docs(pbt): PBT effectiveness study — sensitivity vs realized yield

[cordialApple]

Retrospective on what property-based testing bought Scroll, on TreapOrderIndex. Lives in analysis/pbt/ + a new section in the repo README.

Two questions, two answers

• Sensitivity (could PBT catch faults?): ~100% by 10 cases, and earned — ablations show worst-class detection drop to 80% when the oracle or generator is weakened.
• Realized yield (did it fix real wrong code?): ~1, and unrecoverable from git. Squash-merge + an in-loop agent fixing defects before commit erased the broken intermediate states. That loss is the finding, not a hedge.

What is here

• harness/ mutation sweep (20 mutants, 5 classes, budgets 1–3000 × 10 seeds) + ablation, in-process esbuild + dynamic import vs a pristine array/Fenwick oracle.
• config/ PBT-only Stryker + vitest. results/ raw JSON. data/ star-schema CSVs.
• viz/ seaborn charts + composite dashboard (the README figure). powerbi/ version-controlled PBIP over the same schema.
• Root README.md retrospective section; package.json gains Stryker dev deps.

No app source touched.

🤖 Generated with Claude Code