[Sheraff]

Based on the compiler oracle test suite branch (compiler-tests/baseline), which is included in this diff. That branch adds the oracle (packages/compiler-test-utils) and pins main's bugs as .fails; this PR replaces the compilers and turns the pins green.

Reading order: core README (packages/router-utils/src/compiler/README.md) → planners → hosts → tests. Details in How to review.

Summary

Every TanStack build-time compiler now runs on one shared core built on Yuku: route code splitting and route HMR (@tanstack/router-plugin), and all of Start's transforms (server functions, environment separation, <Hydrate>, RSC CSS, import protection). The route generator is rewritten on Yuku too. Babel, babel-dead-code-elimination and the Prettier hot path are gone.

Why:

• Speed. On main, Babel was ~90% of compiler time and 50–60% of all main-thread time in dev and builds. On a 333-route app, Vite and Rsbuild builds are now 2.1–4.6× faster (webpack 1.5×) and dev cold start 1.9–4.0× faster.
• Correctness. Main's compilers re-parsed and re-printed each other's output, recognised APIs by name, and kept order-dependent state across transforms. 29 open issues are fixed, and the compiler oracle's known-bug pins went from 874 to 45. 36 of those 45 keep main's behaviour on purpose: root routes are not split, and route CSS loads with the route module.
• One model. One parse and one plan per module and environment, with every output emitted from the original source and exact source maps. Split chunks, Hydrate chunks and server-fn providers no longer depend on the order the bundler asks for them.

Architecture in brief

Full map: packages/router-utils/src/compiler/README.md.

• One analysis per (env, module, text) → one plan → many outputs: the env module plus derived modules (?tsr-split=…, ?tss-hydrate=…, ?tss-serverfn-split) and facts (import edges, server-fn registry, diagnostics, dependencies). Each job is a planner that adds edits and roots to the shared plan, in a fixed order: start-env → hydrate → rsc-css → route-hmr → route-split → import-edges.
• Rewrite tree over the original UTF-16 ranges (retained slices, structured glue, claims). User code is never re-printed.
• Reachability = splitting = dead-code removal, at cluster granularity. Code is removed only when our edits orphaned it (provenance), so user side effects never disappear.
• Emit through our own OutputBuilder (direct VLQ maps, composed with input maps).
• Recognition by binding with an abstract value domain. Cross-module summaries are built from raw text, never through the bundler's transform pipeline.
• Plan cache + dependency ledger + one invalidate(): validated reuse, per-env resolvers, no Yuku objects retained (heap test).
• Hosts we own: a native Vite plugin, a Rspack/webpack loader, and Start Rsbuild loaders per env/layer with buildInfo replay for the persistent cache. unplugin is used only for the generator and esbuild.

Results

Compiler oracle (packages/compiler-test-utils/oracle, real bundlers/dev servers/browsers, public APIs only)

	baseline (main's compilers, 80d1890bed)	this branch
tests	3,594	4,114
known-bug pins (.fails)	874	45
non-pin failures	0	0

Heads: router-plugin and start-plugin-core (unit + integration, 3,645 tests, all 45 pins) ran at d0cb0e3ff6. router-generator and react-start-rsc (469 tests, 0 pins) ran at c4cb7e889a; since then only router-generator's test harness changed, and its integration tests also pass at d0cb0e3ff6. fdc48254c8 adds 2 passing tests (per-environment release, below) and changes no other result.

The 45 remaining pins:

#	pin	why it remains
20	route CSS applies only once the route matches (shows-component / -error / -not-found / -pending / -nothing; Vite + Rsbuild, dev + build)	Unchanged from main on purpose: bare style imports of a route load eagerly with the route module. Loading route CSS on match needs a router-core per-route asset loader (F2)
16	a root route's errorComponent / notFoundComponent loads lazily (Vite, Rspack, webpack, Rsbuild; with and without sideEffects)	Root routes are never split, as on main. Splitting them cost SSR CPU on 404/error renders, and the manifest preloaded the chunk on every page anyway
3	rspack/webpack/rsbuild without sideEffects: an exported component another route's chunk imports loads with its chunk	webpack-family chunk graph: only a side-effect-free package lets Rspack bypass the eager route module. The fix is a cross-module optimisation, not done
2	Rsbuild dev: a module of a custom type (.page as javascript/auto) importing a denied file/specifier	Rspack rules select by resource, not module type. A catch-all loader would run on every asset
1	Rsbuild build: an unused exported component with a boundary is not in the client	Rspack drops the unused declaration but still emits the async chunk of the import() inside it (unreferenced file)
1	Rsbuild dev: an edit that adds a route's first server fn to its component shows on the open page	Upstream (Rspack's HMR runtime re-executes the updated chunk before the new async chunk it needs is loaded; web-infra-dev/rspack#11324, #11922). A plain Rsbuild app fails the same way
1	Vite dev: a Start library's import of a CommonJS dependency in a client route	excluded libraries' undeclared CJS deps are not pre-bundled. Not supported for now
1	Vite bundledDev: a factory that becomes a server-fn builder makes the client drop the handler	needs Vite full-bundle API to re-transform importers. Open

