Normalize snippet quality scoring: pass_rate + stable-only gating + artifact upload

[Copilot]

The snippet quality workflow was gating PRs on an absolute passed - failed score, making results highly sensitive to skip volatility, version churn, and stale baselines. A 5% real quality drop could be masked or a mass-skip could look fine.

Changes

conftest.py — normalized score metrics

• Adds executed = passed + failed and pass_rate = passed / executed to snippet_score.json
• Writes the file whenever executed > 0 (was: only when snippet_score > 0)
• Keeps snippet_score for backward compatibility
• Adds --stable-only pytest CLI flag to filter preview versions from test parameterization

test_snippets.py — stable-only version filtering

• Renames VERSIONS → ALL_VERSIONS
• In pytest_generate_tests, respects --stable-only by excluding versions ending with -preview

compare_score.py — tolerant, normalized comparison

PASS_RATE_TOLERANCE = 0.05   # allow up to 5% drop before failing
MIN_EXECUTED_FRACTION = 0.50  # require ≥50% of baseline executed count

• Checks minimum executed count first (catches mass-skip / environment failure)
• Compares pass_rate within tolerance instead of raw score delta
• Falls back to legacy snippet_score comparison when baseline predates this change

test_stub_quality.yml — split gating and preview jobs

• test_snippets (gating): adds --stable-only; uploads results as snippet-results-stable-{run_id}
• test_snippets_preview (non-blocking): new job, continue-on-error: true, runs all versions; uploads as snippet-results-preview-{run_id}
• Both jobs set permissions: contents: read

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.