poll: return closed when the review ended before the poll started

[BioInfo]

The bug

endSession hands {"status":"closed"} only to pollers already registered on the entry key. But the loop SKILL.md prescribes is apply-then-poll, so the poll that follows a delivered batch normally starts after the human has clicked End review. That poll registers on a target nothing will ever close and blocks for its whole --timeout.

Hit it live today reviewing a Markdown draft: poll returned a batch, I applied it, the reviewer clicked End review, and the next poll --ack had nothing to end it.

Reproduction

open a session on a file
POST /api/session/<id>/end      # no poll in flight
human-review poll <file> --timeout 20

Before: exits after the full 20s with {"status":"timeout"}.
After: exits in under 100ms with {"status":"closed"}.

The new test in test/poll-handoff.test.js asserts exactly this. Run against main it fails at 20,125ms; with the change it passes in ~2.1s. Full suite 91/91.

The change

endSession leaves an end mark in the store when the last session for a target goes away, and /api/poll consumes it before registering a poller. Three ordering details:

• The mark is checked after the pending batch, so feedback the human already sent is never hidden by it.
• Opening the target again clears it, so a resumed review waits for real feedback rather than reporting closed twice. The test covers this.
• endSession consumes the mark for each poller it drains directly, so the two delivery paths cannot both fire.

It is persisted rather than held in memory because the server idles out after 45 minutes (IDLE_SHUTDOWN_MS), and a poll arriving after that would otherwise hang against a fresh process for the same reason. Verified by killing the server between the end and the poll: the fresh server answers closed in 0s.

No response vocabulary changed, so no SERVER_PROTOCOL bump and agents need no update. The closed payload moved to one SESSION_CLOSED constant since two paths now send it.

Notes

• No reformatting; the diff is +61/-12 in src/ plus the test.
• Written with Claude Opus 5. Tests were run locally on Node 26 (91 passing), and the before/after timings above are measured, not estimated. Happy to re-run or restructure anything.