[mcdemarco]

So my agent still liked its version better. Here is an actual PR in case you want to consider its improvements, probably after more of a shakedown of your current changes. The chatty details:

Results

Measured with upstream's scripts/bench-acity.ts, plus a small timing script for the roads display:

	Before	After
moves() on a fresh board	~120 ms	~12 ms
Rendering the roads display	~12 ms	under 0.2 ms

What changed

src/games/acity-road.ts

• New structuralConstraints(board) function. In one pass it finds every empty lot where building would break the road, and every lot where building would cut a structure off from it.
  • Road-breaking lots are the "cut points" of the empty-lot network: lots that are the only link between two parts of the road. They're found with Tarjan's articulation-point algorithm.
  • Isolating lots are lots that are the last empty neighbour of some building.
• It assumes a board reached by legal play, where the road is always one connected network.

src/games/acity.ts

• canPlace(), the off-colour dome check and the roads display all read one cached result instead of rebuilding the road graph for every candidate lot.
• The cache refreshes whenever the board's contents change. It compares contents rather than relying on move(), because tests sometimes edit the board directly; upstream's distance cache works the same way.
• Upstream's roadNetworkAfterPlacement is kept as the reference implementation.

test/games/acity.test.ts

• New test group checking, lot by lot, that structuralConstraints() gives exactly the same answers as the reference:
  • on every golden fixture;
  • at every position of two complete, repeatable test games (about 6 seconds).

test/fixtures/acity-golden.ts and acity-golden-oracle.json

• Removed the road-pinch-closed fixture, whose board has the road already split. Legal play can't produce that, so it's no longer supported.
• Its entry was removed from the recorded results by hand rather than by re-recording the file, so every other fixture is still checked against the results upstream originally captured, not against the new code.

Checks

• Upstream's golden-oracle tests, which compare against the original pre-refactor rules, still pass.
• Full npm test passes (2518 tests); lint and typecheck are clean.

Not changed

• test/fixtures/acity-bench-results.json, which records upstream's before/after numbers for their own refactor.
• Two smaller possible improvements, neither of which showed up as slow:
  • the distance cache rebuilds a board signature on every lookup;
  • claimWouldEndConstruction builds a whole new game for each claim it checks.