Add invariant tests for colorGuess and the reducer

[jrr]

Second of three test-focused PRs (#439 was the first). Tests only — no production code changes, no new dependencies. The one non-test change is a one-line fix inside test-util.ts, explained at the bottom.

Where the existing tests pin down specific cases, these state the rules those cases are circling and check them across many inputs, using a small seeded LCG so failures reproduce exactly.

colorGuess

Three rules, checked over all 144 pairings of a duplicate-heavy word set (EERIE, GEESE, TEETH, MUMMY, KAYAK, …) and 2000 seeded pairs drawn from the real word lists:

• one tile per guessed letter, in the order guessed
• green marks exactly the positions the guess already got right
• Wordle's duplicate-letter rule — a letter earns min(count in guess, count in solution) coloured tiles, and no more

That third one is what the eight existing examples approximate. Before writing it as a test I brute-forced it across 92,600 real word pairs to confirm it actually holds rather than merely looking plausible.

Deliberately not asserted: "a gray letter is absent from the solution". That isn't true — a letter goes gray once its occurrences are used up.

reducer

Games are played out with random actions and stop as soon as they end, so the new-game key can't reset the boards mid-sequence and legitimately break these.

Per state: the current row fits in a word; every guessed row is a full word.

Per move: rows never disappear; a move adds at most one row; an already-solved board takes no more guesses and stays solved.

The guess-limit rule is split in two. Single-board games stay inside the limit. The multi-board case is another it.fails for finding 1.1 — the broad form of the targeted case from #439: no random play of a multi-board game stays inside its limit.

Verified by mutation

An invariant that passes against broken code is worthless, so each was checked against a deliberate break:

Mutation	Caught by
colorGuess stops consuming a matched letter	duplicate-letter rule (both invariant tests; only 1 of the 8 examples)
won boards no longer frozen	"an already-solved board takes no more guesses"
row-full clamp removed	"the current row fits in a word"

One invariant I first wrote was simply wrong — "a solved board takes no guesses" fired on the winning move itself, since a board becomes won on the same transition that appends the winning row. The test caught it; it's now expressed over state transitions rather than snapshots.

Why expectEqual changed

Assertions carry their inputs alongside the value being checked, so a failure inside a 2000-iteration loop names the pair that broke it:

{ "coloured": 1, "guess": "KNEES", "letter": "E", "solution": "BASTE" }

That made expectEqual's swapped arguments (finding 3.3) actively misleading — the correct value was labelled "Received" and the buggy one "Expected". Since diagnosable failures are the point of this PR and test-util.ts was already being edited for the RNG, the one-line fix is included. Outcomes are unchanged; only the labels move.

Checks

yarn build, yarn lint, yarn prettier --check . and the --test midgame --quit smoke run all pass. 30 → 36 passing plus 3 expected failures, suite still under a second.

Next, as a separate PR: extracting keyToAction and the CLI flag handling into pure functions so they can be tested directly.

---

Generated by Claude Code