Stage 2: engine package over scrabble-solver (registry, bag, Game, replay)

backend/internal/engine wraps the sibling scrabble-solver library in-process:

- Registry: versioned DAWG load via dafsa.Load, keyed by (variant, dict_version),
  latest-per-variant; English / Russian / Эрудит handled uniformly.
- Bag: own deterministic, seeded tile bag with Draw + Return (for exchanges),
  since the solver's self-play bag cannot return tiles.
- Game: pure rules engine — deal, play/pass/exchange/resign, refill, per-move
  scoring, turn order, and end-condition detection (empty bag + empty rack, six
  scoreless turns, resignation) with end-game rack adjustment.
- decode/ReplayBoard: dictionary-independent MoveRecords and board replay via
  scrabble.Apply (no internal/encoding), realising ARCHITECTURE §9.1.

Wiring: go.work gains "replace scrabble-solver => ../scrabble-solver"; backend
requires scrabble-solver (placeholder) and github.com/iliadenisov/dafsa directly.
Both Go CI workflows clone the public solver sibling (master HEAD, no token) and
set BACKEND_DICT_DIR.

Docs: ARCHITECTURE §5/§14, TESTING engine layer, backend README, and PLAN
refinements + deferred TODOs (publish/version solver; split engine vs dictionary
generator).

docs/TESTING.md

@@ -15,10 +15,17 @@ tests or touching CI.
   by a separate CI workflow (`integration.yaml`; Ryuk disabled, serial). Slow.
 - **UI** *(introduced with the UI in Stage 7)* — Vitest (unit) + Playwright
   (e2e), mirroring the chosen plain-Svelte + Vite toolchain.
-- **Engine** — correctness of scoring and move generation is owned by
-  `scrabble-solver`'s own GCG-backed tests. The backend adds regression tests
-  for end-conditions, the 24-hour timeout / auto-resign, robot balance and
-  margin targeting, and **dictionary-independent history replay**.
+- **Engine** *(Stage 2+)* — correctness of scoring and move generation is owned
+  by `scrabble-solver`'s own GCG-backed tests. `backend/internal/engine` adds, on
+  top of the embedded solver: per-variant smoke tests (load all three committed
+  DAWGs and validate a known word, including Эрудит), bag draw/return determinism
+  and exchange accounting, the `Game` end-conditions (empty bag with an empty
+  rack, and six scoreless turns) with end-game rack scoring, and
+  **dictionary-independent history replay** (`ReplayBoard` reproduces a full
+  greedy game's final board from decoded records alone). The engine tests read
+  the DAWGs from `BACKEND_DICT_DIR` (or the sibling `scrabble-solver` checkout)
+  and fail loudly when it is absent. The 24-hour timeout / auto-resign and robot
+  balance/margin regression tests arrive with those stages.
 
 ## Principles