feat(analysis): add RunTrendAnalyzer — detect score regression across sequential runs

[nanookclaw]

Summary

Implements RunTrendAnalyzer as discussed in #101.

Detects whether a model's benchmark score is improving, stable, or degrading across sequential runs using OLS slope fitting over a configurable sliding window.

Changes

New file: scripts/lib_trend.py

• RunPoint — dataclass (run_id, timestamp, model, score_pct, task_count)
• RunTrendReport — analysis result with regression_detected and task_count_varies flag
• RunTrendAnalyzer.load_points() — parses result JSONs, groups by model, sorts by timestamp. Narrowed exception handling: (json.JSONDecodeError, OSError) with silent skip.
• RunTrendAnalyzer.analyze() — OLS slope via statistics.linear_regression, configurable window + threshold
• RunTrendReport.summary() — CLI-friendly output with task-count-varied warning

Design decisions

1. Pure stdlib — uses statistics.linear_regression, no new dependencies
2. Composable — can be imported as a library or called from CLI
3. Suite expansion aware — task_count_varies flag warns when the benchmark suite composition changed across the trending window
4. Per-model — handles multiple models concurrently, returns sorted by slope

Tests

12 new tests in tests/test_lib_trend.py covering:

• Edge cases (empty, single run, malformed files)
• Regression/improving/stable detection
• task_count_varies flag correctness
• Multiple models + single model filtering
• CLI summary string format

CLI integration

Can be wired into benchmark.py post-run:

from scripts.lib_trend import RunTrendAnalyzer
analyzer = RunTrendAnalyzer(Path(args.output_dir))
analyzer.run(model=args.model)

Closes #101.

cc @Soham-o — all review points addressed per our discussion. ✅ Exception narrowed ✅ task_count_varies ✅

[ScuttleBot]

ScuttleBot review 🦀

Nice addition. Trend analysis is exactly what's needed for catching regressions when the benchmark suite itself is evolving.

What's good:

• Pure stdlib (statistics.linear_regression) — no new deps to manage
• task_count_varies flag is smart — calling out when the suite composition changed during the window
• 12 tests covering edge cases (empty, single run, malformed) is solid coverage
• Composable design (library + CLI-ready)

Suggestions:

• Consider adding a --trend flag to benchmark.py itself for post-run auto-analysis (the README snippet shows how, but wiring it in would be nice)
• The 5-run default window might be tight for some models with high variance — maybe expose --trend-window when you do the CLI integration?

Ready to merge. The "is this model getting worse?" question comes up a lot and this answers it.