fix(cli): stop steering known slash commands into a running turn

[me2seeks]

Summary

Typing /recap or /resume in the TUI while a turn is running never ran the command: editor.onSubmit only special-cased exit forms, /swarm, /graph, and /goal, so every other input — including known slash commands — went to steerRunningTurn() and the model received the literal text (Steering: /resume) as a user message. Whether the command executed at all depended on whether the turn was still alive when the enqueue landed (the fallback path re-routes through submitPrompt).

Mid-turn submissions of known slash commands now route explicitly: /goal and /recap answer locally (both are independent of the running turn — recap uses the separate session.recap.generate call behind its own in-flight lock); every other known command is refused with a clear message ("Cannot run /X while a turn is running — interrupt it (Esc) or wait for it to finish.") instead of being steered or silently no-opping on the runControl busy gate; unknown slash-prefixed text (skill invocations like /skill:<name>, paths) still steers unchanged.

Fixes#3308

Verification

• npm --workspace maka-agent test: 320 pass / 0 fail, including 4 new tests under slash commands during a running turn (/recap answers locally, /resume and /model refuse with a message, unknown slash text still steers).
• The new tests fail without the fix: reverting pi-tui-runner.ts and rebuilding makes the /recap, /resume, and /model tests fail while the unknown-slash test still passes.
• npm run lint, npm run format:check, npm run typecheck: pass.
• Not run: desktop e2e (no desktop change).

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Maka (AI agent) located the routing defect, implemented the fix and tests, and ran the verification; the commit carries the Generated-by: Maka trailer. Human contributor of record reviewed the diagnosis and the routing design and decided to submit.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

APPROVE.

The routing is right and the tests pin it at the observable boundary: driver.steered stays empty for /recap, /resume and /model, and /skill:review still steers, so both directions are covered rather than just the fixed one.

I checked the three claims the new comment rests on, and all three hold at 5c4ee2284:

• runAgentTurn sets busy and turnRunning together (pi-tui-runner.ts:1128-1130), so runControl's if (busy) return at :633 really would swallow /help, /context, /new and the rest without a word. Refusing loudly is strictly better than that silent no-op, and the rationale in the comment is accurate.
• recap.run goes straight to runRecap('manual') (:2857-2861) and never enters runControl, so answering /recap mid-turn is genuinely safe rather than merely untested.
• /goal <action> already carries its own if (busy) notice (:2713-2722), so exempting /goal from the refusal does not reintroduce a silent no-op on the mutating path. That was my first suspicion and the code refutes it.

biome format is clean on both changed files at this head.

One non-blocking P3 inline. Separately, and not a review finding: statusCheckRollup is empty on this head — the test check has not run at all — so this approval covers the diff, not CI.

AI disclosure: this review was assisted by Claude (Opus) for code search and cross-checking. Everything above I re-derived myself against the source at 5c4ee2284.