Work tail findings from #3047: CPU wedge, parked cancel, markdown inspect output

[ScriptedAlchemy]

The benchmark restructuring in #3047 surfaced three product defects. None was fixed there.

1. The work tail wedges a CPU after work_adjudicate_leak runs.
2. A cancel that lands before mark_running leaves the attempt parked instead of cancelled.
3. worktree_cleanup_inspect returns markdown where callers expect JSON.

Details are in the #3047 description and benchmark notes. Each needs a reproduction test before the fix.

[devin-ai-integration]

Devin is working on this (all three findings: work-tail CPU wedge, pre-mark_running parked cancel, worktree_cleanup_inspect markdown output). Session: https://app.devin.ai/sessions/8c3b6020316140699db560903cebf1aa

[devin-ai-integration]

Findings 2 (parked cancel before mark_running) and 3 (worktree_cleanup_inspect markdown) are fixed in #3068, which refs this issue rather than closing it. Finding 1, the CPU wedge after work_adjudicate_leak, is still open: I could not reproduce it, and the post-adjudication producer/rollup/activity paths I traced have no unbounded loop. Releasing my claim on finding 1.

Written by Devin

[devin-ai-integration]

Claiming finding 1 (work-tail CPU wedge after work_adjudicate_leak); findings 2–3 are covered by #3068. Plan: find the spinning/blocking loop, measure before/after, fix at the root.

Written by Devin