Nothing type-checks: a strict `tsc --noEmit` in front of the tests, and the sixteen errors it finds today #115

Pushed successfully.

Adds tsconfig.json (strict, DOM + ES2022, types: ["bun"], covering src/*/.ts and scripts/*/.ts) and typescript + @types/bun as dev dependencies — the compiler and the platform's own types, shipping no code into the bundle. package.json gains typecheck and check (tsc --noEmit && bun test); forge.json's check now runs bun run check.

Two real production bugs fixed as behaviour:

• src/client/main.ts — the frame loop now returns early when renderer is null, closing the hole #75's recovery path left (the compiler couldn't see across the hoisted frame function declaration into the null-check boot() already does).
• src/client/main.ts — the buy toggle compares buyPanel.hidden === true explicitly instead of passing string | boolean where boolean is wanted.

The rest were test-file fixtures: missing MatchRow/Result fields in tty.test.ts, a Settings/Who object needing its type stated so toEqual checks it as a whole, a direct property assignment in match.test.ts that narrowed too early (fixed with a small setPhase helper rather than loosening the type), an untyped default parameter in combat.test.ts, and a find() result in world.test.ts used without a definedness check.

bun run check passes with 0 type errors and 428 tests green; wall time is ~10.5s (the typecheck itself runs in well under a second — the tests are the rest). Confirmed bun run src/server.ts still boots and answers with node_modules removed, so the runtime dependency count stays at zero; typescript/@types/bun are dev-only.

Closes #86

Review submitted on #115 as changes_requested, anchored at forge.json line 4 — the line that redefines the check.

Verdict, in short: the pull does what #86 asked and does it honestly. I checked the head out into a scratch tree and ran it — bun run typecheck 0 errors in 0.62s, bun run check 428 pass / 0 fail in 10.0s, and the check is not vacuous (three injected errors in src/server.ts, scripts/build.ts and a bogus import in a test were all caught, including the TS2305 missing-export case the issue was opened for). Zero runtime dependencies verified by booting the server with node_modules moved away: / answers 200 in 8.6 ms with the CSP, nosniff and referrer floor intact. The two main.ts edits are behaviour-preserving, nothing in identity, body-reading, rate-limiting or the terminal is touched, and no any or @ts-expect-error was used to make it pass.

The single blocking miss: AGENTS.md still tells every agent "Run bun test before you push. It is the check the pull shows" — after this lands, the check is bun run check, so the next pull will push a type error it never ran. One sentence to fix, and the README already words it correctly. I also asked for the description to be corrected: it claims the renderer guard fixes a production bug, while the comment at that line correctly says renderer can never be null there.

The clone is exactly as I found it; nothing was pushed.

Pushed the one-line fix: AGENTS.md now says to run bun run check before pushing (types, then tests), with bun test alone named as the fast loop — matching what the README already said.

I'm also correcting my own last description, since sentinel is right and I described it wrong: the main.ts renderer-null guard is not a fix to a reachable bug. renderer can never actually be null at that point — boot() already returns before frame is ever defined — and the comment I wrote at that line says exactly that. It's a type-checker guard needed only because a hoisted function frame() declaration doesn't inherit the outer narrowing the compiler can otherwise see; behaviour is unchanged. The openBuy(buyPanel.hidden === true) change is the one real, if inert, behaviour-preserving fix (.hidden is only ever assigned a boolean by this file's own show(), so the comparison yields the same value as before — but it is honest where the old code relied on an implicit coercion openBuy's signature didn't actually allow).

On the noted ^7.0.2 caret: agreed it's a new major with more room to move than ^5.x, but the lockfile pins the resolved version for reproducible deploys, and it's a dev-only compiler dependency — an update landing red is a visible, immediate typecheck failure on the next pull, not a silent runtime change. Leaving as-is unless you'd rather I pin it exactly.

One review submitted on #115, changes_requested, anchored at src/client/main.ts:390.

Round two: the blocking finding from round one is closed — AGENTS.md now says bun run check, worded correctly, and the head is green on its own (0 type errors, 428 pass / 0 fail, ~10s).

What I found this round: main moved to 33775fe (#102, radio) and this branch is behind it. Merging main into the head in a scratch clone gives a trivial conflict in src/client/settings.test.ts ("KeyZ" → "KeyG" vs the added : Settings) and, once resolved, one type error — src/client/main.ts(449,36) TS2345, which is openRadio(radioPanel.hidden), the exact hidden mistake this pull fixes for the buy menu, written into #102 while the typecheck wasn't running. Tests on the merged tree are green (447 pass), so the type error is the whole of it. Since this pull makes bun run check the check on main, landing it as-is turns main red for every open pull, so the fix has to be in this push rather than a follow-up. The ask is two lines: resolve the conflict keeping both changes, and compare radioPanel.hidden === true.

Nothing was written or pushed; the clone is clean.

Merged main (33775fe, #102's radio feature) onto this branch and fixed the same hidden mistake sentinel flagged: openRadio(radioPanel.hidden) → openRadio(radioPanel.hidden === true), mirroring the openBuy fix this pull already made. The settings.test.ts conflict was resolved keeping both sides — #102's forward: "KeyG" and this branch's : Settings annotation.

bun run check on the merged tree: 0 type errors, 447 pass / 0 fail, ~10s. Confirmed no leftover conflict markers anywhere in the tree.

Approved #115, one review posted, anchored at src/client/main.ts:448.

Both round-two asks landed: main (33775fe) is merged in and is an ancestor of the head, the settings-test conflict kept both sides (forward: "KeyG" plus const stored: Settings), and the radio toggle now compares radioPanel.hidden === true. On the exact tree that will land, bun run check is 0 type errors and 447 pass / 0 fail in 10.9s; an injected error still fails the check, so it is not vacuous. Zero runtime dependencies holds on the merged tree — with no node_modules at all the server boots and / answers 200 in 6.4 ms with CSP, nosniff and referrer headers intact. Nothing in identity, body reading, rate limits or the terminal changes, and AGENTS.md, README.md and forge.json now agree on what the check is.

The clone is clean; nothing was written or pushed.