Unit tiers: router-utils core tests (pipeline, placement, rewrite, output-builder, values, facts, edges, identity, heap) and the planner tests in start-plugin-core.

e2e

CI (PR workflow, Test job = test:pr: affected lint, unit, integration, e2e, types and build targets) at 6b90fe7291: 885/886 Nx tasks pass. The one failure, vue-start dev-ssr-styles (dev), was Vite's dependency-optimizer reload navigating the page away mid-test. It is fixed since (Vue Start cold-start fix, below), and the test passes locally.

Compiler Platforms (Linux x64 musl, Linux arm64, macOS arm64, Windows x64) is green at 6b90fe7291 and d0cb0e3ff6. CodSpeed at d0cb0e3ff6: the SSR assets regressions are gone (route CSS loads eagerly, as on main). The remaining deltas are small memory benchmarks (±1–2.5 KB).

Performance (c4cb7e889a vs main da2adec76d; harness benchmarks/compiler-perf, 15 cells, 1 warmup + 5 runs, interleaved cell by cell)

Median ms (memory: MB), main → this PR. Medium = 333 route files, large = 1,485.

cell	build	dev listen	dev cold → hydrated	compile-all	dev heap after GC	build peak RSS
react-start Vite medium	6,196 → 1,634	627 → 212	6,146 → 2,561	7,176 → 2,038	212 → 87	1,382 → 1,238
react-start Vite large	29,007 → 8,530	1,447 → 306	23,679 → 7,592	31,251 → 9,665	704 → 200	2,901 → 1,856
react-start Rsbuild medium	6,566 → 1,949	220 → 168	7,930 → 2,279	–	218 → 106	1,271 → 1,166
react-start Rsbuild large	26,450 → 6,301	219 → 170	31,932 → 6,983	–	733 → 284	3,083 → 2,486
react-start-rsc Vite medium	8,899 → 2,447	703 → 314	7,269 → 3,788	7,931 → 3,282	299 → 145	1,781 → 1,470
react-router Vite medium	2,802 → 612	576 → 184	3,254 → 1,046	6,476 → 1,303	64 → 47	872 → 689
react-router Vite large	11,185 → 3,109	1,457 → 273	12,469 → 2,909	28,303 → 5,168	130 → 88	1,428 → 1,214
react-router Rsbuild medium	2,728 → 603	208 → 145	2,892 → 724	–	62 → 47	832 → 680
react-router webpack medium	6,633 → 4,410	342 → 277	3,746 → 1,513	–	162 → 141	3,146 → 3,193
solid-start Vite medium	9,185 → 4,375	641 → 266	7,562 → 3,406	10,669 → 4,624	191 → 90	1,563 → 1,349
solid-router Vite medium	4,202 → 1,982	601 → 248	3,886 → 1,490	9,998 → 4,049	63 → 54	947 → 855
vue-start Vite medium	8,347 → 3,780	654 → 277	10,462 → 3,312 ¹	9,217 → 4,053	178 → 102	1,283 → 1,297
vue-router Vite medium	3,966 → 1,624	608 → 246	3,750 → 1,334	8,854 → 3,054	67 → 57	940 → 836

Dev heap after GC is measured once / has hydrated. Rsbuild and webpack open their port before compiling, so their listen times are not comparable with Vite's. Across all 15 cells: builds −34…−78%, listen −19…−81%, cold start −15…−78%, compile-all −46…−82%, heap after GC lower on every cell (−3…−82%). Peak RSS is lower on every cell but three: React Start small dev cold +4.6%, and Vue Start build +1.1% and webpack build +1.5% (both noise).

