### Description

Rewrites `internal-docs/lazy-compilation/{design,implementation}.md` to match the current implementation. The docs dated back to the initial POC (split out in #9826) and had drifted from the code in several load-bearing places. Every claim was re-verified against current `main` (file/line spot-checks, cross-checked with the playground specs).

#### Corrections to stale / wrong content

- **Proxy templates**: both inline code blocks are now the real `proxy-module-template{,-fetched}.js` verbatim. The stub evicts its runtime cache entry and ends with `return await loadExports($STABLE_PROXY_MODULE_ID)['rolldown:exports']` — the two-level promise chain from #9981 that makes init errors catchable at the consumer's `await import()`. The fetched template returns `loadExports($STABLE_MODULE_ID)` instead of the import namespace (export-name preservation in shared chunks, #9132).
- **Module IDs**: design.md's TL;DR claimed absolute paths are used "consistently throughout the runtime" — runtime lookups actually use stable (cwd-relative) ids; absolute paths survive only in the `/@vite/lazy?id=` param and the fetched template's `import($MODULE_ID)`. Also documents the four template placeholders and their JSON-quoted rendering (#9102).
- **Error handling**: replaced "`Err` or panic is fine for POC" with the actual contract — unknown ids are rejected as a cache-key-only security gate (#9969), napi surfaces `Failed to compile lazy entry:` → HTTP 500, and init errors are catchable on both cold and warm paths (#9975/#9981).
- **Dedup**: the race-condition section's "potential future enhancement" runtime guard is long implemented — lazy chunks render `createEsm/CjsInitializer(stableId, factory, 1)` with a dedup flag (HMR patches deliberately omit it so updates re-execute). Rewrote the section around the two-layer dedup: per-client `executed_modules` pruning + the runtime flag.
- **Rebuild id normalization**: the old note said the `ModuleChanged` rebuild "should resolve or normalize" the proxy id to a real module id — the raw proxy id is deliberately correct (it is the incremental-cache invalidation key for the proxy whose content changed; normalizing would leave the stale stub cached). Note removed and the actual mechanism documented.
- Misc: `lazyMagic` helper (never existed), `__export` → `__exportAll`, `snippet` → `ast_factory` in the Lessons Learned snippets, `registerModule(stableId, { exports })` shape.

#### New coverage

- **Enabling / wiring**: `experimental.devMode.lazy: true`, inner-plugin registration, the shared `LazyCompilationContext` handed to the `DevEngine`
- **Lazy chunk rendering**: initializer wrappers + `__rolldown_module_id__` param, the nested-lazy `import()` → `/@vite/lazy` rewrite, entry init call, `lazy_compile_{n}.js` naming
- **Emitted assets**: delivered via `onAdditionalAssets` before the code returns (#9815)
- **Build output refresh**: `update_watch_paths()` runs first (what makes post-fetch edits watchable at all), silent output swap for connected clients, failure paths (#9903)
- **Editing a fetched lazy module**: per-client HMR outcomes; a non-accepting proxy escalates to FullReload → HmrRebuild with a deferred reload
- **Client sessions**: implicit creation on `hmr:module-registered`, removal on ws disconnect, the `"rolldown-tests"` escape hatch; clientId prunes the patch — nothing is "routed"
- **Known limitations**: link-stage-synthesized exports (JSON/text/base64/dataurl register `{}` inside lazy chunks until rebuild + refresh), CSS error deferral, assets require load-hook plugins, sourcemaps (inline-only for lazy chunks — the map asset is discarded on the lazy path, unlike `HmrPatch`)
- **Test coverage table**: the 7 playground specs + `dev-lazy-compile.test.ts` and what each pins

Docs only — no runtime change.

🤖 Generated with [Claude Code](https://claude.com/claude-code)