fix: give reconnect-in-progress submissions a bounded retry window

[YishanCoding]

Fixes #224

Problem

RunExecutor.submit() rejects immediately with reconnect-in-progress any time activeRuns.newRunsPaused() is true. bot/channel.ts's IM flush handler treats that rejection as terminal — it sends the "当前 bot 正在重连，稍后会继续处理新消息。" notice and discards the message, with no retry. Since reconnect windows are frequently just a few seconds (a WS ping timeout that self-heals — see the keepalive's 15s ping-timeout logic), any message unlucky enough to land in that window is silently lost rather than actually being handled "稍后" as the notice claims.

Fix

• ActiveRuns gains waitForResume(timeoutMs): Promise<boolean>, resolving true as soon as the current pause is released (all nested pauseNewRuns calls have unwound) or false once timeoutMs elapses first. Implemented with a small waiter list notified from the existing pauseNewRuns release callback — no polling.
• RunExecutor takes a new optional reconnectWaitMs (default 0). When > 0 and the submission isn't nowait, a paused submit() call waits up to that long for the pause to clear before falling back to the original reconnect-in-progress rejection. nowait submissions are unaffected — they already opt out of waiting for anything (pool included).
• bot/channel.ts wires reconnectWaitMs: 8000 for the production RunExecutor, matching the keepalive's ping-timeout window.

Compatibility

reconnectWaitMs defaults to 0, which reproduces the exact prior behavior (synchronous, immediate rejection) — every existing test in tests/integration/executor/run-executor.test.ts and tests/unit/runtime/run-executor.test.ts passes unmodified. This is purely additive/opt-in.

Testing

• New unit tests: tests/unit/bot/active-runs.test.ts (6 cases — resolves immediately when unpaused, resolves on resume before timeout, resolves false on timeout while still paused, non-positive timeout, nested pause/resume only fires once fully unpaused, no waiter leak across cycles).
• New integration tests in tests/integration/executor/run-executor.test.ts (4 cases — holds and then runs a submission once a short pause clears, still rejects once the wait window elapses, nowait fails fast even with reconnectWaitMs configured, and the default/unset case keeps the immediate-reject behavior).
• pnpm test:unit, pnpm test:integration, pnpm typecheck, pnpm build all pass. (Pre-existing failures in tests/unit/observability/logger.test.ts and three tests/integration/** files are unrelated — they hardcode the log filename bridge-20260525.jsonl and fail on any other date; confirmed unchanged before/after this change.)