fix: hold completion notifications until the parent run settles

[vincelwt]

Fixes #185.

The problem

emitIndividualNudge re-checks record.resultConsumed before sending, but that check runs when the message is handed to pi.sendMessage — 200ms after completion (NUDGE_HOLD_MS). Delivery happens much later: deliverAs: "followUp" is only delivered once the agent has no more tool calls, so a notification emitted mid-run sits in pi's follow-up queue until the run ends, where it can no longer be withdrawn. The orchestrator usually joins the agent with get_subagent_result somewhere in that gap, so the notification is already stale by the time it lands — after the final answer, costing a turn to answer "already incorporated". With pi's default followUpMode: "one-at-a-time", a batch drains one turn each.

The change

src/nudge-queue.ts (new, ~100 lines) takes over the timing that was inline in index.ts:

• Parent idle → deliver after the existing 200ms window. Unchanged behaviour, and the case where a notification is genuinely useful.
• Parent mid-run → park in-process instead of in pi's queue, and flush at agent_settled.

agent_settled rather than agent_end or turn_end: those fire with a retry, auto-compaction, or another tool-calling turn still ahead, so a notification emitted there is parked by pi exactly as before. agent_settled is the first point where pi won't continue on its own.

The suppression logic itself is untouched. Every send closure already re-checks resultConsumed (and the group callback re-filters unconsumed records), so deferring the call defers the check — an agent joined in the meantime simply never notifies. No new suppression path to keep in sync.

Busy state is read through ctx.isIdle() at delivery time rather than tracked as a local flag, so an unbalanced lifecycle event can't strand notifications; an absent ctx falls through to immediate delivery, i.e. today's behaviour.

dispose() drops everything undelivered, matching the existing session_shutdown semantics (results are undeliverable once the session is gone, which is why shutdown already calls abortAll()).

Notes

• scheduleNudge / cancelNudge keep their signatures and call sites; they're now thin wrappers, which keeps the diff in index.ts small.
• Group notifications go through the same queue under their group: key. Their callback already re-filters on resultConsumed and returns early when nothing is left unconsumed, so a fully-joined group now correctly produces no notification at all.
• Extracted rather than left inline so the timing is testable in isolation, matching how group-join.ts and status-note.ts are structured.

Tests

10 new tests in test/nudge-queue.test.ts (fake timers, in the style of group-join.test.ts): idle delivery, mid-run parking, flush, cancellation while parked (the actual regression), cancellation inside the hold window, busy state re-read at delivery rather than schedule time, same-key replacement, flush-once, one throwing send not blocking the rest, and dispose dropping both parked and in-window notifications.

npm run lint, npm run typecheck, npm test (765 passed, 5 skipped), npm run build all pass.

I left CHANGELOG.md alone per CONTRIBUTING. Nothing in the README describes the timing being corrected here, so I didn't touch it either — happy to add a line if you'd like one.