Investigate unexplained cancellation before base refresh

[0xroylee]

The mixed-effort sandbox validation at e2a9de4 produced one unexplained cancellation before base refresh. The cause remains unconfirmed.

Evidence:

• Issue Routing proof: format elapsed time #71 completed through PR Routing proof: format elapsed time #74 on codex/role-proof-1788962745785.
• Issue Routing proof: summarize token usage #72 completed Scout high, Implement medium and independent Review high. PR Routing proof: summarize token usage #73 passed both lint and the Routing acceptance functional check.
• At 2026-09-09T14:11:03.163Z the coordinator saved a refresh intent. The local log immediately recorded TASK_EXECUTION_FAILED, phase awaiting_merge.
• At 14:11:13.255Z the checkpoint became needs_replan with Task cancelled; execution requires replan. No refresh result was saved. The task worktree remained clean at a7b07c9 and no refresh backup ref existed.
• The controller did not intentionally stop productive work; it stopped when the run needed attention. The runtime returned without a top-level error. All Pi clients closed successfully.
• A second pair, Routing proof: format elapsed time (traced repeat) #75 and Routing proof: summarize token usage (traced repeat) #76, ran on the same source/fixture seed with test-only exception tracing around refreshBase. PRs Routing proof: summarize token usage (traced repeat) #77 and Routing proof: format elapsed time (traced repeat) #78 merged, including a successful rebase and fresh Review. No refresh exception was captured in that repeat.

Investigate cancellation origin, authority snapshots and error propagation across GitHubTaskPool.awaitSelection, GitHubTaskRunner.coordinate, and refreshBase. Do not assume the cause is a Git conflict, model effort or GitHub index lag without a reproduction.

Acceptance: obtain a deterministic reproduction or a trace identifying the initiating condition; preserve a useful safe reason when cancellation follows another failure; retain the existing approval, refresh-intent, lease and ownership checks. Do not reset #72's checkpoint or merge #73 merely to make the validation look successful.

Local evidence: .scratch/role-routing-live/execution-first.json, state-first.json, first-routing-proof.json and the isolated checkout's .agile/runtime/agile.log.