docs: add project CLAUDE.md (architecture, layout, build/test guide)

CLAUDE.md

@@ -0,0 +1,49 @@
+# scrabble-solver — project guide
+
+A Go library that, given a dictionary, a board position and a rack, returns every legal
+play ranked by score, and scores/validates arbitrary plays. The move generator is the
+**DAWG** algorithm (Appel & Jacobson) over `github.com/iliadenisov/dafsa` — a bit-packed,
+minimised DAWG with a compact ≤63-symbol alphabet. A GADDAG generator was also built,
+measured by self-play, and **removed**: DAWG won for this scoring-solver workload
+(~7× smaller, comparable speed) — see `RESULTS.md`.
+
+Module `scrabble-solver`, Go 1.26. Rulesets: English Scrabble, Russian Scrabble, and
+Russian **Эрудит** (`rules` package); Эрудит has no Ё tile and folds Ё→Е in its dictionary.
+
+## Layout
+
+- `scrabble/` — the public API: `Solver` (`NewSolver`, `GenerateMoves`, `ScorePlay`,
+  `ValidatePlay`), the `Move`/`Placement`/`Word` types, the DAWG generator and scoring.
+- `board/`, `rack/`, `rules/` — board grid (+ transpose), rack as per-letter counts,
+  and rulesets (geometry, premium layout, tile values/counts, alphabet, bonus):
+  `rules.English()`, `rules.RussianScrabble()`, `rules.Erudit()`.
+- `internal/` — `dictdawg` (build/load/serialise DAWGs over dafsa), `wordlist`
+  (encode/filter/sort/dedupe + `FoldYo`), `graph`, `dict`.
+- `cmd/builddict` — word list → serialised DAWG (`-alphabet latin|russian`).
+- `cmd/stress`, `selfplay/` — the self-play stress harness behind `RESULTS.md`.
+- `dawg/` — **committed** dictionaries: `en_sowpods.dawg`, `ru_scrabble.dawg`,
+  `ru_erudit.dawg` (Ё→Е folded). Rebuild with `make dawg`.
+- `dictionaries/` — `kamilmielnik/scrabble-dictionaries` git submodule (English source).
+- `dictprep/` — self-contained tooling that turns the Russian academic orthographic
+  dictionary into a common-noun word list. See `dictprep/README.md`. Committed output is
+  `dictprep/russian/{all,scrabble}.txt` (+ `orfo_dict_2025.{pdf,txt}`, `manual_confirm.txt`).
+  Running Stage 2 needs a Python venv with `mawo-pymorphy3` and the `libmorph` apt packages
+  (see `dictprep/README.md`).
+
+## Build & test
+
+    go test ./...            # all packages green; also run go vet ./... and gofmt
+    make dawg                # rebuild dawg/*.dawg from the word lists
+
+Scoring and move generation are validated against **real tournament games** in GCG format
+(`scrabble/gcg_test.go` + `scrabble/testdata/*.gcg`, including the 700+ club): for every
+move the test checks the score, the running total, and that the generator actually
+produces the played move with that score — canonical play, not invented cases.
+
+## Key facts
+
+- Compact byte encoding: low 6 bits = alphabet index; `0x80` = blank/wildcard (board, rack
+  and output bytes only — never inside the graph). The public API is byte-indexed.
+- DAWG is the production generator; the GADDAG was removed after measurement.
+- Detailed docs: `ALGORITHM.md` (the algorithm — single source of truth), `PLAN.md`
+  (design and decisions), `RESULTS.md` (DAWG-vs-GADDAG), `dictprep/README.md` (RU pipeline).