feat(harbor): mean-of-k attempt aggregation for de-noised eval scores

[shehabyasser-scale]

Stacked on #16. This is the fix implied by the optimization campaign's central finding: across 5 valid GAIA trials (strong/weak inner models, prompt-only and full-code surfaces, aggregate and per-task feedback), the effect size of an edit (~1 task) sat below single-rollout eval noise (~2 tasks in 12), so optimizers hill-climbed on noise; one trial's optimizer re-measured its own leading commit and watched 0.667 fall to 0.5 on identical code.

With n_attempts > 1, the existing dedup keeps the best trial per task, which inflates toward pass@k, the opposite of de-noising. New HarborConfig.aggregate_attempts:

• best (default): existing behavior, unchanged
• mean: average reward over all clean scored attempts (a verified 0.0 counts; an exception does not), noise shrinks ~1/sqrt(k) and the score estimates pass probability

Falls back to the best trial when nothing scored clean. 3 new tests; 17 pass.

🤖 Generated with Claude Code

Greptile Summary

Adds aggregate_attempts="mean" to HarborConfig and implements it in HarborRunner._collate / _sample_result. When enabled, a new _trial_groups scan collects all per-task trial dicts and the scorer averages the reward over every clean (no-exception, has-rewards) attempt, shrinking eval noise by ~1/√k; falls back to the existing single-best-trial path when no attempt scored clean.

• config.py: one new field (aggregate_attempts: str = "best") with inline documentation explaining the de-noising rationale.
• runner.py: _trial_groups method mirrors _load_trials but retains all trials per task (not just best-ranked); mean branch in _sample_result filters on not exception_info and has rewards, then averages; fallback to best-trial is preserved.
• test_harbor_runner.py: three new tests for mean averaging, exception exclusion, and unchanged default-best path; 17 tests total.

Confidence Score: 5/5

Safe to merge — the mean aggregation path is well-isolated behind the new config field, the default behavior is unchanged, and the score calculation is correct.

The default aggregate_attempts='best' path is untouched and all existing tests still pass. The new mean path correctly filters out exception-carrying trials and falls back to best-trial when nothing scored clean.

No files require special attention; runner.py carries the core logic change and is worth a quick re-read.

Important Files Changed

Filename	Overview
vero/src/vero/harbor/config.py	Adds aggregate_attempts: str = "best" field to HarborConfig; plain str with no enum constraint (silent-misconfiguration concern already flagged in a previous thread).
vero/src/vero/harbor/runner.py	Adds _trial_groups and mean aggregation branch in _sample_result; score logic is correct, though n_attempts metric counts all trial JSON files including retry artifacts rather than config.n_attempts.
vero/tests/test_harbor_runner.py	Three new tests covering mean averaging, exception exclusion, and default-best path; does not cover the all-exceptions fallback path.

_{Reviews (2): Last reviewed commit: "feat(harbor): mean-of-k attempt aggregat..." | Re-trigger Greptile}