Add reducer tests; require solutions to be passed into newGame

[jrr]

First of three test-focused PRs. Small refactors + new tests only, all using the existing vitest setup — no new dependencies or tooling.

newGame now requires its solutions

newGame takes one solution per board, and the board count derives from the array's length:

newGame({ solutions: ["CIGAR", "REBUT"], numGuesses: 3 })

game-states.ts no longer imports game-logic.ts, so the state layer is pure data construction with no dependency on the word lists. Randomness moves to the two call sites that actually want it, via a new pickSolutions(n) helper. Three call sites changed: ui.tsx, the reducer's new-game branch, and display-rows.test.ts (which previously depended on a random word — harmless, since it only checks row shapes, but a latent flake).

This is code organization, not testability. The reducer's new-game branch still reaches Math.random() — explicitly now rather than transitively. Making that path deterministic means putting the solutions in the action, which splits the "N only works once the game is over" rule across the reducer and the key handler. Better designed alongside deterministic daily mode, so it's deliberately not here.

While changing the ui.tsx call site, useReducer switched to its lazy-init form. The initial-state argument is re-evaluated on every render, so picking words inline drew throwaway solutions on every keystroke.

src/state/reducer.test.ts

The reducer is the core state machine and had no coverage at all. Because the bug below needs four actions in sequence to appear, these drive the reducer through action sequences rather than testing it one call at a time:

const guess = (state: GameState, word: string) => submit(type(state, word));

Covered: typing and row-full clamping, backspace bounds, submission before the row is full, valid and invalid words, note lifecycle, per-board colouring, single- and multi-board wins, guess exhaustion, give-up, and the N/Q keys in terminal states.

Two known bugs recorded as it.fails

Rather than leave these undocumented or turn CI red, they're written as it.fails. Vitest reports them as "expected fail", so CI stays green — and each one starts failing the moment the bug is fixed, which forces the flip to it.

• Finding 1.1 — a two-board game runs past its guess limit when board 0 is solved first. onFinalGuess counts rows on gameBoards[0], which stops growing once that board is won. The renderer then throws RangeError on a negative blank-row count.
• Finding 1.4 — the invalid-word note survives every later keystroke.

Both were confirmed to fail for the intended reason, not incidentally.

Checks

yarn build, yarn lint, yarn prettier --check . and the --test midgame --quit smoke run all pass. Tests go from 11 to 30 passing plus 2 expected failures.

Since the smoke run takes the initialState path and never touches newGame, the real path was verified separately: board counts, the numGuesses default (1→6, 2→7, 3→8, unchanged), the --num-guesses override, lowercase normalization, and the reducer's N path all behave as before.

Note that newGame({ solutions: [] }) still yields zero boards and crashes on the first render — finding 1.3, unchanged and simply relocated by this PR.

Next up, as separate PRs: invariant tests for colorGuess and the reducer, then extracting keyToAction and the CLI flag handling into pure functions so they can be tested directly.