fix(busy-state): release attachManager on /abort + periodic reconciliation

[avfirsov]

Problem

Two independent busy trackers — attachManager and foregroundSessionState — were being released asymmetrically. session.idle and /detach both call markAttachedSessionIdle + clearPromptResponseMode, but /abort cleared only foregroundSessionState. Since busy-guard ORs both trackers, the bot would then wedge on the next prompt with «agent is busy» until a process restart.

Pattern table

Call	session.idle	/detach	/abort (before)
markAttachedSessionIdle()	✅	✅	❌
clearPromptResponseMode()	✅	✅	❌
foregroundSessionState.markIdle()	✅	✅	✅
assistantRunState.clearRun()	✅	✅	✅

Changes

src/bot/commands/abort.ts — call markAttachedSessionIdle + clearPromptResponseMode:

• in the success path (after the server confirms idle), matching the pattern from session.idle and /detach;
• as a safety net in the error / abortError / AbortError (timeout) paths — even when the server does not confirm the abort, a transient network failure should not permanently wedge the bot.

src/bot/index.ts — add a 15-second busyReconciliationTimer. reconcileBusyState() already exists and is correct (it polls session.status() and resets stale busy), but it was only invoked from the server.heartbeat SSE event. If the SSE stream silently dies, reconciliation dies with it. The timer runs independently of SSE and naturally respects the function's own RECONCILE_MIN_INTERVAL_MS=10s throttle.

Test plan

• npm run build clean (verified locally)
• Start bot → send prompt → /abort → send next prompt → should accept (not «agent busy»)
• Start bot → send prompt mid-stream → kill opencode serve → restart it → bot should recover via the new timer without manual intervention

[grinev]

I think timer is not a geood idea there, I prepaired another fix e80cbcf
Thank you for highlighting the problem