Follow-up to #4404 (closed). The lazy-only config-table deferrals are done (#4407/#4408/#4409, ~330 KB raw off the eager path). What remains is a different, harder class: large data pinned eager by eager boot services, not by lazy consumers. Each needs a behavioral refactor (lazy-load the table out of a synchronous boot path), so each is /ce-plan-worthy with real regression surface — not a clean lazy-cache like the shipped tables.

⚠️ Pursue only if a post-#4409 prod measurement shows a mobile LCP/TBT gap. If the ~330 KB already shipped hit the target, the risk here isn't worth it.

Targets

Out of scope (confirmed not deferrable)

panels.ts (~57 KB, panel-layout at boot), map-layer-definitions (~28 KB, sanitizeLayersForVariant() at App init), entities.ts (~20 KB, marginal + eager via cached-risk-scores).

Approach when prioritized

1. Measure prod mobile LCP/TBT (median-of-3, cold, throttled) after perf(dashboard): defer geo-map table bulk off the eager path (#4404) #4409 deploys — establish the gap, if any.
2. If a gap exists, /ce-plan the largest tractable target (likely feeds.ts via a boot-critical/deferrable catalog split).
3. Per-target: lazy-load the table inside the eager service (empty-until-loaded + retryable, the perf(dashboard): defer the ai-datacenters config table off the eager path (#4404) #4408/perf(dashboard): defer geo-map table bulk off the eager path (#4404) #4409 pattern), verify build (off dashboard.html modulepreload, not imported by main, 0 circular chunks) + runtime (the boot feature still populates) + the deferred path on Vercel preview where API-gated.