¹ After the Vue Start cold-start fix (c271282ac5), measured separately with 14 cold starts per side. At c4cb7e889a the page stalled after Vite's dependency-optimizer reload in 12 of 14 cold starts (median 19,424 ms). The cause was the pre-bundled @tanstack/vue-router/hmr subpath: requested during the re-optimization, it got a 504 "Outdated Optimize Dep" and the entry never ran. With the fix: 0/14 stalls and 0/14 reloads. Main reloads the page in 14/14 cold starts.

HMR (save → DOM, p50): route component/loader edits −17…−66% on Vite. Server-fn edits are −62…−84% on Vite (RSC −28%, Rsbuild −31/−40%). Solid Router and Vue Router route edits no longer full-reload: main does 20/20 full reloads at 254–264 ms, this PR does hot updates at 14–16 ms. Rsbuild HMR is noisy on both sides (p90 up to 2× p50): the <Hydrate> edit reads +97% / +21% (medium/large) and the large route-component edit +6%. Shared-module edits (a component one route imports) are at or below main in every cell and don't grow with app size: large React Start −2.40 ms [95% CI −2.70, −2.10], Solid Start −2.15 [−2.40, −1.90], React Router −0.10 [−0.40, 0.20]; React Start medians stay at 23.2/23.6/23.5 ms small→large while main grows 23.4/24.1/25.6.

Leak check (100 edit + revert cycles, large apps): no growth. Heap slope after cycle 20 is −10.8 / +0.4 MB per 100 cycles for Start / Router.

The generator no longer runs Prettier. At 498ac9be17, the react-start build floor (every compiler output replayed from a recording) dropped from 1,629 to 779 ms, and the generator from 715 to 94 ms.

Memory soak (500 edit + revert cycles on one dev server, at 18ef570edc)

The edits rotate over: route component, loader, shared module, server-fn handler, <Hydrate> child, module add → delete, route add → delete, route rename, a *.server.ts import added and removed, and a factory that stops or starts exporting createServerFn. Heap after GC is sampled every 25 cycles, with heap snapshots at cycles 100 and 500. Slope over cycles 125–500, KB per cycle [95% CI]:

run	heap after GC	main (reference)
react-start Vite	−0.7 [−14.5, 13.1]	11.2 [−11.7, 34.1]
react-router Vite	2.3 [1.8, 2.8]	6.3 [3.7, 9.0]
react-start Rsbuild	20.9 [2.6, 39.3], plus 262 of ArrayBuffers	–

The soak found five records that grew with deleted or renamed files: plan-cache generations, dependents-index buckets, raw facts, import-protection records, and Rsbuild per-resource records. All five are fixed and tested. After the fixes, no data object of our packages grows in count between snapshots, and every compiler cache stays bounded by the live files. Replaying our recorded transform outputs (no compiler work) gives the same slope within the CI, so our transforms add no measurable growth. Two caches the soak showed to be useless were removed afterwards: per-env resolutions, and the emitted-output memo (1 hit in 4,959 lookups for 12.9 MB).

The remaining growth comes from upstream. Vite 8.3.4 keeps the module-graph nodes of deleted files (≈2 KB each; most of the Router slope). Rsbuild's dev server keeps every hot-update asset in its in-memory output file system (≈260 KB per edit).

Memory issues (#8614, #8794)

• Import protection in dev (Start: import protection keeps a parsed AST per module in dev (hundreds of MB in large apps) #8614), the reporter's repro with 1,500 modules, heap after GC with import protection on minus off: main +321 MB, this PR +2.4 MB. Main kept a Babel AST per module; no AST is retained here.
• Production builds ([Start] vite build keeps every environment's server-fn compiler alive until the build ends (memory stacks across client → ssr → server) #8794), a synthetic Start + Nitro app (3,000 modules, 600 server-fn routes), heap after GC when the Nitro environment starts: main 506.5 MB, main with the issue's buildEnd fix 98.7 MB, this PR 64.3 MB (fdc48254c8). When an environment's production build ends, its plans and import-protection state are freed; dev and watch are unchanged.

