fix: typing while a tool runs re-armed the stall watchdog and reset the turn

[saucam]

Diagnoses the reported symptom: send a message while Claude is working and the session goes quiet, then needs a "please continue" to come back.

Root cause

The mid-turn push path carved out waiting_approval when re-asserting status, but not tool_running:

if (this.#status !== "waiting_approval") this.#setStatus("thinking");

Both are real states the push does not end, and both are load-bearing twice over:

• Frontends key off them — the approval bar off waiting_approval, the running-tool card off tool_running. The flip made the session header say "thinking" while the tool card still said "executing", which is exactly what the reported transcript shows.
• #watchdogPaused() keys off them. A genuinely-executing tool is a legitimately silent run — that is precisely why the stall watchdog pauses there.

So typing during a long tool silently re-armed the watchdog against it. Once the stall window lapsed, the run was force-recovered:

⚠️ Turn timed out — no activity for 0s. The session was reset;
   send your message again to continue.

killing a healthy tool and the message the user had just queued into it. The longer the tool — a warehouse sweep, a big build — the more certain the kill. Typing while it ran is what shortened its leash.

Fix

Keep tool_running for the same reason waiting_approval was already kept. Anything else still becomes thinking — the push does start model work there. When the tool finishes, the existing tool-completion path moves it to thinking and the watchdog re-arms normally.

Tests

Two, both verified to fail without the fix (I reverted the one-line change and re-ran to confirm they aren't vacuous):

1. Status is preserved across a mid-turn push during a running tool.
2. The damage that actually mattered — sitting past the stall window with the tool still running yields no timeout message, no provider teardown, no second turn. Without the fix this one fails with the exact ⚠️ Turn timed out message above.

The existing waiting_approval mid-turn test and the "long tool on a backend without mid-turn injection" test both still pass, so this closes the gap between them rather than widening either.

Full suite: 2313 pass, 0 fail; lint + typecheck clean.

Scope note

This fixes one confirmed mechanism with this symptom. If quiet sessions persist after it lands, the next thing I'd look at is #pendingMidTurnCount: it is incremented per querying push and only decremented on a turn_done, so if a push ever fails to produce the extra turn boundary it expects, the turn's real terminal turn_done gets absorbed as an intermediate one and the consumer waits forever. That path is currently backstopped only by the stall watchdog — the same watchdog this PR stops mis-firing. Worth a look if the symptom recurs, but I did not want to speculatively change reconciliation logic in the same PR as a proven one-line state fix.