Packet 3: canonical dataset evaluation

[howdymary]

Summary

• define the strict autopredict.dataset.v1 JSON-manifest + canonical-JSONL contract
• separate forecast-safe observations from resolution labels and fail closed on leakage/integrity errors
• add deterministic validate, evaluate, and deprecated canonical backtest CLI paths
• add provenance-rich baseline evaluation reports and exact dataset documentation

Safety and integrity

• enforces UTF-8/LF canonical bytes, hashes, counts, UTC timestamps, capture bounds, ordered non-crossed books, stable market-to-event identity, and resolution joins
• partial datasets validate structurally but cannot be used as performance evidence
• live execution remains disabled; no venue, credential, or order path is touched

Verification

• 330 tests pass
• all tracked/untracked Python files compile
• Black and scoped mypy pass
• isolated wheel import/CLI audit passes
• independent reviewer approved; independent tester passed adversarial and installed-package matrices

Parent: #2
Closes #6
Depends on: #13

This is a stacked draft PR based on codex/packet-2-release-integrity; retarget to main after its prerequisites merge.