[appleweiping]

Summary

• Add a bounded local what-if API and CLI to compare up to eight independent expert-capacity and declared-conflict scenarios against one unchanged baseline.
• Re-run the existing matching engine and independent assignment audit for each scenario; report assigned/unmet demand, diagnostic reasons, workload, total score, and added/removed assignment pairs.
• Reject malformed/unknown/no-op changes and input/output aliases; include a synthetic example, documentation, and CI smoke.

Scope and limitations

This is a deterministic local sensitivity comparison, not causal attribution, an interactive web UI, OpenReview write-back, or parity with the frozen OpenReview Matcher solver families. It never infers whether removing a real-world conflict is appropriate. The precomputed embedding-provider slice already exists in this repository; this PR addresses a separate roadmap gap.

Validation

• Windows Python 3.13: 925 passed, 1 skipped (file symlink unavailable); targeted what-if tests: 13 passed.
• Pure branch coverage: 2267/2504 = 90.54%.
• Ruff, strict mypy, Bandit, uv audit --locked, wheel/sdist build, real synthetic CLI smoke (three scenarios), and clean-wheel import/CLI smoke passed.
• Signed and DCO-signed commit.

Assisted-by: OpenAI Codex