--- title: "Day five: nothing in Ursprung transpiles" date: 2026-08-05 summary: >- Asking who the transpile was actually for produced no answer, so the transpiler left and the three oxc wasm packages went with it. What reads the module graph is now a single 27-kilobyte lexer, Tailwind became a command the producer runs, and five tickets were ruled out of scope in one day. It is the first time the map has been redrawn rather than extended. --- # Day five: nothing in Ursprung transpiles Day four ended with the build living inside a Worker, three oxc wasm packages in the isolate, a 128 MiB memory ceiling bisected against production, and a paid-plan-only clause. Day five deleted all of it — not by finding a better bundler, but by asking who the transpile was for and getting no answer. **Ursprung is no longer the TypeScript→JavaScript transpiler that builds it.** That is a change to the Destination rather than an amendment to an axiom, and it is the first time this map has been redrawn rather than extended. Six tickets resolved, eight opened, **five ruled out of scope** — a disposition that had never been used before today and was used five times. Five throwaway spikes, no research surveys: the day ran entirely on grilling sessions and code. The ticket count moved from twenty-eight with fourteen resolved to **thirty-six with twenty resolved**. The map itself grew from 851 lines to 1,471. And for the first time in three days the repository changed as much as the plan did: 53 files outside the planning directory, 1,132 lines added and 454 removed. The Worker now compiles with TypeScript 7 and uploads unbundled, `ursprung` ships the compiler settings every app extends, and the project has its first ADR. This post is mostly a ledger of reversals, because that is mostly what the day was. ## The question that dissolved Ticket 34 asked whether TypeScript 7 should own emit. The spike had already run and the numbers were good: `allowImportingTsExtensions` plus `rewriteRelativeImportExtensions` without `noEmit` is legal on 7.0.2, the output matched ticket 11's hand-written specification on **all eight assertions** — automatic JSX runtime as real ESM with no `require()` trap, type-only imports elided, one source map per module — and it ran in **169 ms emit-only, 450 ms emit-plus-check** over 201 modules against ticket 26's 1,044 ms cold for the oxc chain. `--noCheck` even preserved ticket 11's central result, keeping the build a pure transform that never consults types. **None of that decided it.** TypeScript 7 ships 21 platform-native Go executables, has no wasm target, and its JS API is an IPC proxy to a spawned process — so _tsc emits_ and _the build runs inside workerd_ are mutually exclusive. The ticket could not be answered on emit evidence at all. It had to go at the axiom. And the axiom turned out to have no recorded justification. Both previous forms of it argued about _how_ to guarantee "the build works in a Worker". **Neither recorded who wanted the property.** > Nobody wants a build inside a Worker. They want the **app** inside a Worker. Ad hoc Ursprung applications, written by agents, running in a **dynamic worker**, spun up with no deploy cycle. That is what the build-host requirement was for the whole time. And an agent producing such an app either writes JavaScript directly or runs `tsc` in the sandbox it already has — so **the modules arriving are already JavaScript, and the transpile has no caller.** The axiom **stands**, in an operative form that is a capability bar rather than a location: _whatever is left as tasks for the build needs to be able to run inside the worker._ A sandbox may run the residue too, so ticket 24's one-code-path property is kept rather than traded. What changed is how much has to clear the bar. **Ticket 24 had the distinction and drew the wrong half of it.** It wrote the Worker Loader _out_ of the surface — "built in a Worker does not imply served through the Worker Loader" — while keeping the half that followed from it. The Loader **is** the surface. That clause is reversed, and it is the largest single addition the public-API patch has taken. ## What day four bought and day five returned The oxc apparatus was the map's most expensive single arrangement, and almost everything costly on the map descended from it rather than from anything Ursprung wanted. It left in one patch: - **Three packages, 2,343 KiB gzipped** — `oxc-parser`, `oxc-transform`, `oxc-minify` — gone. With no TypeScript, no JSX and no splicing, what the walk needs from a module is a **specifier list**, which is a lexer's output rather than a parser's. - The **forked wasm loaders**, which ticket 26 had proved were the difference between working and not working on production. - **wasm