games_hub: rummy (#245)

[aaylward]

Closes #245. UI: muchq/muchq.github.io PR on branch claude/rummy-game-backend-ui-fvxk0l.

Basic rummy as the hub's third game, on castle's terms:

• Engine libs/cards/rummy: 10 cards at 2 seats, 7 at 3–4; draw stock or discard; melds (sets, runs, ace low or high, no wrap) and lay-offs onto anyone's meld; discard ends the turn; the card taken from the discard can't go straight back unless it's the last in hand; an empty stock is the discard turned over under its top; first empty hand wins and scores the others' deadwood. Versioned serde for the games row.
• Wire model/rummy.smithy: rummy envelope member; moves name cards (castle: the wire should name cards, not the engine's slots #1505). The view redacts other hands and the stock, and a stock draw's lastMove names no card.
• Hub: castle's move loop generalised to TableEngineMove<Engine> / LifecycleMove<Move>; castle behaviour unchanged (castle_e2e green). Rummy points ride gameEnded and stay out of the room's running total (golf's lower-is-better scale).
• Vocabulary: stats hubVariants, deja VARIANTS, otel_contract pin, prom_proxy Rummy tiles.

Tests:

• engine, meld and serde unit tests (frozen payloads, hostile rows)
• rummy_e2e_test: whole games against an engine mirror at 2/3/4 seats; also redaction, refusals, refill, abandonment, resume
• rummy_wire_test: raw-byte goldens
• model/counter pins in hub_e2e_test
• Postgres-gated: pg_hub_store_test and pg_hub_e2e_test (restart mid-turn). Ran locally against Postgres 16.

Locally: bazel test //domains/games/... with Postgres, stats/otel_contract/deja/prom_proxy tests, --runs_per_test=10 on the rummy stream suites, and a mutation-check over hub redaction and engine rules. The only surviving mutant is equivalent.

The review panel ran against the first commit; its survivors are the second commit. Cursor's panel fold-ins are the last.

Rummy variants are planned separately (#1609–#1611).

---

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	iili	7a71af6	Commit Preview URL Branch Preview URL	Sep 28 2026, 08:05 PM

[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	7a71af6	Commit Preview URL Branch Preview URL	Sep 28 2026, 08:05 PM

[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	7a71af6	Commit Preview URL Branch Preview URL	Sep 28 2026, 08:06 PM

[cursor]

Review panel + Bugbot (bbbbfc07)

Four lenses + Bugbot. CI green. Bugbot found no bugs. Correctness and resource-safety clean — author-panel fold-ins in 06b13b6d (serde lastMove shape, stuck-row reject, turn-before-stale-card) still hold. Altitude is right: rummy-on-castle-terms, TableEngineMove/LifecycleMove earn their keep at N=2, points stay off golf’s running total.

Verdict: LGTM with small fold-ins. Engine/wire/e2e/PG/vocab look solid; leftovers are docs drift and a soft dashboard ask.

Survivors (verified):

1. Docs — cards/rummy/README.md drops the last-card discard exception. It says the taken discard “may not go straight back,” full stop. game_state.h, rummy.smithy (RummyDrawDiscard), and Discard.TheTakenCardMayGoBackWhenItIsTheLastInTheHand allow the exception. The README claims to restate the GameState contract.

2. Docs — game_events.h still says variant is "golf" or "castle". FinishedOf / GameKindName include rummy; dedicated rummy finished tests exist. Fold the comment.

3. Tests — EveryEventsLineIsTextWithNothingToEscape still asserts variant == golf|castle. The test name claims “every ending of every game”; rummy endings are covered elsewhere, but that loop never names "rummy". Add the third variant (and a GameStartedLine golden if those stay variant-specific).

4. Altitude (soft) — prom_proxy tiles. Day-one rummy_commands/rummy_events are declared and vocabulary-pinned, but registry.go still selects only golf_|castle_|lobby_. Castle’s tenant landing (games_hub: castle joins the room stream (#77 phase 2) #1492) shipped tiles with the series. Equating this to voice_* undersells it — voice is room-layer, not a GameKind. Fold the Castle-sized tiles/selector, or a same-day follow-up; not a playability blocker.

Thin (won’t hold LGTM): hub e2e never asserts "that card does not fit that meld" (unit only) or the empty-stock-with-nothing-under-top refusal (WithNothingUnderTheDiscardTopThereIsNoStock). Refill and "no such meld" are covered.

Line comment below. Panel complete.

[cursor]

Recheck at 7a71af62

Closed:

1. README — last-card exception restored.
2. game_events.h — "golf", "castle" or "rummy".
3. EveryEventsLineIsTextWithNothingToEscape — rummy started + completed/abandoned endings in the loop; variant assert includes "rummy".
4. prom_proxy — Rummy tiles, selector rummy_, allowlist, handler tests.

Thin e2e refusals staying at unit level is fine.

LGTM.

[aaylward]

sanitize (tsan) on 7a71af6 died fetching opal_cpp before any test ran: Unknown host: github.com (runner DNS). Not this PR's; the same job passed on bbbbfc0, and 7a71af6 only touches docs, a test and prom_proxy. I can't re-run it (403) — needs a manual "Re-run failed jobs".

---

Generated by Claude Code

[cursor]

CI note (7a71af62)

sanitize (tsan) failed before any test ran: Bazel Unknown host: github.com fetching opal_cpp (exit 37). Sibling asan/ubsan on the same tip passed — infra/DNS flake, not a rummy race. LGTM still holds; rerun that job when convenient.