build: turn on no-floating-promises and no-misused-promises, and fix what they found

[zizzfizzix]

Closes #277.

The pair type-aware linting was switched on for in #7, staged in DEFERRED with 128 and 35 violations. Both are now enabled with no options at all, in a new PROMISE_RULES group beside ALREADY_CLEAN, and pnpm lint is clean.

trackEvent returns void

Its body was already inside one try/catch, so the promise it handed back could not reject and no caller had anything to do with it — only the type said otherwise. The awaitable half is now captureEvent, which the tests use; the wrapper discards it. That settles 65 of the 128 floating promises with one signature rather than 65 voids.

queueEvent behind it needed no such change: every call site already awaits it, so the rule never had anything to say about it.

The rest were read one at a time

Work that genuinely has nobody to await it goes through the new src/utils/report-rejection.ts:

• reportRejection(task) — for a caller that cannot await: the synchronous background entrypoint, a JSX event handler, the message router that must return true before its dispatcher finishes.
• reportingListener(handler) — the same thing shaped for a callback position that ignores the return value: browser.*.addListener, storage.watch, addEventListener.

Both report the rejection, which is what the rules were warning about and what a bare void would still lose. A bare void is kept only where the callee already logs its own failures and resolves either way — getConsentState, getAllPresets, setupUninstallUrl — and each of those sites says so, following the comment already in install.ts.

Six that should have been awaited

Not discards — bugs the rules found, which now cannot come back:

• trackEvent from a content script never caught a failed sendMessage: the try wrapped an un-awaited call, so the catch was decorative.
• injectContentScriptToAllTabs on install could reject into nothing.
• Five E2E assertions ran against a tab bringToFront() had not finished raising.
• Two E2E fixture helpers returned before their storage write landed.
• The content-script test harness resolved as soon as the listeners had been called, not when they had finished answering.

No configured exceptions

no-misused-promises first went in with checksVoidReturn.attributes off, on the argument that the rewrite it wanted — onClick={() => void handle()} — dropped the same rejection as the async handler and only added syntax.

That argument does not survive the helper added alongside it. onClick={() => reportRejection(handleSave())} reports what the handler threw instead of discarding it, so the rule is asking for a real handler, not ceremony; a rejected click handler is exactly as invisible as the service-worker listener this rule is usually credited with catching. The third commit turns the option back on and routes all 23 JSX attributes through reportRejection. Two long inline handlers became named copyRow functions on the way, in DataTable and FullDataViewApp, rather than growing a wrapper around a 15-line arrow.

SidePanel passes one resetSystemPresets to both of its Footers. Wrapping at each site would have given the splash branch its own arrow that no test clicks; the two sites always wanted the same discard, so it is written once.

Two helpers, named late

Both went in under names that described the situation rather than the behaviour, and the last two commits fix that.

fireAndForget named the thing it exists to replace: void promise is fire-and-forget, and the entire argument for the helper — and for turning checksVoidReturn.attributes back on — is that a bare void drops the rejection and this does not. Its doc comment had to end with "plus the report that makes it worth writing" to correct its own name. It and asListener are now reportRejection and reportingListener.

settle, the test helper, had the milder version of the same fault. Settled is the Promise spec's word — it is what Promise.allSettled waits for — so beside the other two it read as promise vocabulary while settling nothing; and it promised an outcome that one macrotask turn cannot deliver once a handler waits on a real timer. It is now flushMicrotasks, which names the mechanism, and its comment says what it does not do.

Both renames are behaviour-free; each is the whole of its commit. The smell test is CLAUDE.md's own: when a comment is needed to explain what something does, the name is what should change.

Also here

AsyncMessageHandler for the two message dispatchers, since the router relies on them being async; tests/support/flush-microtasks.ts for tests that trigger a listener reportingListener left unawaited; CLAUDE.md updated.

The second commit is unrelated to the rules and fixes something that has been failing on main: ensureVisible in ConfigForm scrolls from inside a requestAnimationFrame and nothing waited for that frame, so whether its "nothing to scroll to" arm ran at all was a race. Since #285 that has been failing pnpm test:coverage intermittently — reproducibly with six busy cores, three runs out of three at 1749/1751, on main as much as here. The branch was always reachable; only the test was racy, so the fix drives the frame by hand rather than touching the component. It covers existing branches rather than adding any, so the denominator is unchanged at 1751.

Verification

fmt-check, lint, compile and build pass. test:coverage is at 100/100/100/100 with the fidelity check green, and holds under the load harness that reproduced the flake.

The E2E suite was not run locally: this environment's pre-installed Playwright browser does not match the version the repo pins, and the matching build cannot be downloaded here, so all 117 specs fail at launchPersistentContext before any spec code runs. The six E2E edits type-check and lint under the type-aware config, but CI is the first place they actually execute. Worth a closer look there than usual, because the JSX rewrites in the third commit touch handlers the E2E suite drives.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LakMrJTG2mdy4N6s1AHEnm