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 per-target executor-model overrides to the Harbor verifier's finalize phase, enabling "transfer probes" — scoring the selected champion under a model it was not optimized on, so model-specific couplings (e.g. hardcoded temperature=0 crashing on an executor that rejects it) surface at finalize rather than weeks later in a manual probe.

• TargetSpec.model / _TargetCfg.model / VerificationTarget.model thread the override from build.yaml through the compiler into serve.json and on to the verifier.
• engine.evaluate_admin(model=...) creates a shallow model_copy of run_constraints with task_params[\"harbor_model_override\"] set without mutating the shared object; HarborRunner._build_command reads that key and passes it to harbor run -m, taking precedence over the task's configured model.
• The baseline is scored under the same override so the champion-vs-baseline comparison stays like-for-like; selection is untouched (override applies only at finalize).

Confidence Score: 5/5

Safe to merge — the override only activates at finalize, selection incentives are unchanged, shared run_constraints are never mutated, and the baseline receives the same override keeping comparisons like-for-like.

The core data flow is simple and well-tested: four targeted tests cover the engine immutability guarantee, the runner precedence logic, the verifier both-sides override, and the compile-to-serve.json round-trip. No existing eval paths are modified; the change is purely additive and opt-in via a nullable field.

vero/src/vero/harbor/build/config.py — TargetSpec.model is accepted by Mode A without a load-time guard, which could silently produce a do-nothing transfer-probe column for Mode A tasks.

Important Files Changed

Filename	Overview
vero/src/vero/evaluation/engine.py	Adds model param to evaluate_admin; uses Pydantic model_copy to build a per-call task_params dict without mutating the shared run_constraints. Immutability verified by tests.
vero/src/vero/harbor/runner.py	Reads harbor_model_override from task_params with fallback to configured model; injects -m flag correctly. Magic string duplicated with engine.py (P2 noted).
vero/src/vero/harbor/verifier.py	Passes target.model to both the champion eval and the baseline eval, keeping the comparison like-for-like. Logic is correct and tested.
vero/src/vero/harbor/build/config.py	Adds TargetSpec.model; shared by Mode A and Mode B, but Mode A silently ignores it. No load-time guard to catch Mode A misconfiguration (P2 noted).
vero/src/vero/harbor/build/compiler.py	One-line addition passes t.model through into the compiled serve.json targets dict. Correct.
vero/src/vero/harbor/serve.py	Adds model field to _TargetCfg; model_dump() round-trip into VerificationTarget(**...) correctly propagates the field to the verifier.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant BC as build.yaml (TargetSpec.model)
    participant Compiler
    participant SJ as serve.json (_TargetCfg.model)
    participant V as Verifier
    participant EE as EvaluationEngine
    participant HR as HarborRunner

    BC->>Compiler: compile_task()
    Compiler->>SJ: writes model field per target
    SJ->>V: "load_serve_config to VerificationTarget(model=...)"
    V->>EE: "evaluate_admin(model=...)"
    Note over EE: model_copy(run_constraints)<br/>+ task_params[harbor_model_override]
    EE->>HR: evaluate(params with harbor_model_override)
    HR->>HR: task_params[harbor_model_override] or c.model
    HR-->>HR: harbor run -m override-model
    Note over V: baseline scored under same override

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant BC as build.yaml (TargetSpec.model)
    participant Compiler
    participant SJ as serve.json (_TargetCfg.model)
    participant V as Verifier
    participant EE as EvaluationEngine
    participant HR as HarborRunner

    BC->>Compiler: compile_task()
    Compiler->>SJ: writes model field per target
    SJ->>V: "load_serve_config to VerificationTarget(model=...)"
    V->>EE: "evaluate_admin(model=...)"
    Note over EE: model_copy(run_constraints)<br/>+ task_params[harbor_model_override]
    EE->>HR: evaluate(params with harbor_model_override)
    HR->>HR: task_params[harbor_model_override] or c.model
    HR-->>HR: harbor run -m override-model
    Note over V: baseline scored under same override

Loading

_{Reviews (2): Last reviewed commit: "feat(harbor): transfer targets: per-targ..." | Re-trigger Greptile}