only** and **sync throughout**, both dead rather than amended — the first because there may be no wasm, the second because its ground was oxc's `wasi.thread-spawn`. - The **128 MiB single-shared-memory ceiling**, the **183.87 MiB floor**, _"exactly three memories fit and a fourth kills the isolate"_, and the `SharedArrayBuffer` dependency. - **Paid-plan-only**, of which more below. `node:crypto`'s `createHash` survives as the emit hash. _TypeScript 7 can never run inside the build_ survives too and **stops mattering**, because tsc is not in the build any more; it is in the agent's sandbox, outside Ursprung's scope entirely. Ticket 26 — the production deploy probe that was day four's sharpest empirical work, run without a Cloudflare account through `wrangler deploy --temporary` and a probe that _threw_ its results so wrangler would surface them — **becomes history rather than a live constraint.** The map kept its paragraphs for the reasoning and for what they proved about workerd, marked as no longer binding. That is the right disposition and it costs nothing; deleting them would have made the same mistake look fresh in six months. Three tickets went out with the apparatus, each closed in its own body rather than left to inference: **source-map composition** (no chain to compose), **oxc wasm distribution** (no wasm to distribute), and **a non-threads oxc build** (no floor to relieve). The replacement evidence is weaker than ticket 26's, and the map says so rather than pretending otherwise. It does not need to be stronger: **both local/production divergences this map was burned by were wasm memory facts, and there is no wasm.** ## One package, and it is not really asm.js Ticket 36 asked what reads the module graph now that nothing transpiles. The answer is **`es-module-lexer` 2.3.1, its `./js` subpath, 27,504 B, zero dependencies** — against the _"three packages, all oxc, 2,343 KiB gzipped"_ the entry started as. **The build isolate holds one package.** The candidate to beat was confirmed rather than beaten, and the interesting part is what the `./js` entry actually is: **V8 refuses to validate it as asm.js** (`Invalid asm.js: Invalid return type`) and runs it as ordinary JavaScript. So adopting it bets on **no engine feature whatever** — not wasm, not `SharedArrayBuffer`, not asm.js AOT. There is no `init` and no async entry point; `parse()` is synchronous by construction. Twenty of twenty correctness cases passed inside a real workerd isolate on the first run, malformed input is refused with a `line:column`, an analysable `import("./page.js")` is distinguishable from an unanalysable one, and it produced **zero disagreements with the wasm entry across 201 modules**. The whole residue — `lex → walk → refuse → hash → manifest` — runs in **28 ms cold, 8 ms warm** for 201 modules, against ticket 26's 1,044 ms. **workerd costs nothing over Node**, where wasm cost 5.35×, and the pure-JS entry's ~1.5× slowdown against the wasm entry is the entire price. Pin it exactly, for the same reason as `capnweb`: `./js` is a subpath nobody upstream has promised to keep engine-independent. ## Tailwind leaves the framework This is the largest change the styling axiom has taken, and it retires three days of work that was correct and is now irrelevant. **Ursprung no longer ensures Tailwind runs. It does not run Tailwind at all.** Tailwind is a **CLI in the producer's hands**, alongside `tsc` — the same move ticket 34 made for transpilation, for the same reason: an agent producing an app already has a sandbox, and one more command in it is cheaper than a library in the isolate. Day three proved Tailwind runs unmodified inside workerd with no `nodejs_compat` and no wasm. Day four priced it at 79.9 KiB gzipped and 8 ms, found `build()` monotonic, measured a **cross-request CSS leak** from a module-scope compiler, refused the CSS minifier on 519 B of savings, and established that `compile()` is **undocumented and has broken in every v4 minor**. All of it real, all of it now moot. What survives is the axiom's second half, and it is sharper for the move: **the compiled stylesheet is an _input_ to Ursprung's build, never a product of it.