perf(webui): lazily load module descriptors

docs/WEBUI_BUNDLE_BUDGETS.md

@@ -0,0 +1,71 @@
+# WebUI Loading And Bundle Budgets
+
+The Core WebUI host owns the loading boundary for installed module packages.
+Vite discovers configured packages at build time, but emits an asynchronous
+loader for each package's `src/module.ts` contribution descriptor. At runtime,
+Core imports only descriptors whose backend manifests are enabled and identify
+the matching `frontend.package_name`.
+
+The direct descriptor entry is intentional. A package root may re-export pages
+for consumers; importing that barrel as module wiring can cause those pages to
+be evaluated before navigation. Route pages and substantial panels should use
+`React.lazy`, and Core wraps routes in the shared loading/error boundary.
+
+## Enforced Budgets
+
+`webui/bundle-budget.json` contains the production limits:
+
+| Measurement | Raw limit | Gzip limit |
+| --- | ---: | ---: |
+| Initial JavaScript static import closure | 512 KiB | 160 KiB |
+| Largest individual asynchronous JavaScript chunk | 384 KiB | 110 KiB |
+
+`npm run build` writes a Vite manifest, measures the entry and its recursive
+static imports, writes `dist/bundle-metrics.json`, and fails when either budget
+is exceeded. `npm run test:module-permutations` applies the same gate to every
+permutation and records the collected results in
+`dist/module-permutation-bundle-metrics.json`. In CI, each result is also added
+to the step summary.
+
+Budgets are limits, not targets. A change that approaches a limit should add a
+new lazy boundary or remove unnecessary entry code instead of raising the
+limit without measurement and review.
+
+## 2026-07-30 Baseline
+
+Measurements use the same full-product source tree and Node 22 runtime. The
+post-change build additionally includes the Search module in the default and
+full-product sets.
+
+| Initial-load measurement | Before | After | Reduction |
+| --- | ---: | ---: | ---: |
+| JavaScript assets in initial static closure | 1 | 1 | 0% |
+| Raw JavaScript | 1,387,043 B | 453,769 B | 67.3% |
+| Gzip level 9 | 364,767 B | 141,725 B | 61.1% |
+| Brotli quality 11 | 254,797 B | 106,701 B | 58.1% |
+| Parse proxy median | 18.776 ms | 7.750 ms | 58.7% |
+| Parse proxy p95 | 21.980 ms | 8.739 ms | 60.2% |
+
+The parse proxy constructs a fresh `node:vm` `SourceTextModule` from the entry
+source 30 times with a randomized source marker. It is useful for a controlled
+before/after comparison, but is not enforced in CI because absolute timings
+vary across runners. Transfer budgets use deterministic raw and gzip byte
+counts.
+
+The first budgeted full-product build reported:
+
+- initial JavaScript: 453,769 B raw / 141,725 B gzip;
+- largest async chunk: `CampaignWorkspace`, 353,724 B raw / 98,142 B gzip.
+
+## Verification
+
+```bash
+cd /mnt/DATA/git/govoplan-core/webui
+npm run build
+npm run check:bundle-budget
+npm run test:module-permutations
+```
+
+The build gate also catches accidental eager imports: a page pulled into the
+entry closure consumes the initial budget, while an oversized page or module
+descriptor consumes the asynchronous chunk budget.