Add deterministic Polymarket recording and replay

[howdymary]

What changed

• adds an immutable autopredict.recording v1 capture contract for Gamma metadata, CLOB books, public trades, explicit gaps, and later resolution
• separates point-in-time snapshot capture from resolution capture and restores identity/completeness state across restarts
• validates Gamma market IDs, condition IDs, YES token IDs, CLOB market/asset identities, paginated trade filters, hashes, sequence continuity, and source timestamps
• adds deterministic replay into the canonical dataset contract with sticky partial-completeness semantics
• publishes capture/replay companion files through exact temporary-candidate validation and atomic directory rename
• constrains the default public transport to official HTTPS Polymarket hosts, GET-only behavior, no redirects, no ambient credentials, and strict public query/header allowlists
• adds official-shaped immutable fixtures, adversarial lifecycle/transport/atomicity tests, and recorder documentation

Why / root cause

AutoPredict had no first-party point-in-time recorder, so realistic evaluation and shadow replay could not distinguish complete evidence from polling gaps. Venue identifiers and public transport state also needed an explicit fail-closed boundary before captured data could be treated as reproducible evidence.

User and developer impact

Developers can capture public Polymarket observations, append resolution later, verify immutable hashes and completeness warnings, and replay the result deterministically. Partial or malformed captures remain ineligible for performance claims; no market data or missing sequence is fabricated.

Schema and migration

Introduces versioned recording/capture contracts and a documented replay projection into autopredict.dataset.v1. Existing evaluation datasets are unchanged. Recorder artifacts with unsupported completeness values, identity mismatches, discontinuities, or conflicting immutable targets are rejected with actionable errors.

Verification

• independent reviewer: APPROVED after repeated adversarial review
• focused recorder suite: 42 passed
• full Packet 5 suite before stack: 372 passed
• post-rebase Packet 4 + Packet 5 suite: 389 passed
• Black, compileall, scoped mypy, and git diff --check: passed
• sdist/wheel build: passed
• clean-wheel 98-module import audit: passed
• no authenticated request, live order, or production credential was used

Residual risks

• REST offset pagination cannot prove a transactionally frozen upstream trade set; artifacts therefore record polling sequence completeness as unavailable and surface truncation/gaps explicitly
• publication is atomic at directory-rename level but does not claim fsync durability or protection against a malicious local symlink race
• public recording remains read-only and is not evidence that live execution is safe

Tracks #2
Closes #9

Depends on #15.