Stage S spread runner + post-run analysis/media toolkit

[jkbennitt]

Summary

Codifies everything built during the 2026-06-10 Stage S 6-model spread so the post-run pipeline is repeatable:

• scripts/run_spread_n1.sh — the spread runner (was untracked since Stage S prep)
• scripts/analyze_spread.py — cross-model leaderboard: day-aligns against the no-agent baseline (tick-alignment is wrong: baseline ticks ~30x/day vs agent 2-3x/day), ranks by mean composite, reports raw + ex-artifact action success with a harness-failure marker list
• scripts/build_footage_index.py — maps console-log model markers + ticks to wall clock and OBS recording offsets for frame-accurate video cuts
• scripts/capture_run_media.sh — during-run loop: game-window stills (gdigrab) + per-model latest_tick.json snapshots (dashboard replay source)
• scripts/replay_ticks.py / scripts/record_dashboard.mjs — replay a finished run into the React dashboard and record it headless (Playwright; run with node, not bun)
• .claude/skills/run-analysis — the whole pipeline as a repo skill: Phase 0 capture setup → analysis → story mining with mandatory verbatim re-verification → footage index → EDL/ffmpeg/Remotion videos → publish prep
• benchmark_history.jsonl — the 6 spread run entries

Spread context

Leaderboard (N=1, mean composite): grok43 0.831 > opus48 0.826 > gemini35 0.825 > fable5 0.803 > gpt55 0.722 > deepseekv4 0.704. Actual OpenRouter spend $7.12. Key findings + queued fixes in #33; capture improvement idea in #34. Detailed artifacts in results/spread/ (gitignored; story bank + analysis also mirrored to the rle-media repo).

Test plan

• ruff check src/ tests/ scripts/ clean, mypy src/ strict clean, 458 tests pass
• analyze_spread.py smoke-run post-cleanup reproduces the leaderboard byte-for-byte
• Capture/replay/record scripts are the exact code used live during the spread (parameterized paths only)

🤖 Generated with Claude Code