exec: v1.4 notify only where kept jobs survive idling; crash-duplicate fix split out

[jaredLunde]

Two changes to the v1.4 "Sandbox job notify" spec. As before, the spec, ARCHITECTURE.md, the fixtures and the coverage test change together. Nothing is implemented yet.

Notify only where kept jobs survive idling. Releasing a job removes the heartbeat that kept its sandbox busy, so a sandbox that stops or kills idle processes (a vendor's auto-stop) would lose the job because of v1.4.

• An endpoint MUST NOT advertise notify unless its deployment says a kept job survives its sandbox going idle. This is an endpoint-side setting, off unless the operator turns it on, and on for the platform's own native endpoint.
• Native endpoints advertise as before: {"jobs": "v1.3", "notify": "v1.4"}.
• An endpoint without native jobs advertises notify for the baseline by answering the job probe with {"notify": "v1.4"} and no jobs. A client that predates v1.4 reads that as no native jobs, as it does today.
• The baseline offers notify only then, and only if it also finds curl and the endpoint honours stdin_base64.
• Coverage: new probe vectors, including a baseline_notify expectation, and a conformance check, a kept job survives idling.

The job ledger leaves v1.4. v1.4's own guarantee is now that notify adds no window in which an end is reported twice or not at all, beyond those today's code has (N34).

• The crash window in which a sandbox job's end is reported twice exists today for held streams too: a replica that dies after the turn is persisted but before collect_garbage removes the record.
• ARCHITECTURE.md now describes that window as a known gap, together with its fix as a separate future change: ledger entries in the transcript's fenced segment batch.
• Requirements go from 50 to 47, and the section gets shorter.

exec_notify_vectors passes 10/10. The previous test fails on the new probe fixture (2 of 10).

🤖 Generated with Claude Code

https://claude.ai/code/session_01JimHGjsfk2Ktm5GxyZJKKk