuild addendum found the Android proof sufficient. The independent head phone run repeated the automatic return; its settled frame differs by 2,746 pixels (0.105%) in clock, cellular indicator, and the older control's later status label. Independent recording, rebuild script, and transcript are public. Cold-start blank UI and local persistence warnings were observed during setup; the verified transition used a loaded, live connection.

Fixture matrix, full evidence index and capture recipe, server completion transcript, diff generator, and compiler reproduction helper.

For each recording, send Run sleep 30 in shell then reply Done. Do not edit files., observe the thread working, then leave the list visible across completion and grace without tapping or navigating. Exact UTC milestones, raw recordings, narrated derivatives, scripts, and device details are in the index.

Before: Done target stuck in Working. After: target returned above control. Right: absolute difference including clock.

Android tablet sidebar, Pixel Tablet landscape, light, 2560×1600 captured pixels: head recording. The collapsed shelf retains the selected target during work, then returns it automatically and removes the zero-count header. The final reselection recording ends with the conversation still loading; loaded content and identity after that action are supported by the linked accessibility transcript, not by a settled final video frame.

What else changed in the captures: wall clock/status-bar indicators differ, the target has one additional real turn at head, and the client/Metro restarted between revisions. Intended visible differences are shelf removal and fresh return ordering. No style change or pixel-identical claim is made. Android stops encoding a static tail: the base recording covers the running-to-Done transition, while the linked later unmodified still and persisted completion show it remained stuck 31 seconds after completion.

Unverified: native iOS tablet sidebar (A3). Its cached client installed, but pool orientation/launch/snapshot operations repeatedly timed out. No iOS runtime pass is claimed. Unchanged attention, placement, and continuation rules remain covered by existing tracker tests; no broader provider or transport sweep was performed.

Models and harnesses: Claude Opus 5.5 through Claude Code; GPT-6.1-Sol through Codex in T3 Code.