Bytes

Client (pnpm benchmark:bundle-size:run, 18 scenarios, at d0cb0e3ff6 vs main da2adec76d; byte-identical at fdc48254c8):

scenario	gzip Δ	note
12 of 18 (all Router scenarios, Start minimal / query / deferred-hydration)	0	byte-identical
react-start.full / solid-start.full / vue-start.full	−47 / −38 / −36	shorter server-fn ids (22-char base64url)
react-start.rsbuild.full	−35	same
react-start.rsbuild.minimal / minimal-iife	+1 / +3 (raw −22)	accepted exception, see below

The Rsbuild minimal gzip increase is the only client gzip regression. Our loader request changed, so Rspack renumbers its numeric module ids; raw size is −22 B and our emitted code is not larger. The maintainer explicitly accepted it as a one-off.

react-start.full brotli is +92 B, while raw (−42) and gzip (−47) are lower. This is brotli q11 noise: random 22-char ids span 174 B of brotli size. Accepted.

Server (scripts/benchmarks/bundle-size/server.mjs, checkout paths normalized, at d0cb0e3ff6 vs main; identical at fdc48254c8):

• Every Vite scenario: −12…−823 B raw.
• react-start.rsbuild.minimal: +7 raw / −17 gzip.
• react-start.rsbuild.full: +432 raw / +132 gzip, accepted exception. Per D6, the server-fn handler stays in its module (no double init; main ran server-fn modules twice). So the resolver requires __root.tsx, which becomes a second concatenation root. Rspack then cannot scope-hoist the router modules __root shares with the route tree (+1,042 B of module wrappers), partly offset by removing main's duplicated provider code (−585 B). A cache-group fix is blocked by an Rspack panic and is a follow-up.

Cold start (spawn → first SSR response, measured at c4cb7e889a; server output is identical at d0cb0e3ff6) is within noise (−2.1…+3.5%). The *.full Vite servers load one more chunk (the route module shared with the resolver), per D6.

Behaviour changes / migration notes

• routeTree.gen.ts formatting changes once. The generator prints the tree directly in your quoteStyle/semicolons. enableRouteTreeFormatting now defaults to false; set it to true to keep Prettier formatting (Prettier is then loaded lazily). Custom scaffolding templates are inserted as written.
• Server functions only server code calls are no longer HTTP-callable in production builds (also those only returned to the client as data). The dev server answers them and warns once per function.
• serverFns.generateFunctionId receives the stable key (greetUser) as functionName, not greetUser_createServerFn_handler. Default production ids are shorter.
• esbuild + autoCodeSplitting: true throws; the generator still works with esbuild. Docs and the esbuild quickstart examples are updated.
• router-plugin requires Vite ≥ 6 (vite peer >=6). CommonJS consumers (a webpack.config.js / rspack.config.js that require()s the plugin) need Node 20.19+ or 22.12+.
• Removed experimental, Babel-typed hooks: codeSplittingOptions.compilerPlugins; Start compilerTransforms, serverFnProviderModuleDirectives, StartCompilerPlugin and related types; @tanstack/react-start-rsc/plugin/rscCssTransform. RSC integrations pass rsc: { loadCssExpression, providerDirective }.
• router-utils no longer exports the Babel helpers (parseAst, generateFromAst, deadCodeElimination, …). It gains @tanstack/router-utils/compiler. router-generator no longer exports format, inferFullPath, checkRouteFullPathUniqueness or writeIfDifferent, and GeneratorPlugin.onRouteTreeChanged receives { model }.
• New dev-only subpaths @tanstack/router-core/hmr and @tanstack/{react,solid,vue}-router/hmr (the route HMR runtime; nothing reaches production).
• Route CSS is unchanged from main: bare style imports of a route load eagerly with the route module; CSS modules follow the code that references them.
• Root routes are never split (as on main): their component, errorComponent and notFoundComponent stay in the root route module.
• Route callbacks reading this (methods/function expressions) stay unsplit with a warning (Handle route callbacks that reference this during code splitting #8688).
• Import protection: per-edge mocks exporting the target's names, build verdicts after tree-shaking, and onViolation returning false gets the real module (fixed-point rebuild). Traces follow the bundler graph, and reports point at the import.
• Rsbuild: server-only routes are pruned from the client tree, the client manifest lists only a route's chunks, the persistent cache is supported, spa.enabled dev serves the shell, server.compress is no longer forced, and the server env shares async chunks.
• Solid Start server envs split route components like the client (hydration keys agree).
• Vue Start dev: a cold start no longer reloads the page once after dependency optimization (@tanstack/vue-router > @tanstack/vue-store is pre-bundled, as React Start does for its store).

