feat: [no-ticket] harbor - transfer targets (per-target executor-model override at finalize)

[shehabyasser-scale]

Stacked on #37. Together they are the response to the wave-1 transfer-matrix finding: three of five champions independently hardcoded temperature=0 (a variance-reduction trick on their home model) and scored 0/72 on claude-opus-4-8, which rejects that parameter, while looking healthy on every eval the optimization loop ever ran. #37 makes such a crash legible after the fact; this PR makes it measurable DURING every run.

What it does

VerificationTarget gains an optional model: a finalize target with an executor override scores the selected commit under a model it was not optimized on, alongside its home-model reward in the same reward.json:

targets:
  - split: test
    reward_key: reward            # home executor
  - split: test
    reward_key: reward_opus       # transfer probe
    model: anthropic/claude-opus-4-8

A champion with a baked-in model coupling then ships reward_opus: 0.0 with the crash cause in target_errors (#37), instead of the coupling surfacing one substrate away, weeks later, in a manual probe.

Design notes

• The baseline is scored under the same override, so score_baseline comparisons stay like-for-like per target.
• Plumbing rides task_params["harbor_model_override"] (an existing per-eval passthrough): TargetSpec (build.yaml) → compiler → _TargetCfg (serve.json) → VerificationTarget → engine.evaluate_admin(model=...) → HarborRunner -m. Mode A ignores the key; the shared run_constraints are copied, never mutated.
• Selection is untouched: the override applies to finalize targets only, so the optimizer's incentives during the run are unchanged; portability is measured, not optimized against (that would be its own experiment).

Tests

• Engine: override rides task_params without mutating shared constraints; no-override passes constraints through unchanged.
• Runner: -m honors the override over the configured model.
• Verifier: the override reaches BOTH the champion and baseline evals.
• Build: TargetSpec.model round-trips through the compiler into serve.json and validates.
• Full affected suites green (153 passed after widening 11 strict-signature test fakes of evaluate_admin).

🤖 Generated with Claude Code

Greptile Summary

This PR adds an optional model override to VerificationTarget (and the full TargetSpec → compiler → serve.json → verifier pipeline) so finalize can score a champion under a model it was not optimized on — a direct response to finding that three champions independently baked in temperature=0 and scored 0/72 on an executor that rejects that parameter. It also adds bounded within-eval infra retry (infra_retry_rounds, off by default), dead-attempt classification with infra labeling, a key-budget exhaustion alarm, and an n_dead_infra metric.

• Transfer probe plumbing: evaluate_admin(model=...) creates a shallow copy of run_constraints (never mutates shared state) and injects the override into task_params["harbor_model_override"]; the baseline is scored under the same override so comparisons remain like-for-like.
• Infra retry (opt-in): samples whose every attempt died of a transient infrastructure exception can be retried in a fresh sibling jobs dir with a linear backoff; the qualifying predicate is restricted to transient causes only, and each recovered sample receives an infra_retry audit marker listing the discarded rounds.
• Dead-attempt classification: a conservative allowlist of infra exception names labels deaths as [infra] or [infra:llm-key-budget]; bracket-injection in candidate-authored exception names is neutralized before classification; n_dead_infra surfaces the count without affecting the score.

Confidence Score: 5/5

Safe to merge. The transfer-probe plumbing is minimal and non-invasive; infra retry is opt-in and off by default; no shared state is mutated.

The model override creates a copy of run_constraints rather than mutating the shared object; the infra retry is gated behind infra_retry_rounds > 0 (default 0); the adversarial re-roll vector is correctly neutralized by requiring ALL attempts to carry a transient-infra label. All new paths are covered by tests. No logic errors or data-integrity issues were found.

No files require special attention. runner.py is the largest change but is thoroughly tested by test_harbor_runner.py.

Important Files Changed

Filename	Overview
vero/src/vero/evaluation/engine.py	Adds model parameter to evaluate_admin; correctly uses model_copy to create a shallow copy of run_constraints rather than mutating shared state.
vero/src/vero/harbor/build/compiler.py	One-line addition to propagate t.model into the serialized target dict for serve.json.
vero/src/vero/harbor/build/config.py	Adds `model: str
vero/src/vero/harbor/config.py	Adds infra_retry_rounds and infra_retry_delay_s fields to HarborConfig with validation: rejects negative rounds and rejects zero/negative delay when retries are enabled.
vero/src/vero/harbor/runner.py	Largest change: adds dead-attempt classification, key-budget alarm, bounded infra retry with sibling-dir isolation and audit stamping, model override in _build_command, and n_dead_infra metric. Well-documented and tested; no shared-state mutation; adversarial-candidate guard is correctly opt-in.
vero/src/vero/harbor/serve.py	Adds `model: str
vero/src/vero/harbor/verifier.py	Adds model to VerificationTarget, threads it through _admin_eval_score, and correctly applies the same override to both champion and baseline evals to keep comparisons like-for-like.
vero/tests/test_harbor_runner.py	Adds TestInfraResilience with 9 tests covering classification, key-budget alarm, retry eligibility, multi-round backoff, audit-marker content, and adversarial no-retry invariants.
vero/tests/test_harbor_verifier.py	Updates 11 _admin fakes to accept the new model keyword arg; adds test verifying the override reaches both champion and baseline calls.
vero/tests/test_engine.py	Adds two tests confirming the model override injects into task_params without mutating run_constraints, and that a no-override call passes constraints through as the same object.
vero/tests/test_harbor_build.py	Adds a round-trip test verifying TargetSpec.model survives compilation into serve.json and is correctly deserialized.

_{Reviews (3): Last reviewed commit: "Merge pull request #39 from scaleapi/har..." | Re-trigger Greptile}