Fleet claims: parked/cancelled agents hold write claims indefinitely — no release path blocks sibling work

[Hmbown]

What happens

A child agent parked mid-turn (parent turn ended before it settled) keeps holding its write claim after the parent cancels it. Every subsequent write-capable spawn with an overlapping scope — including the SAME worktree path the claim names — is refused with write-scope contention with agent_X, even when the claimed path no longer exists on disk. There is no release path a sibling or parent can invoke: children may only cancel their own descendants (agent cancel → refused as non-descendant), and the coordinate/release surface named in the error does not exist as a tool.

Evidence (2026-09-05, three incidents)

1. Parked child agent_4361a6fb (session fix-5896-win-cu) claimed worktrees/cw-5896-win-cu-20260905. That worktree was later deleted from git entirely (git worktree remove), yet the claim still refused: (a) a resume spawn under a new name; (b) a fresh implement agent on a different worktree; (c) per that agent's report, EVERY execution surface it had (bash, Run, MCP shell) for the entire session — 'blocking peers: agent_4361a6fb'.
2. Its replacement sibling (fix-5896-v2) reported the same blanket bash blockage and had to route all shell work through its own turn-owned children to finish (fix(computer-use): win32 backend reports PowerShell failures truthfully #5903's branch).
3. The fleet-menu implement agent (feat(tui): prioritize the Fleet menu surface (#5888) #5905) lost its entire execution window to the same claim: could not create its worktree, run cargo, or commit; delivered a prepared patch + apply script instead, which the operator then executed manually.

Expected

• Cancelling a parked/cancelled agent releases its write claims (or reaps the record promptly).
• A claim on a path that no longer exists cannot block anything.
• The parent (spawner) has a working release path; the error message should name one that exists.

Relation to #5575

Same family as the posture single-source-of-truth issue: authority state (claims) has no single owner and no cleanup contract. This instance is concrete, reproducible from today's receipts, and blocks parallel agent work at the workspace level.

[Hmbown]

Founder-visible symptom (2026-09-05/06, live TUI): parked/interrupted child records render in the fleet view as agents 'waiting for input' indefinitely — there is no visible lifecycle state for parked-vs-really-waiting, no cleanup path shown, and the parent's cancel leaves the husk record in place. Three husks accumulated in one operate session (two from the park/resume cycle in this session's receipts, one stale claim holder that blocked spawns while showing as waiting). The fleet surface should distinguish parked/interrupted from actively-waiting, and settled husks should age out or be dismissible.