Round 2 of the eager-main.js diet. #4395 was closed by #4403, but #4403 only did round 1 — it deferred the search palette + Sentry init policy (~80 KB raw / ~15 KB br off main.js). That's ~4% of the eager JS. The score-moving bulk is still eager and untracked; this issue tracks it.

Current state (measured on prod after #4403, 2026-06-24)

• Deployed main-BX3KB0N8.js: 1,012 KB raw / 287 KB br (was 1,091 / 306 pre-perf(startup): lazy-load dashboard search #4403).
• Cold mobile (fresh Chrome, Slow 4G + 4× CPU, median of 3): FCP/LCP ~872 ms, CLS 0.007, TBT ~705 ms, DCL ~6.3 s.
• Delivery is solved (Brotli). The remaining bottleneck is JS parse/execute — TBT ~700 ms and DCL ~6.3 s are dominated by eager JS that compression doesn't help.

The targets (still eagerly modulepreload-ed, by raw size on prod)

≈ 780 KB raw eager — ~10× what #4403 moved. All four land in the same ~865 ms wave as main.js and feed the ~700 ms TBT.

Approach

Defer each behind a dynamic import() on the same pattern #4403 / #4386 (settings/checkout) / #4390 (map) already use. For each:

1. Audit what drags it onto the eager path — trace the static import chain from the main entry (and the modulepreload list / LAZY_HTML_PRELOAD_CHUNKS in vite.config.ts).
2. Confirm it is NOT first-paint-critical before deferring. Some may genuinely be needed early (e.g. if data-freshness drives an above-the-fold freshness indicator). Don't trade a TBT win for a visible-content regression — same discipline as the CSS-import-removal reviews. Where a piece is needed early, split the first-paint-critical slice from the deferrable bulk.
3. Buffer-and-replay any data that arrives before the deferred module loads (the latestSearchAdsb/latestTechEvents pattern from perf(startup): lazy-load dashboard search #4403).

Acceptance

• Measurable reduction in eager main.js transfer and its unused-javascript bytes.
• Mobile TBT drops from ~700 ms toward ~400 ms; DCL drops materially.
• Mobile Lighthouse Performance moves from low-80s toward ~90 (the perf: cut homepage LCP from 9.9 s → < 3 s (Lighthouse 50 → 90) #3987 target).
• Verified by a median-of-≥3 cold mobile run (single throttled samples are too noisy — TBT varies hundreds of ms run-to-run) plus the deterministic bundle-size delta.

Refs

#4395 (closed, round 1 = #4403), #4379/#4386 (prior diet), #3987 (homepage LCP→<3s / Lighthouse→90 tracker), #4335 (split/stage initial JS).

https://claude.ai/code/session_015Ae5Xavw1ZV4GMCWEsgKwe