[minglong51]

Phase 0's Codex input/recovery proof depended on temporary scripts that disappeared with the private captures. Add scripts/codex_qualification.py to repeat the bounded scenario inside an isolated clilane task: verify host isolation, open one synthetic input request, expire and restart the observer while keeping the provider alive, require a new source read, and cancel the input.

The probe uses the existing pinned collector, compares its private journal against the provider capture, and removes its credential copy before child-process cleanup, including on TERM/HUP/INT. Recurring thread/read responses maintain freshness while the final turn completes. The analyzer validates the isolated skills/list exchange, and the transport accepts an explicit sanitized environment.

This remains a Phase 0 tool for absolute private sockets. Its report is explicitly unqualified; the 24-cell matrix stays unknown and the shipped runtime is unchanged.

Validation:

• All 122 existing tests and the evidence artifact check pass; repository tests and fixtures are unchanged.
• 53 disposable checks cover malformed isolation messages, cleanup failures, blocked pipes, signal escalation, delayed completion, and missing source responses.
• Fresh live Codex 0.155.1 capture passes all 10 recovery checks: 43 messages, 4 thread reads, exact journal/source agreement, and no retained credential copy.