refactor(chat): one centering module, one closed loop

[chrisnojima]

Problem

"Center the thread on message X" was a five-module conversation with three representations of the target and three independent closed-loop scroll correctors.

• center-context.tsx derived an ordinal from a messageID by identity cast — not the same as the real resolver in thread-message-state.tsx, which neither it nor thread-context called.
• list-area/index.tsx ran a desktop corrector, a native corrector, and a second native loop driven by onViewableItemsChanged on a [50,250,500,900] timer ladder — each with its own cancellation token, retry budget and deadband.
• Nothing outside list-area could observe whether centering succeeded, so search.tsx advanced its n of m counter on !!message?.id — for a hit the list might be unable to reach. The clamp was silently absorbed by a pinnedChecks >= 3 check.

Change

The interface returns the outcome:

centerOn(messageID, highlightMode): Promise<'centered' | 'clamped' | 'not-found'>
useConversationCenter() // → {centeredOrdinal, centeredHighlightOrdinal, hasCenter}

Three correctors become one loop behind a single adapter — {measureTarget, scrollToIndex, scrollToOffset} — with exactly two implementations, desktop (LegendList) and native (FlatList). measureTarget returns measured | offscreen | pending; pending is native's answer while its viewable range is stale, so the shared loop waits rather than correcting off a stale reading.

The identity cast is deleted; the target resolves through the real getOrdinalForMessageID, and the row must also be present in messageOrdinals. pinnedToEndRef and the end-anchor loop become the module's own ScrollOwner state.

Ported verbatim

8px desktop deadband; native's half-row deadband and 0.9 damping; 3 settle checks; 3 pinned checks → clamp; 3s budget; native's 12-correction cap; the onScrollToIndexFailed retry (cap 5, reset per target); abort on wheel / onScrollBeginDrag; the end anchor's 2000ms/50ms/2-correction budget.

Bug found and fixed

Re-centering on the message you are already on left the promise hanging until the not-found watchdog fired — the effect was keyed on the ordinal alone, which does not change. Covered by a test that fails on either half of the fix.

⚠️ Needs a device check

On native the [50,250] and [50,250,500,900] ladders and the onViewableItemsChanged corrector are replaced by the shared 50ms poll loop. It is a superset of the old firing schedule under the same caps, but this has not been verified on a simulator or device. The desktop path is covered by the corrector tests.

Validation

lint:all clean — 0 bailed out, 0 whole-props deps, tsc clean both projects.
jest --runInBand — 230 suites / 2258 tests.

The file had no tests. Added 17 (deadband, settle count, clamp, offscreen, stale/pending, abort, correction budget, timeout, end anchor, outcomes through the provider, the ordinal resolution the identity cast got wrong) plus 3 in search.test.ts. Every one mutation-checked.