Codex: fail fast on pinned unavailable alpha model and surface offending model in summaries

[Copilot]

Scheduled Codex workflows were failing 100% red because sub-agent sampling still pinned gpt-5-codex-alpha-2025-11-07, which 404s at the API proxy. The harness already classified this as non-retryable, but only after wasted in-CLI retries and without clear model attribution in summary output.

• Harness preflight guard for known-bad model pins

  • Added a known-unavailable model blocklist in codex_harness.cjs.
  • Preflight scans resolved Codex args/prompt payload (including sub-agent model instructions) and exits before spawning Codex when the blocked snapshot is present.
  • Step summary now includes explicit remediation guidance (move to a GA Codex model such as gpt-5.4-mini).

• Clearer invalid-model attribution

  • Added model-name extraction from common invalid-model payloads (AIC Model not found ... and generic model ... not found/is not available forms).
  • When invalid-model is detected, the harness writes the offending model to GITHUB_STEP_SUMMARY and includes it in stderr diagnostics.

• Repo-wide regression protection

  • Added a workflow-source guard test to fail if any .github/workflows/*.md reintroduces gpt-5-codex-alpha-2025-11-07.
  • Added harness unit coverage for blocked-model detection, offending-model extraction, and preflight short-circuit behavior.

const knownUnavailableModel = detectKnownUnavailableModel(resolvedArgs);
if (knownUnavailableModel) {
  log(`fatal: known-unavailable model "${knownUnavailableModel}" detected ...`);
  appendStepSummary(`### ❌ Codex model preflight failed
- Offending model: \`${knownUnavailableModel}\`
- Remediation: use a supported GA Codex model (e.g. \`gpt-5.4-mini\`).`);
  process.exit(1);
}

[pelikhan]

@copilot fix the models .json by removing that model and ensure that this kind of model name pattern is refused