v3.1.5 — codex-executor: surface response.failed error reason + retry once

Improvements

codex-executor: failed reviews now say why (#40)

When the mantle Responses API returns a response.failed terminal event, the executor used to discard response.error and post a generic "failed before producing output" — giving no signal whether it was the PR, our code, or AWS. (Surfaced live: gpt-5.5 on mantle was 500-ing on every request with server_error in both us-east-1 and us-east-2, while gpt-5.4 / gpt-oss-120b were healthy — an AWS-side model outage.)

• Surface response.error (code + message) — printed as a ::error job annotation and shown in the ❌ sticky, with a note that server_error is typically an AWS/mantle service-side issue, not the PR.
• Retry once on a transient response.failed (same params). Recovers one-off 5xx blips; won't rescue a full model outage.
• New failed outcome → ## ❌ Codex Review — model service error header + reason; the job fails (red ✗) for any non-(ok|truncated) outcome.

No consumer interface change. Consumers on @v3.1.4 should bump to @v3.1.5.

Follow-up (not in this release): auto-fallback to a secondary model when the primary is down.