fix: weight static + runtime in execution correctness scoring

[lezama]

Problem

Execution correctness was computed as the unweighted fraction of passed runtime assertions (_score_assertions):

passed = sum(1 for a in assertions if a.get("passed"))
return round(passed / len(assertions), 4)

This had three issues:

1. The static-analysis score was discarded entirely. The runtime computes a weighted static.score (required/forbidden patterns), but the harness never used it. Code matching zero required patterns could still score full correctness.
2. Per-pattern / per-assertion weights were ignored. The datasets define "weight" on every check; the runtime honors them, but the Python side recomputed an unweighted average.
3. Forbidden-pattern hard fails were ignored. A model tripping a forbidden_patterns rule (e.g. a direct SQL query) still got full correctness.

Fix

_score_correctness(raw, test) averages the static and runtime sub-scores — each already weighted by the WordPress runtime — and counts a dimension only when the test actually defines checks for it (so a test with only runtime assertions is scored purely on runtime, and vice versa). Applicability is read from the test definition, not the runtime output.

Hard crashes (fatal/execution error before any assertion runs) force correctness to 0.0: the runtime is ground truth, so static pattern matches must not rescue code that doesn't run. Code that runs but fails some assertions still earns partial credit.

Before / after (real e-hooks-001: 4 weighted static patterns + 2 runtime assertions)

Case	before	after
Registers both hooks but wrong meta key, no get_post_meta (runtime 2/2, static 2.0/3.0)	1.0000	0.8334
Works but trips a forbidden pattern (e.g. direct SQL)	1.0000	0.5000
Static perfect but code crashes	0.0000	0.0000

Tests

New python/tests/test_scoring.py (10 cases): static+runtime averaging, single-dimension scoring, forbidden hard-fail, crash detection (both signals), and the control case (runs but half assertions fail → keeps partial credit). Full suite passes (17), ruff clean.

Not included

The quality / LLM-judge dimension (30% of overall) is still unimplemented in the runtime — tracked separately.

🤖 Generated with Claude Code

[github-actions]

The following accounts have interacted with this PR and/or linked issues. I will continue to update these lists as activity occurs. You can also manually ask me to refresh this list by adding the props-bot label.

If you're merging code through a pull request on GitHub, copy and paste the following into the bottom of the merge commit message.

Co-authored-by: lezama <migueluy@git.wordpress.org>
Co-authored-by: JasonTheAdams <jason_the_adams@git.wordpress.org>

To understand the WordPress project's expectations around crediting contributors, please review the Contributor Attribution page in the Core Handbook.

[JasonTheAdams]

The quality / LLM-judge dimension (30% of overall) is still unimplemented in the runtime — tracked separately.

Can you elaborate on this, @lezama? I thought the LLM-judge piece was removed in my previous, giant PR.