[mstaats-nvidia]

Compute equally weighted mean signed baseline-to-mutant degradation for each fixed task subset. Negative effects remain valid and receive inspection diagnostics; missing required pairs produce unavailable scores with missing_score_reason rather than zero or a reduced task set.

compute_mutant_scores.py owns the CLI and calls the scoring library for one case. It takes --cases, --analysis, --case, --results-root, and --output, checks the saved case identity, and selects recorded attempts by execution number. With signed case inputs, scoring needs only metadata, subsets, and results indexes. Unsigned sources must remain available to establish content identity; sign before removing them. Native validation is a separate tool.

The command guide documents arguments, the library API, and a worked JSON-output example. Public scoring functions have short docstrings. The shared score schema uses missing_score_reason throughout code, examples, validation reports, and synthetic fixtures.

Validation at f7a71eb:

• Full local suite with the Docker smoke enabled: 3,413 passed, 35 skipped, 290 warnings. Skips cover unavailable optional runtimes, Node-dependent rendering checks, Windows-only behavior, and explicitly gated live-model execution.
• Nine real local Harbor trials passed, followed by completed-job reuse and relocation. Scoring the saved native results produced the expected three subset scores; CLI/library output was byte-identical and execution evidence stayed unchanged.
• Downstream benchmark suite at feat(metric-benchmark): orchestrate configured experiments through stage libraries #99 f81f82a: 180 passed, 1 skipped (the Docker opt-in test, exercised separately above). This covers coverage/correlation consumption, coordinator replay, and standalone CLI parity.
• Ruff, formatting, type checks for Python 3.12/3.13, copyright, license, DCO, and whitespace checks passed.
• Hosted CI passed on this exact head, including Python 3.12/3.13 and all three native Gym compatibility routes.

Both inline review comments are addressed and resolved. The new regression cases cover rejection of a different cases_sha256 and aggregation across three tasks and two mutants. No scoring algorithm changes were needed during final validation.

Stack: main → #95 mutant scores → #96 coverage → #97 correlation → #99 coordinator. Merge preparation is complete; this PR remains open and has not been merged.

[aleckhoury]

The scoring math checks out. Each subset's score is the equal-weight mean over mutants of the mean paired task degradation, sign * (baseline - mutant). Negative values are kept and flagged in diagnostics. A missing, incomplete or metric-absent pair makes that mutant effect and the subset score null with a reason; nothing turns into zero or is silently dropped. Attempts are selected by execution number, never by outcome.

Edge cases probed:

• With latest, a failed first attempt followed by a later success scores correctly.
• lower_is_better produces the right sign.
• A wrong metric name makes every subset unavailable.
• With signed inputs, scoring works after task dirs, bundles and native trial dirs are removed.

At this head: 139 passed, 1 skipped. Two small notes inline.