Changesets: compiler-core-yuku (router-utils minor), route-generator-yuku (router-generator minor, router-cli patch), router-plugin-yuku (router-plugin minor), route-hmr-runtime (router-core + react/solid/vue-router patch), start-compiler-yuku (start-plugin-core + router-core minor; start-client-core, react/solid/vue-start patch), react-start-rsc-compiler (react-start-rsc minor), dev-cold-start-and-invalidate-index (start-plugin-core, vue-start, router-utils patch), router-plugin-library-build-externals and router-plugin-order-check-splitting-only (router-plugin patch), start-plugin-core-file-url-sources (start-plugin-core patch), release-finished-environment-plans (start-plugin-core, router-plugin, router-utils patch).

Follow-ups (out of scope)

• F1 Server functions sent to the client as data: today rejected in production, and their client middleware never ran (on main either). Idea: an explicit createServerFn({ serializable: true }).
• F2 Fully lazy route CSS (Unnecessary CSS being imported #4396): a router-core per-route asset loader awaited in loadRouteChunk. Route CSS placement is unchanged from main; the 20 route-CSS pins wait on this.
• F3 Prettier packaging: whether it becomes an optional peer of router-generator.
• F4 Start: the Vite plugin and tsr generate write different routeTree.gen.ts (Start footer added, then stripped) #8437 CLI half (tsr generate writing Start's footer); server routes with a .ts extension gets the wrong boilerplate while creating the file. #4462 (.ts server-route scaffolds); Vite HMR blank page when editing __root.tsx (Solid SSR) #6331 (Solid SSR root HMR: solid-refresh bug, reported upstream as solid-refresh#91, nothing in this repo); [@tanstack/react-start/plugin/rsbuild] tanstackStart() eclipses the client environment's output.assetPrefix, blocking downstream plugins #7827 (Rsbuild final-config ordering).
• F5 ssr: false routes still bundle components in server output #6400 (ssr: false route code pruned from the server: needs inherited-ssr knowledge, design in notes); with RSC, a server-fn module's state exists once for server fns and once for server components (allowed; docs note).
• Ports after merge: perf(start-plugin-core): avoid repeated work in Rsbuild compilation scans #8799, perf(start-plugin-core): reuse the Rsbuild dev client build when a rebuild cannot change it #8800 (Rsbuild perf), fix(start): prune statically disabled SSR route implementations #8291, [feat]: Add stable IDs for server functions #7693, feat(router-plugin): add native Bun bundler integration #8804 and the rest of the "implementation obsolete" list below.

Upstream issues filed

✅ = fixed upstream.

• Yuku (fixed): ✅ #202 purity comments with leading whitespace, ✅ #203 codegen SIGBUS on invalid object method, ✅ #204 JSX references for _/$/Unicode names, ✅ #229 import X = ns.Y reference space, ✅ #230 decorators printed after export, ✅ #231 JSXText.value entity decoding, ✅ #232 false duplicate export (type + value), ✅ #237 comments in empty JSX containers, ✅ #244 WASM stale bytes for empty source (WebContainers), ✅ #247 unwrap typing, ✅ #248 module.copyAst(), ✅ #251 export {} from "m" module record, ✅ #252 parenthesized/template import() specifiers, ✅ #253 big-endian AST decoding, ✅ #254 JSX attribute value mapping, ✅ #255 postfix ++ parentheses before a line comment, ✅ #256 for (init;;) no-in leak, ✅ #257 private method overloads, ✅ #258 import type {} from 'x', ✅ #259 super() in a nested computed key, ✅ #260 JSX text after </>, ✅ #287 (a?.b)!.c printing
• Yuku (open): #288 JSXText escaping, #289 stale JSXText raw, #290 strip keeps type-only exports, #291 diagnostic span splits a surrogate pair, #292 Annex B block functions
• Vite: #23719 saves within 50 ms ignored (chokidar throttle), #23907 module graph keeps deleted files
• @vitejs/plugin-vue-jsx: vite-plugin-vue#866 removed export breaks HMR accept
• Rolldown: #11054 this.load() returns null during bundled-dev HMR, #11223 watch mode keeps stale moduleSideEffects: false, #11300 bundled-dev MissingFactoryError (this PR carries main's importer fallback)
• Rspack: #16288 HMR runs a module before its new async chunk loads, #16289 splitChunks panic "Chunk not found in ChunkByUkey"
• Rsbuild: #8736 dev output FS keeps every hot-update file
• solid-refresh: #88 rename invalidate loop on webpack/Rspack, #89 sticky first registry (Vite), #90 importer stale after rename, #91 document component remount (TanStack Vite HMR blank page when editing __root.tsx (Solid SSR) #6331)
• magic-string: #368 maps count only \n as a line break
• babel-dead-code-elimination: #50 a typeof x type query keeps x alive
• WebContainer: webcontainer-core#2173 TextEncoder.encodeInto('') returns the previous result
• tanstack.com: ✅ #1353 example playground drops Start's compat config

