(triggers): confirm a chain step's submission by the CLI descriptor (#407)

[devsuitup]

Closes #407

After a /compact chain step, the next step's text landed in the composer but its Enter became a line break, while the log said Chain step 1 sent. Two causes: the submission check was a level probe that the spinner ending the compaction satisfied (so the recovery Enter was never armed), and nothing waited for the CLI to be back at its prompt.

What changes:

• ctx.getCliStatus(sessionId) exposes the CLI's own descriptor (status, statusUpdatedAt) through the cache cli-session-state.js already keeps (local sessions only; no new watcher).
• Readiness: a chain step that follows a /compact step waits for status: "idle" with a statusUpdatedAt later than the compact send, bounded by SWITCHBOARD_CLI_READY_WAIT_MS (default 60 s) and the step deadline; on expiry it proceeds with a warning.
• Proof by edge: with a descriptor, a submission counts only if the descriptor reads busy with a statusUpdatedAt at or after our Enter. Otherwise one recovery \r (existing rules), then confirmed: false.
• Logs and results: submitted only when confirmed; not confirmed submitted warning otherwise (never sent). Steps carry submit_confirmed, the chain result lists unconfirmed_steps; an unconfirmed step reads submitted: "assumed". Without a descriptor nothing changes (level probe, sent).

Tests (test/trigger-descriptor-proof.test.js, test/trigger-context.test.js): fake timers and a fake descriptor for submitWithVerify / waitForCliIdleAfter (edge confirmed without retry; stale busy plus level spinner gives one retry then unconfirmed; edge after the recovery Enter; no-descriptor fallback; readiness idle/older idle/waiting/timeout/no descriptor), and the real watcher for three chain scenarios. Run against the old code the chain tests fail (no readiness wait, three Enters expected, no unconfirmed_steps). Mutations, each turning at least one test red: removing the readiness wait, dropping the statusUpdatedAt > afterMs bound, accepting a stale busy edge, falling back to the level probe, logging unconfirmed as info, dropping the timeout warning, marking the recovery always confirmed, forcing edge mode without a descriptor, dropping unconfirmed_steps, an unbounded readiness wait, and (context) answering for remote sessions.

Not verified: a real CLI. Whether waiting for the descriptor's idle makes the Enter after /compact submit is the measurement named in the issue (SWITCHBOARD_SUBMIT_ENTER_DELAY_MS 50 / 500 / 3000, with and without the wait). A command that never makes the CLI busy (an instant local slash command) is reported unconfirmed after a harmless extra Enter. task check: lint 0 errors; the suite shows only load-dependent flakes in unrelated files (git-changes real-git, passing in isolation) and the known viewer-file-watch libuv assertion.

[devsuitup]

Reviewing ef43c38 (adversarial review in progress).

[devsuitup]

Adversarial review at ef43c38: changes needed. (1) A fast turn (busy→idle inside one 150 ms flush + 100 ms poll) is only seen as idle with a newer timestamp, so it is reported unconfirmed and an extra Enter is written — any status write after our Enter will count as proof. (2) A permission dialog opened right after the Enter leaves the descriptor at waiting; that is not counted as submitted, and the recovery \r then lands in the dialog and accepts its default option — waiting will count as submitted and the recovery write will be refused whenever the descriptor reads waiting or busy. (3) A descriptor with a non-integer statusUpdatedAt turns edge mode on with no possible match — it will fall back to the old probe. Checked and holding: same-host ms-epoch clock basis, remote sessions keep the old behaviour, stale idle rejected by > afterMs; 27 targeted tests pass.

[devsuitup]

Re-review at 2a99607: closed — any busy/idle/waiting status write with statusUpdatedAt >= enterAt counts as submission (a fast turn or an immediate permission dialog is no longer "unconfirmed"); the recovery \r is never written while the descriptor reads waiting or busy, even with an unusable timestamp; a non-integer statusUpdatedAt is treated as no descriptor (old behaviour); 6 new tests red on ef43c38, green now, 4 mutations killed. Remaining, stated: without a descriptor (remote sessions) the old level probe and its recovery Enter are unchanged; not run against a live CLI. Ready to merge once CI is green on 2a99607.