fix: adopt upstream kernel, idle-wait and Codex continuity reliability fixes #69

Implement the first reliability batch from the September 14 review under #1 and pylon-code/pylon#114:

• Prime upstream fix(coding-agent): absorb kernel pipe errors instead of crashing the worker PrimeIntellect-ai/prime-agent#2213 (90ca4a457e0fef0bd925215d4c973abd6264f067): absorb kernel pipe errors rather than crashing the session worker.
• Prime upstream fix(coding-agent): park waitForIdle while the session-input pump is blocked PrimeIntellect-ai/prime-agent#2151 (8eaa00c37385c6b25fed9d46584b50ffc2e87f9c): park waitForIdle while the input pump is blocked.
• Prime upstream [RSI] fix(ai): reset the codex continuation chain on previous_response_not_found PrimeIntellect-ai/prime-agent#2314 (7d1913969d33fa5a31366321e3f5db0070dcf2bf): retry stale Codex continuation IDs once with full context before emitting output.

Adopt these focused fixes on current pylon, retain fork cancellation, correlated lifecycle, ownership and recovery behavior, and record selective adoption without advancing the complete upstream integration checkpoint. Verify focused kernel/queue/Codex regressions, npm run check, and hosted checks. Merge is authorized by the maintainer in the active Pylon conversation after review and green checks. Managed publication/update is a separate follow-up; merging source does not update installed runtimes.