Defer uncommon admin icons and Block Kit chart code

[richardjoo]

Defer uncommon admin icons and Block Kit chart code

Motivation

The admin PluginRegistry entry currently includes the full @phosphor-icons/react namespace used to resolve plugin navigation icon names, plus Block Kit chart code even when no chart is rendered. This makes every admin visit download and parse code needed only by uncommon plugin icons or chart blocks.

Benchmark

I built the same production Astro/Cloudflare site twice with source maps and Rollup graph metadata, changing only the EmDash workspace packages. Sizes below sum the PluginRegistry entry and its static import closure; gzip uses level 9 and Brotli uses quality 11.

Metric	Baseline	Candidate	Delta
Entry raw	2,855,961 B	1,907,368 B	-33.2%
Entry gzip	796,756 B	485,414 B	-39.1%
Entry Brotli	633,278 B	381,927 B	-39.7%
Initial closure raw	3,362,679 B	2,796,350 B	-16.8%
Initial closure gzip	953,375 B	757,531 B	-20.5%
Initial closure Brotli	766,262 B	610,568 B	-20.3%

The complete admin-reachable graph stays effectively flat: gzip changes from 3,073,718 B to 3,068,896 B (-0.2%). This is code deferral, not code removal.

A separate repository fixture on the candidate reports a 1,861,446 B raw PluginRegistry, 2,745,418 B initial static closure, 153 initial Phosphor definitions, 32 lazy icon buckets (largest 217,820 B raw), and one lazy chart chunk.

Proposed design

1. Generate 32 deterministic Phosphor buckets from the exact installed @phosphor-icons/react export surface. Preserve canonical names, *Icon aliases, and aliases exported from differently named modules.
2. Resolve plugin navigation icons through the bucket matching the canonical module name. Existing first-party icons remain statically imported. Unknown names still fall back to Plug; a failed bucket load also falls back to Plug and does not unmount the admin.
3. Lazy-load Block Kit chart rendering behind React.lazy and Suspense. Reserve the configured chart height while loading; retain that placeholder if the chunk fails.
4. Pin the Phosphor version because generated deep imports and consumer-resolved root imports must use the same export surface.
5. Add a production-fixture bundle guard that inspects the emitted chunk graph and source maps. It enforces entry/closure ceilings, catches eager or full-catalog Phosphor regressions, requires all generated buckets and the chart chunk to stay lazy, and rejects stale CI builds.

Compatibility and failure behavior

• No public API or persisted-data change.
• Existing plugin icon-name behavior and aliases are preserved.
• Unknown icons retain the current Plug fallback.
• Chunk failures degrade to the icon or chart placeholder rather than an error boundary taking down the route.
• The generated files are checked during admin typecheck so a dependency update cannot silently stale the mapping.

Verification completed on the candidate

• Admin icon/Sidebar tests: 23 passed.
• Block renderer tests: 29 passed.
• Core Vite integration tests: 10 passed.
• Admin, blocks, and core typechecks pass.
• Generated icon mapping check passes.
• Production bundle guard passes with the metrics above.
• Formatting and git diff --check pass.

Repository-wide Oxlint currently aborts inside oxc_allocator on this machine even when constrained to one changed file and one thread; no lint diagnostic is emitted. I would rely on upstream CI for that gate before merge.

If this direction is acceptable, I have a single-commit branch based directly on current main, including tests and a patch changeset, ready for review.