Unified stage follow-ups: phone agent list after restart, goal loops on pooled agents, dictation with an empty lane

[Juliusolsson05]

Follow-ups from the #1013 parity sweep (surfaces). These are not regressions against a working feature in the sense of the two MAJORs, which are fixed on the PR, but each is a visible change the unified stage makes more common.

1. After a restart, the phone lists far fewer agents. Boot now starts a process only for the focused lane's agent (sessionOwnership.ts:117-131); main started every tile. The phone lists only agents with a running process (SessionFeedSource.ts:92-133, RemoteServer.ts:938-960), and it has no way to wake one (RemoteServer.ts:771-786). A user with 9 grid panes restarts and sees 1 agent on the phone. Either the phone lists hibernated agents and can wake one, or the narrower boot is documented as a phone-visible change.
2. A goal loop running on a pooled agent is invisible. GoalLoopPane is mounted only by TileTree.tsx, and the index has no loop indicator. Main's classic Dispatch had the same gap, but new agents now land in the pool (orchestration children always do, plan §9.2). The index row should show that a loop is running, and allow stopping it.
3. Dictation with an empty focused lane goes to another lane's composer. No leaf mounts in an empty lane, so the hotkey falls back to the last focused composer (dictationHotkeyRegistry.ts:218-228). Clear Lane (⌥⌫) makes this easy to reach.
4. No phone test uses a v3 workspace file. src/main/remote/workspaceProjection.test.ts:53-64 builds only v2 fixtures.

Each fix needs a fail-first test through the real entry point. The sanitized live workspace in testing/fixtures/workspace-v2/ is the recorded input.

[Juliusolsson05]

Reopening: GitHub closed this off "Closes #1031 item 3" in #1064, but the issue is a list of four and two are still outstanding.

Item	State
1. After a restart the phone lists far fewer agents, and cannot wake one	OPEN — needs an owner decision, see below
2. A goal loop on a pooled agent is invisible	Done — #1063 (merged)
3. Dictation with an empty focused lane goes to another lane's composer	Done — #1064 (merged)
4. No phone test uses a v3 workspace file	In review — #1065

Item 1 needs a product decision before it can be implemented. The issue itself names the fork: either the phone lists hibernated agents and gains a way to wake one, or the narrower boot is documented as a phone-visible change. Those are very different features — the first means the phone can start a process on the desktop, which is a new capability and a new failure surface, not a bug fix. I am not picking that on my own.

Everything else here is either merged or in review.

[Juliusolsson05]

Reopening again — #1065 said "Closes #1031 item 4" and GitHub parsed the Closes #1031 prefix, exactly as #1064 did. My mistake twice; a partial-item PR should say Refs #1031 (item 4), and I will use that from here.

Item	State
1. After a restart the phone lists far fewer agents, and cannot wake one	OPEN — needs your decision
2. Goal loop on a pooled agent is invisible	Done — #1063
3. Dictation with an empty focused lane crosses lanes	Done — #1064
4. No phone test uses a v3 workspace file	Done — #1065

Item 1 is the only thing left, and it is a product call, not a bug fix. The issue names the fork itself: either the phone lists hibernated agents and gains a way to wake one, or the narrower boot is documented as a phone-visible change.

Those are not the same size. Waking an agent from the phone means the phone can start a process on the desktop — a new capability with a new failure surface (what happens when the wake fails, or the desktop is asleep, or two phones race it), not a follow-up to a layout change. Documenting the narrower boot is a paragraph.

I am not choosing that on your behalf. Tell me which and I will build it.

[Juliusolsson05]

Audit note (2026-09-20): 3 of the 4 items are done. Only item 1 remains, and it is blocked on a product decision, not on work.

• Item 1 — phone lists far fewer agents after a restart: NOT DONE. SessionFeedSource.ts:98-105 still seeds only from manager.list() (live processes), and there is no wake path in RemoteServer.ts. The root cause is now closed as a fix: perf(boot): bound and prioritise rehydrate spawns instead of launching every pane at once #774 deliberately reduced boot to spawning the focused lane's occupant and nothing else, so on the phone a 9-pane workspace becomes 1 agent. The choice stated in this issue stands — either the phone lists hibernated agents and can wake one, or the narrower boot gets documented as intended.
• Item 2 — goal loop invisible on a pooled agent: DONE. c84ea01e (PR feat(dispatch): show a running goal loop on the agent index #1063): a loop 8/25 chip on the agent index row, and selecting it makes that agent the command target.
• Item 3 — dictation with an empty focused lane: DONE. 2ab5bd35 (PR fix(dictation): keep a hold inside the focused lane #1064): dictationHotkeyRegistry treats the workspace-focused session as authoritative, with three states (unknown / focused-empty / target).
• Item 4 — no phone test on a v3 workspace: DONE. 0ee7be89 (PR test(remote): cover the phone projection on a REAL v3 workspace #1065): workspaceProjectionV3.test.ts against a sanitized real workspace.json, fail-first proven by deleting v3 parsing.

Housekeeping worth recording: GitHub auto-closed this issue twice from Closes #1031 prefixes in #1064 and #1065, and it was reopened both times. A multi-item issue should be referenced with Refs, not Closes, until its last item lands — otherwise the tracker says done while item 1 is untouched.