Make turn recovery progress-aware

[morganmcg1]

Summary

• classify Senpai's one-hour OpenHands lease boundary as a resumable paused_timeout while preserving genuine provider/runtime failures
• measure terminal stalling from the last successful persisted tool observation and reset consecutive retry/recovery allowance only on genuine progress
• quarantine only the poisoned inbox turn, continue later student/GitHub events once on a fresh branch, and publish queued post-quarantine work as degraded-but-live health
• migrate existing inbox state with chain-aware recovery counts and normalize naïve persisted SDK timestamps as UTC
• make recovery selection, progress-budget evaluation, reset, and quarantine one atomic transition while preserving processed/quarantined terminal states and making stale duplicate requests idempotent
• document the durable recovery contract in the README and specification

Verification

• uv run pytest -q — 904 passed, 6 skipped
• focused recovery/controller/OpenHands/supervisor suite — 111 passed
• targeted red/green proofs for lease pauses, quarantine continuation, chain migration, timestamp normalization, concurrent progress, terminal states, and stale duplicate recovery
• uv run python -m compileall -q senpai_agent
• bash -n k8s/*.sh scripts/*.sh plugins/senpai/scripts/*.sh
• uv lock --check
• git diff --check

Two independent read-only reviews completed. Their findings around poisoned-context isolation, prequeued-event health, failed tool observations, timeout-test stability, naïve local timestamps, migration ancestry, recovery races, terminal-state monotonicity, and stale duplicate calls are covered by the final implementation and regression tests.