perf(console): free the declared-lazy marketplace routes, pin the metadata-admin surface that cannot be freed - #6775

Fixes #6681

Four surfaces AppContent declares with lazy() and that the emitted bundle loaded on
every console page load anyway. Each was measured separately, from the emitted
chunk's own module list — never a source-level search, which cannot see the mechanism at
all (objectui#6680 is the recorded reason).

Baseline, re-measured on my own base

The card's pair of numbers is stale on both sides; both were re-read on b98352a15:

Which side moved: both, and by the same commit. faac0d935 (objectui#6735, the
sideEffects array of objectui#6683) lowered MAX_EAGER_CLOSURE_GZIP_BYTES from
3_345_000 to 3_300_000 (−45,000 B) and took more load off than that
(−51.5 KB), which is why headroom widened rather than narrowed. Recorded here for
objectui#6631, which remains open and is not touched by this PR.

Per chunk: the module list, the edge, and whether it was removable

The card measured metadata-admin at 181,228 B and MarketplacePage at 3,129 B eager.
Both moved: metadata-admin is 8,577 B smaller, and MarketplacePage left the eager
set entirely before this branch. That is a finding, not a nuisance — one of the four
was already fixed by objectui#6735, and it is exactly the control that makes the
co-tenancy mechanism legible below rather than a story about two files.

metadata-admin — 172,651 B, NOT removable

views/metadata-admin/index.ts is the target of all six lazy() declarations, and the
emitted graph shows it statically imported by two modules that are eager by
construction:

1. packages/app-shell/src/index.ts — the package barrel, re-exporting eleven runtime
   values from it (registerMetadataPreview, useMetadataClient, …). The console entry
   imports that barrel.
2. packages/app-shell/src/services/builtinComponents.tsx — which the barrel
   BARE-imports (import './services/builtinComponents.js';) and which imports
   MetadataDirectoryPage and MetadataResourceRouter by value, because a registry
   entry naming a component must hold the component.

Decisive for the chunking lever: the six pages the declarations name have
dynamicImporters: [] in the emitted graph. They are reached only through the
barrel's static re-exports, so no advancedChunks policy separates them from it. The
module also performs five load-bearing top-level registrations
(registerBuiltinAnchors, registerDefaultMetadataSchemas, registerDatasourceResource,
registerBuiltinPreviews, registerBuiltinInspectors) — which is why
@object-ui/app-shell's own sideEffects array names it.

Freeing it means changing what registerAppComponent accepts (a component VALUE today)
and what the package barrel re-exports: a published-contract decision, not a bundling
one. It is pinned in DECLARED_LAZY_VIEWS_STILL_EAGER with that reasoning, and the
follow-up is filed separately.

The two marketplace chunks — chunk co-tenancy, removable

Neither page had a static importer. Their chunks did:

• assets/MarketplacePackagePage-*.js also held components/SuggestedBindingsPanel.tsx
  (+ services/suggestedBindingsApi.ts), which the eager
  views/studio-design/StudioDesignSurface.tsx imports statically.
• assets/MarketplaceInstalledPage-*.js also held
  console/marketplace/InstalledListWidget.tsx, bare-imported by the barrel for its SDUI
  registration.

The repair is an app-shell-eager-leaves group in apps/console/vite.config.ts that
isolates those co-tenants, so all three pages chunk by their own dynamic-only
reachability.

The opposite grouping was measured FIRST and was worse. A marketplace-routes group
over the three declared-lazy pages swept the bare-imported widget in with them, so the
barrel's side-effect import made the GROUP eager; it became an attractor for 47 modules
including runtime-config.ts, providers/MetadataProvider.tsx and
@object-ui/plugin-form, emitting a 44 KB eager chunk. Net −1,697 B against −8,888 B,
and three other lazy boundaries destroyed. Both the config comment and the plugin header
record it, because a chunk group decides co-tenancy and co-tenancy runs both ways.

Result

eager closure  3180.2 KB -> 3171.5 KB gzipped   (-8,888 B)
eager chunks     48/513  ->   45/513
headroom         42.5 KB ->    51.1 KB
per-chunk ceilings (vendor-objectstack, framework, ui-components) unmoved

The ledger, widened

scripts/vite-declared-lazy-views.ts deliberately parsed only ../views/NAME.js.
It now parses every relative lazy() specifier — twelve distinct files, up from
eight — so the directory barrel and the marketplace routes are inside the ledger and
cannot drift back. The package specifier (@object-ui/plugin-designer) stays out, with
the reason recorded: it needs a resolver to become a path, and its chunk was measured NOT
eager.

Widening the ledger meant widening what gets declared moduleSideEffects: false, and the
existing purity guard was not enough for that. bareSideEffectImport reads the module
source for a bare import './x.js';; views/metadata-admin/index.ts has no bare
import at all — its five side effects are CALLS — so that guard returns null for the
one file where a false purity claim would drop every built-in preview registration. A
second guard now reads the package's own published sideEffects array and refuses any
module it names. One producer of the fact, two consumers.

Verification

Union re-run on the final commit d853fdedb, tree clean:

GATE eager-closure EXIT=0           GATE changeset-fixed EXIT=0
GATE control-bytes EXIT=0           GATE changeset-no-major EXIT=0
GATE side-effects-array EXIT=0      GATE entry-guard EXIT=0
GATE sdui-registration-pins EXIT=0  GATE esm-specifiers EXIT=0
GATE changeset-presence EXIT=0      GATE phantom-deps EXIT=0
VITEST_SCRIPTS EXIT=0   ESLINT EXIT=0

Verdict lines, quoted from the gates themselves:

• check:eager-closure — Console eager closure is 3171.5 KB gzipped across 45 of 513 chunks (budget: 3222.7 KB, headroom: 51.1 KB).
• check:sdui-registration-pins — All 12 registration(s) a sideEffects array promises are present in the built console (513 chunks weighed…), including marketplace:installed-list 1 chunk(s) — the widget my chunk group moved.
• check:side-effects-array — @object-ui/app-shell: sideEffects names exactly the 10 module(s) that register at load time…
• console build — [plugin declared-lazy-views] 10/12 modules AppContent declares lazy are genuinely lazy; 2 eager, all pinned
• vitest scripts/__tests__/vite-declared-lazy-views.test.ts + check-eager-closure-budget + console-vite-alias-closure-4925 — Test Files 3 passed (3) · Tests 114 passed (114)
• pnpm --filter @object-ui/console type-check — exit 0 (tsc --noEmit && tsc -b tsconfig.node.json --force echoed, so it ran)

Reverse verification (ablation). The new purity guard was proved to fire by deleting
the metadata-admin ledger line and rebuilding. Mutation confirmed on disk before the
run (git hash-object e1971ad3… → ae9bce55…, pin lines 1 → 0); build exited 1 in
buildStart with [declared-lazy-views] refusing to declare … side-effect-free: its own package says otherwise; restored via git checkout HEAD -- and proved restored
(git diff HEAD empty, blob hash back to e1971ad3…). Without that guard the deletion
would have declared the module pure.

Lint was narrowed and the narrowing is measured, not assumed: the diff's lintable
population is the three source files it changes (the fourth changed file is the
changeset, and eslint.config.js configures no markdown linting); --format json
reports files linted: 3, all errorCount 0; and eslint.config.js enables no
type-aware linting (no projectService, no parserOptions.project), so this diff cannot
move the verdict on any file it does not touch. The full farm runs in CI regardless.

Generated by Claude Code