feat(action): add configurable model fallback

[l30t]

Summary

Workflows that require their requested model can opt into model_fallback: fail. Today the Action may replace a blocked or invalid model with a concrete tier-alias stand-in or the organization's default model. Fail mode stops instead, while leaving model selection and the default fallback behavior unchanged.

• Expose the input on the main, review, security, and base actions, and forward it through preparation and execution of both review passes.
• Use one shared parser to validate the input before policy lookup or execution.
• Reject known policy denials during preparation and stop recognized model failures after one execution, using the existing retry predicate and error classifiers.
• Keep transient retries, usage-limit and turn-limit protections, and current model/reasoning defaults.

Scope and impact

This is independent Action hardening, not a prerequisite for CLI-2122 or Sonnet 4 retirement. The default remains organization-default; existing workflows do not opt into strict mode automatically.

No changes to security or validator model selection, reasoning settings, debug artifact names, CLI policy enforcement, or GitLab component inputs. In fail mode, an older CLI that rejects a tier alias fails instead of using the concrete-model compatibility fallback. Tier aliases still resolve in the CLI, and a failed policy lookup still defers enforcement to the CLI/server.

Testing

Validated on 33026ce:

• bun test: 754 passed, 0 failed.
• bun run typecheck: passed.
• bun run format:check: passed.
• actionlint .github/workflows/*.yml: passed.
• Red-first regression run against unmodified dev: 11 failures covering strict policy rejection, eager configuration validation, action input wiring, invalid aliases, and policy errors reported through stderr, result events, and agent-loop events. These pass after implementation; coverage also checks invalid concrete models and unchanged arguments on transient retries.
• Process-level tests use a fake CLI to exercise real subprocess and retry behavior. Live Factory CLI/GitHub Action end-to-end execution was not run.

[factory-droid]

Droid finished @l30t's task —— View job

---

Droid review summary

Solid wiring change overall, with tests covering the new input across action variants. The main gaps are around fail-closed ergonomics: fail mode still retries deterministic model failures, and invalid fallback configuration is only validated on the disallowed-model path.