core/scraper: typed game-record types for scraped-game persistence (SQLite schema)

[valmathieu]

Problem

The roadmap requires the scraper to materialize observed games into contrai_core objects before persisting, feeding supervised learning. But no serializable record of a game exists today: PlayState is a live in-memory structure full of object references, Trick holds (Player, Card) tuples, and nothing round-trips to disk. The SQLite schema for scraped games is an open question (CLAUDE.md §11).

Proposal (to design when scraper observation work opens)

Frozen dataclasses in contrai-core — e.g. BidEvent, PlayEvent, RoundRecord, GameRecord — built on Position, TeamSide, Card, Bid. The record types then are the schema:

• The SQLite table design falls out of the dataclass fields (games / rounds / bids / plays / players).
• Later they double as the web server's wire format and the supervised-learning dataset row format.
• Scraper stores raw DOM payloads / structured events first, then parses into these records (reproducibility).

Timing

Deliberately not now — bidding/play observation in the scraper are still # FUTURE LOGIC placeholders, and the right record shape depends on what is actually observable. Design this together with that work. Blocked by / benefits from: Position enum, TeamSide enum.

Acceptance criteria (when picked up)

• Records are frozen, fully typed, and round-trip to/from SQLite without loss.
• A scraped game replays through the engine's rules without contradiction (validation harness).
• Schema documented in docs/packages/scraper/.