chore: drop 19 MB of committed per-case scores.jsonl from incident-triage demo

[Chang Liu (changliu2)]

What

The four incident-triage variant runs committed their full per-test-case judge outputs (\scores.jsonl, ~5 MB each = 19.1 MB total), bloating every clone and CI checkout. This untracks them and re-ignores the pattern.

Kept (the README + viewer rely on these)

• \metrics.json\ — the aggregate numbers cited in the README results table
• \manifest.json, \config.yaml\ — per-variant run metadata
• \ rade_off.png\ — the chart the README embeds (unchanged)

\scripts/render_trade_off.py\ already falls back to the documented case-study numbers when \scores.jsonl\ is absent (see its module docstring), so the chart still re-renders.

Tradeoff

Viewer per-case drill-down on these four pre-computed demo runs now shows summary metrics only (no per-row judge data). Re-run the suite locally to regenerate \scores.jsonl\ (gitignored going forward).

Scope

• 4 files deleted, 19.1 MB
• .gitignore: re-ignore \�xamples/incident_triage_agent/artifacts/**/scores.jsonl\ (the explicit artifact negation above it still allows the small curated files)
• Local working copies preserved (\git rm --cached\ only)

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[Jake Present (jakepresent)]

Approved. I verified the ignore rule behaves the way the PR intends: the generated incident-triage scores.jsonl files are ignored, while the curated metrics.json, manifest.json, config.yaml, and trade_off.png artifacts remain trackable. I also checked there are no tracked incident-triage scores.jsonl files left and ran python scripts/render_trade_off.py --suite incident-triage-agent-v1 --out /tmp/incident-triage-tradeoff-pr233.png to confirm the fallback path still renders without the committed score rows.

One non-blocking follow-up worth considering later: the incident-triage README still mentions scores.jsonl in a couple of artifact path summaries. That is still true for freshly regenerated local runs, but less true for the checked-in demo snapshot after this cleanup.

[Chang Liu (changliu2)]

Addressed your non-blocking follow-up in bc4aede (thanks for the thorough verify on the render fallback).

The two per-variant artifact summaries (Appendix B.1 naive-prompt and B.2 guarded-with-shield-gepa) listed scores.jsonl among the committed files. After this cleanup it's gitignored and regenerated locally, so both summaries now list the committed curated files — {config.yaml, manifest.json, metrics.json} — and note scores.jsonl is regenerated on a fresh run. The render-fallback mentions (the scripts/render_trade_off.py section) and the general "each run writes … scores.jsonl" line are left as-is since they're accurate for fresh local runs.

Docs-only change on top of your approval — re-requesting a quick re-approval per the new commit.

[Minsoo Thigpen (minthigpen)]

Approved docs-only change on top of Jake's first approval.