Issues

• Closes Server functions that are stored in a separate file cause middlewares to throw error #7213
• Closes "Server function info not found" when calling a server function that's *only* called on the server #8731
• Closes Start RSC mode: server functions imported only from 'use client' modules are missing from the production manifest ("Server function info not found") #7943
• Closes Start + Rsbuild: server-only routes are not pruned from the client route tree #8545
• Closes Start: symlinked route files skip the client-side server.handlers strip — server code and node-only imports ship in the client bundle #8514
• Closes Empty .ts route file is scaffolded with JSX, breaking all route generation #8480
• Closes Code-splitter breaks every route when the project path contains a single quote/apostrophe (autoCodeSplitting: false does not avoid it) #7754
• Closes router-plugin: object-form loader breaks the code splitter when loader is split #8753
• Closes Code splitter does not support the method shorthand syntax #8623
• Closes router-plugin: split chunk and shared module miss shared bindings when compiled before the route file #8754
• Closes router-plugin: a split route option that is also a shared binding exports an undeclared Split… when two @babel/traverse copies are installed #8715
• Closes Compiler correctness: split-route exports and distinct re-export aliases #8505
• Closes If you pass createServerOnlyFn(...) directly as an argument to createServerFn().handler(), it will generate a broken server function. #8446
• Closes start-plugin-core: the dev server-fn id validator's recovery can never succeed — it transforms under the one flag that skips the compile #8447
• Closes start-compiler-plugin fails to lazy-load dependencies with Vite bundledDev #8407
• Closes start-compiler-plugin: hotUpdate crashes with TypeError under vite experimental.bundledDev #8062
• Closes Start: Ssr doesn't reflect changes with bundledDev: true #7535
• Closes Server-fn transform filter regex rejects whitespace between . and handler/server/client, silently disabling transform after AST-roundtripping pre-plugins #7280
• Closes Bug: Breakpoints fail to attach in VSCode when autoCodeSplitting is enabled #6667
• Closes autoCodeSplitting on Rsbuild breaks source maps #3828
• Closes [Vite plugin][Module Federation] Multiple routers in the same window cause an Invalid Hook Call when the remote uses @tanstack/router-plugin — regression in 1.168.17 #7921
• Closes Route generation does not allow plain template literals anymore after 1.166.31 #7216
• Closes Start: user router.plugins never reach the route generator (start-plugin-core overwrites them) #7768
• Closes Start + Rsbuild: SPA mode dev server SSRs routes and runs loaders on the server #8543
• Closes Start + Rsbuild: plugin forces server.compress: false, overriding user config #8544
• Closes Handle route callbacks that reference this during code splitting #8688
• Closes Start: import protection keeps a parsed AST per module in dev (hundreds of MB in large apps) #8614
• Closes [Start] vite build keeps every environment's server-fn compiler alive until the build ends (memory stacks across client → ssr → server) #8794
• Closes Crash "TypeError: serverFn is not a function" when calling server function with nested server function multiple times #6451

Partially addresses #8437 (Start's footer follows quoteStyle/semicolons; tsr generate still writes none).

#6451 was a Vite module-runner race (fixed in Vite 8.2.2, vitejs/vite#23009) that main's compiler triggered by dynamically importing a separate copy of the callee module. Main can still hit it on Vite 8.0–8.2.1, which its vite >=8.0.0 peer allows; this PR compiles server fns in place, so it is avoided on any Vite.

Not fixed by this PR: #5721 is already fixed on main (a Rollup warning; Rolldown/Vite 8 emits none). #8139 is a Rolldown chunk cycle (the __exportAll helper is placed in a dynamically imported chunk); workaround: strictExecutionOrder: true on the ssr environment.

