games_hub: rummy dealer's-choice table (#1609)

[aaylward]

Closes #1609. UI: muchq/muchq.github.io PR on branch claude/rummy-dealers-choice.

Every rummy table is dealer's choice:

• Table libs/cards/rummy/table.{h,cc}: startGame opens it choosing, seat 0 dealing. The dealer's chooseVariant deals; the seat after the dealer opens. A deal won by play credits the winner a hand and passes the deal on. Below two seats the table closes. The dealer's chair follows its holder on leaves, including mid-deal. A dealer the room shows disconnected lets any seat deal.
• Rule change: seven cards a seat at every table size (was ten at two).
• Serde table_serde: v2 row. A v1 row (one deal) loads as a one-deal table; ten-card rows still load.
• Wire model/rummy.smithy: chooseVariant; the view gains variant, dealNumber, standings, choosing, lastDeal; phase choosing. A deal's end rides the view's lastDeal; gameEnded is the table's (standings, dealsPlayed).
• Hub: each deal is a game to room stats and the event log (game_started on deal, game_finished on its end), credited in the deal-ending commit. The room listing restages on deal start and end.
• Stats: a member upsert updates presence only; stats move only by a finish's increments. A sibling's stale presence write could erase a remote deal's credit, which rummy now makes common.

Tests:

• table_test, table_serde_test (frozen payload, v1 rows, hostile rows), game_state* for seven cards.
• rummy_e2e_test: next dealer's choice and rotation, away dealer, dealer leaving between deals, table close; whole seeded games at 2/3/4 seats against the engine.
• rummy_wire_test: goldens for the choosing view, refusals, dealt views, a deal's end, the table's end.
• hub_e2e_test: event log records each deal as a game.
• Postgres-gated: pg_hub_store_test (PresenceWritesLeaveStatsToTheirIncrements), pg_hub_e2e_test (restart mid-deal). Ran locally against Postgres 16.

Locally: bazel test //domains/games/... 181/181 with Postgres and CI=1; mutation checks on dealer rotation, away dealer, room staging, event log and the empty-turn guard. Played end to end in headless Chromium against a local hub.

Review panel ran on the first push; its survivors are the last commit. Not fixed here:

• A dealer whose instance crashed stays connected in the room (golf_hub: restored members are never reaped — a deploy turns the 5-minute grace into forever-membership, breaking room share links (muchq.github.io#260) #1295), so other seats can't deal; leaving is the way out.
• A deal-start or deal-end commit that reports unavailable but landed loses its event line (golf's existing unknown-fate behaviour).

Gin (#1610) and Rummy 500 (#1611) are next.

---

Generated by Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	ast-lol	f922348	Commit Preview URL Branch Preview URL	Sep 29 2026, 12:07 AM

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	1d4-web	f922348	Commit Preview URL Branch Preview URL	Sep 29 2026, 12:05 AM

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	iili	f922348	Commit Preview URL Branch Preview URL	Sep 29 2026, 12:06 AM

[cursor]

Review panel + Bugbot (7cdf7552)

Four lenses + Bugbot. Bugbot found no bugs. Correctness and resource-safety clean — review-fix commit still holds (mid-deal dealer pass, presence-only upsert). Altitude is right: TableState over one-deal GameState, deal-as-game stats/events with lastDeal vs table gameEnded, Gin/500 correctly underbuilt. Known won’t-fix (#1295 stuck connected, golf-style unknown-fate event lines) accepted as documented. build-and-test / sanitize still running at post time.

Verdict: LGTM with test fold-ins. Design and the presence/stats race fix look solid; leftovers are unpinned hub refusals / standings.

Survivors (verified):

1. Tests — chooseVariant during a live deal unpinned at hub/wire. Engine has NoChoosingWhileADealIsInPlay (status code only). Wire/e2e pin non-dealer choice and unknown variants, and refuse draws between deals, but never send chooseVariant while phase == "playing". Assert "not between deals" on the stream (and pin the reason in the unit test).

2. Tests — leave mid-deal never pins non-zero standings on gameEnded. ALeaveMidDealClosesTheTableWithItsStandings / LeavingMidGameAbandonsItWithNoWinner close on deal 1 with handsWon: 0. After a won deal, leave mid-next-deal is untested — zeroing getWins() on close would stay green. Compaction while choosing is covered; close-with-prior-wins is not.

3. Tests (soft) — between-deals view has no KeysOf / view-level variant pin. Dealt view has a full golden + KeysOf. Deal-end asserts choosing / lastDeal / standings dumps only; dropping or renaming view-level variant between deals would not fail the wire suite the way a dealt-view rename would.

Altitude (named, ship): Seven-card at every seat size is hitchhiking on dealer’s choice (Gin wants ten at two anyway) — fine as the named companion it already is; serde still loads ten-card rows. Presence-only ON CONFLICT is the right altitude for the stats race; Memory still full-replaces (harmless single-writer).

Line comments below. Panel complete.

[cursor]

Recheck at f922348b

Closed:

1. Mid-deal chooseVariant — unit pins "not between deals"; e2e refuses it on the stream.
2. Close with prior wins — bob leaves mid-deal 2; gameEnded keeps alice’s handsWon: 1, dealsPlayed: 2.
3. Between-deals KeysOf + view-level variant.

LGTM.