Level 7 — Frontier-to-local handoff: export, train, route, and fallback

[jayscambler]

Objective

Stress test the full localization path on a realistic scenario: discover with a frontier/default provider, export data, train a tiny local artifact, route into the local provider, and verify fallback behavior when the local path is weak.

Scope

This issue is the first execution target in the real-world escalation ladder. It should prove the end-to-end frontier-to-local path under realistic persisted state, not just isolated unit tests.

What To Exercise

• Run a short but real discovery loop on a stable scenario such as grid_ctf
• Persist generation outputs, scores, and knowledge artifacts
• Export strategy-level training data from the real run database
• Invoke the real training loop or CLI to produce a loadable MLX bundle
• Route competitor execution into the local provider when harness coverage is strong
• Verify explicit fallback to the primary/frontier provider when the local bundle is absent, disabled, or below threshold
• Confirm persisted history, reports, and artifact metadata reflect the real execution path

Implementation Guidance

• Prefer one high-value end-to-end test plus a small number of targeted support tests
• Use the real export and training runner path where feasible; avoid mocking away the subprocess/training boundary unless the test would become flaky or too slow
• Validate the routed execution path through orchestrator or pipeline integration rather than instantiating the local provider directly
• Keep the scenario deterministic and the time budget short enough for CI
• Capture which provider actually executed the competitor step so the handoff is observable in assertions

Produced Artifacts

• A persisted run database with at least one completed discovery run
• Exported training data bundle
• Loadable MLX artifact bundle
• Test assertions or logs proving local routing occurred
• Test assertions or logs proving frontier fallback occurred when expected

Parallelism And Dependencies

• This should run first
• AC-217 can begin in parallel once the exact scenario fixture and runtime harness assertions are settled, but this issue should merge first because it establishes the real localization path
• AC-218 should wait until this issue lands, because package portability is more meaningful once the exported and trained artifacts are known-good

Review Focus

• Did the test hit the real export/train/routing path, or was a critical boundary mocked out?
• Does the routed execution actually choose local at least once?
• Is fallback explicit, safe, and observable?
• Do persisted artifacts and history match the executed path?

Success Criteria

• Export completes with usable data
• Training produces a loadable local bundle
• Routed execution actually selects the local provider for the target role
• Fallback to the primary provider is explicit and safe
• Match outcomes and persisted history remain coherent across the handoff

Acceptance

• Discovery -> export -> train -> local load path completes
• Routed run uses local provider at least once
• Fallback path works when local artifact is unavailable
• Persisted artifacts and reports reflect the real execution path

[linear]

AC-216