fix: reconcile stale session statuses on boot; name the tool blocking a send

[saucam]

Two fixes for the same confusing experience: returning to a session and being told it's busy or blocked, with nothing to act on.

1. Non-terminal statuses survived daemon restarts

thinking / tool_running / waiting_approval all describe work owned by a live process — the provider's turn loop, and for approvals the in-memory #pendingApprovals resolvers. None of it survives a restart, and a resumed Session starts idle (#status is initialised, never restored from the row).

The row was never corrected, so the session list advertised work that no longer existed. Observed on a live daemon that had restarted many times:

waiting_approval  mpc-policy-eval-paper   Aug 9   (16 days)
tool_running      studio                  Aug 6   (19 days)
thinking          shield_and_policy       Aug 14  (11 days)

Store.reconcileStaleSessionStatuses() clears them at boot, before resumeSessions, so rows and live sessions agree from the first broadcast.

error is deliberately preserved — a terminal state a human may still want to see, not an artifact of the process dying.

2. The mid-turn send rejection named nothing

Sending while an approval is pending correctly throws — falling through would close the live turn queue and auto-deny the approval. But the message only said:

A tool approval is pending — approve or deny it before sending

which tells you something is waiting, not what, and gives you nothing to act on. After stepping away from a session it reads as the model having simply gone silent. Now:

The tool Bash is waiting for approval — approve or deny it before sending (this backend can't queue mid-turn). Your message was saved; send again after deciding.

Names come from the same maps the approval bar already uses (#approvalIdToMessageId → #toolCallMessages), so there's no new bookkeeping to drift. It falls back to the original phrasing when no name resolves: the status is authoritative, the label is best-effort, and a missing name must never turn a clear refusal into a crash.

Scope note

This does not change when the guard fires — that behaviour is correct and unchanged. It only makes the refusal legible. Whether a pending approval should also be re-surfaced to the sending client (so it can be answered in place rather than requiring the user to find it) is a real follow-up, but it's a protocol/UI change rather than a message fix, so it's deliberately not bundled here.

Verification

lint + typecheck clean, 2382 pass / 0 fail. Four new reconcile tests: all three non-terminal statuses cleared, error preserved, idle rows untouched (reports zero), and idempotence across a second boot. The two existing tests asserting on the rejection text still pass — the fallback keeps them honest.

🤖 Generated with Claude Code