Agent runtime: process-ownership fix (#3308) covers Claude chat and headless only — Codex/Gemini keep a stale registry entry, and a failure before the reader starts leaks one #3408

Follow-up to #3308 (fixes #3307), found while validating it for the merge train. Neither blocks that PR.

1. The fix covers two of four runtimes

#3308 makes process_registry.unregister() refuse to drop an entry whose process is still alive, and adds worker-completion cleanup so the executor thread unregisters once the child exits. The registry change is global; the worker cleanup was added only to Claude chat and headless.

codex_runtime.py and gemini_runtime.py still rely on the waiter's finally. When the waiter is cancelled while the child is alive, that finally is now a no-op, and nothing unregisters after the child exits: the entry stays, and log subscribers never receive stream_end.

Found by reading the code; no test covers it.

2. A failure between register() and the reader start leaks the entry

Paths that call registry.unregister() after a failed stdin write (claude_code.py, gemini_runtime.py in two places, headless_executor.py) are now a no-op when the child is still alive, and nothing retries.

Reproduced with a probe (live child, stdin write forced to raise): on the #3308 head the child stays registered, listed as running and allowlisted from the orphan sweep, and the entry and its log buffer remain after the child dies. On dev before #3308 the entry is dropped.

Reachability is unproven: in the realistic broken-pipe case the child is already dead or dying.

Acceptance criteria

• Codex and Gemini runtimes unregister on worker completion after an abandoned waiter, with a test per runtime that executes the path (cancel the waiter with a live child, let the child exit, assert the entry is gone and stream_end is delivered).
• The failed-write paths terminate the process group before unregistering, with a test that drives a live child through a failing stdin write and asserts no entry and no live child remain.
• Decide and document whether an abandoned child should still be eligible for the periodic orphan sweep (after fix: retain subprocess ownership after waiter cancellation #3308 it runs to its own deadline with the result discarded).
• termination_unconfirmed is named in the error detail the agent router returns instead of a generic 500.