Superseded PRs

Supersedes #8805, #8504, #8630, #8721, #8780, #8773, #8756, #8755, #8732, #8686, #8682, #8681, #8061, #8541, #8493, #8481, #8459, #8455, #7892, #8515, #7281, #8269, #8628, #8797, #8662, #8296, #8215, #8454, #8276, #8675, #8540, #8523, #8524, #8680, #8684, #8642, #8749, #8801, #7338, #8656, #8798, #8818

Implementation obsolete, change not delivered here (needs a re-port or rebase after this lands): #8804, #7693, #7766, #7872, #8291, #8293, #8277, #8799, #8800, #6748, #8256, #8057, #7925

How to review

The diff is large: +159k/−75k lines in 1,975 files, of which the compiler-tests/baseline branch is +102k/−42k in 1,468 files. Most of it is tests (+103k/−41k in packages/*/tests, +16k in the new compiler-test-utils) and regenerated route trees. Product code is +39k/−22k in 231 files. Suggested order:

1. Core: packages/router-utils/src/compiler/README.md, then identity → source/analysis → units → rewrite → reach → plan → emit/output-builder → values/summary/resolve/facts → cache/host/pipeline. Read alongside tests/helpers.ts (toy planners, memory host) and tests/placement.test.ts.
2. Planners:
   • Router: router-plugin/ARCHITECTURE-CODE-SPLITTING.md, src/planners/ (route → decisions → split → hmr).
   • Start: start-plugin-core/src/planners/start-env.ts + src/compiler/ (recognition) + src/server-fns/, planners/hydrate.ts, planners/rsc-css.ts, planners/edges.ts + src/import-protection/.
3. Hosts: router-plugin/src/hosts/; Start src/vite/compiler-plugin.ts (+ planning.ts, bundled-dev.ts) and src/rsbuild/ (COMPILER_ARCHITECTURE.md first).
4. Generator and runtime: router-generator/AGENTS.md (layout), analyze/route-file.ts, model/, print/, generator.ts; router-core/src/hmr.ts, {react,solid,vue}-router/src/{hmr.ts,lazyRouteComponent.*}, start-client-core/src/createServerFn.ts (client middleware meta), react-start/src/plugin/*, react-start-rsc.
5. Tests: the oracle contract packages/compiler-test-utils/oracle/README.md, SPEC.md, KNOWN-BUGS.md (what main got wrong, by root cause), then the unit tiers.

How to reproduce the measurements

CI=1 pnpm install --frozen-lockfile
# oracle (normal, then pins as ordinary tests): see packages/compiler-test-utils/oracle/README.md
COMPILER_TESTS_PINS=run CI=1 NX_DAEMON=false pnpm nx run <project>:test:integration --outputStyle=stream --skipRemoteCache --skip-nx-cache
# compiler perf (same Node for both checkouts; candidate and main run back to back)
node benchmarks/compiler-perf/scripts/run.mjs --label <name> --out <dir> --runs 7 --warmup 1 \
  --variants react-start,react-router,solid-start,solid-router \
  --scenarios build,dev-compile,dev-cold,rsbuild-build,hmr,floor,memory,leak --floor build,dev-compile,dev-cold,rsbuild-build
node benchmarks/compiler-perf/scripts/compare.mjs <main>/results.json <candidate>/results.json --md compare.md
# memory soak (500 edit + revert cycles; add --soak-bundlers rsbuild for Rsbuild)
node benchmarks/compiler-perf/scripts/run.mjs --label soak --out <dir> --variants react-start,react-router --scenarios soak \
  --soak-cycles 500 --soak-every 25 --soak-snapshots 0,100,500
node benchmarks/compiler-perf/scripts/soak-analyze.mjs <dir>/soak-react-*.json
# client bytes (18 scenarios) and server bytes / cold start
pnpm benchmark:bundle-size:run --name <name> --baseline <main current.json>
node scripts/benchmarks/bundle-size/server.mjs --repo-root <checkout> --dist-dir <dir> --build --runs 30 --out <json> [--baseline <json>]

The 15-cell matrix above runs the same scenarios per cell (app size × bundler × framework), alternating the two checkouts cell by cell. Its driver is not part of this PR. Details and caveats (load control, floor method, the duplicate-Vite trap): benchmarks/compiler-perf/README.md.

---

This PR was written with AI assistance (Claude Code agents working in lanes under a human architect). Each lane's work was checked against the oracle, e2e, perf and bytes gates above.

🤖 Generated with Claude Code