** It arrives in the same file map as the modules, is treated as **opaque bytes**, is content-hashed into the output store, recorded in the manifest, and linked from the document head by the renderer. **Ursprung never parses it** — which is what keeps the no-CSS-pipeline exclusion intact by construction rather than by care. Three terms ride with it: the stylesheet is **declared, not discovered**; the **producer** runs the CLI, always; and there is **one stylesheet for the app**, not one per route. Everything ticket 13 found living in Tailwind's bundler _plugins_ rather than its engine — content detection, `@source` resolution, URL rewriting, minification, dependency tracking — is exactly what `@tailwindcss/cli` provides. That paragraph was written as a cost of the delegation; it is now the argument **for** the move. The undocumented-API exposure goes with it, because the CLI is one of the four documented routes ticket 13 itself named: an unversioned private API traded for a published contract, and _"budget real work at every minor"_ softens to ordinary version pinning. Four things follow. **Candidate extraction is out of scope** — and it was priced rather than waved away: a scanner over raw module text cost **25 ms for 201 modules, five times the lex**, and silently lost a real class string to a regex-versus-division mistake on the way. **The build becomes synchronous again**, because `compile()` was the one irreducibly async stage. **An app installs a second dependency**, so the honest form of "one dependency" is now _"Ursprung is one dependency; the toolchain that produces the JavaScript and the CSS is the producer's."_ And **staleness between the JavaScript and the CSS becomes a silent failure mode** that ticket 14 now owns. It does not violate ticket 16's _one pipeline, zero permitted differences_: that rule is about the **build**, and the build is identical either way — it receives a file it does not interpret. Only the producer differs, which was already true of `tsc`. ## The reversal ledger The three above were structural. These are the rest, and several of them reverse decisions taken within the last forty-eight hours. **The free-plan exclusion is reversed.** It rested on _"the free tier allows 10 ms CPU per request and a single module transform costs several times that"_ — and there is no transform. The whole residue is 28 ms cold for 201 modules; at **startup** that is charged against the 1 s gate ticket 4 measured, two orders of magnitude of headroom. Kept on the map as a **struck entry** rather than deleted, because the arithmetic is worth not rediscovering — and because which budget a startup build is actually charged against on production is **unmeasured**. **The route table is executed, reversing ticket 7.** Evaluating the module gives each URL, and `Function.prototype.toString()` plus a lex of the thunk gives its page, exactly, per entry, **with no page module loaded**. That rescues ticket 23's two endangered bullets and makes _configuration as an executed module_ a mechanism available generally — which the public-API patch had been owing a shape for. **A client module's content hash is content-local, reversing ticket 8's 2026-08-04 amendment.** No splicing had already removed the dependency cascade; ticket 8 kept the hash dependency-inclusive anyway, routing the dependency through the import map's entries. **That was a choice presented as a necessity.** Reversing it makes the output directory a content-addressed store, makes an incremental cache worth having, and improves browser caching — a leaf edit now invalidates one entry instead of an ancestor chain. It costs ticket 8 its bottom-up emit ordering and leaves **the client-graph import-cycle refusal with no stated ground**. The refusal is probably worth keeping; it now needs a reason written down rather than inherited. **Minification of the app's JavaScript is out of scope — refused on the argument, not the arithmetic.** The arithmetic would not have supported a refusal: minification saves **25.8% gzipped** on real unminified published ESM, against ticket 13's 519 B for the CSS minifier. It is out because it is the **producer's** business, exactly as transpilation is. Ursprung has no more standing to minify the JavaScript it is handed than to transpile it. **No splicing's residue is struck, and the axiom is now literally true.** Day four recorded, honestly, that a bounded check on handler bodies survived and that claiming "no language analysis" would be overclaiming. It turns out **there was never anything to check**: handler identity is `module URL + export name`, so only a module-scope exported function can _be_ a handler, and such a function has nothing in scope to capture but its parameters, module-scope bindings and globals. That is the rule, by construction. A closure returned from a factory has no export name, therefore no identity, and fails where ticket 9's serializer is asked to write a reference it cannot form. **The build performs no language analysis whatever.** **Ticket 31 is out of scope on measurement.** It existed because oxc's wasm linear memory grew ~90 KiB per module and never shrank, giving roughly 600 modules per isolate. The chosen reader's buffer is a **high-water mark sized by the largest single module** — `nextPow2(max(2·chars + 256 KiB, 1 MiB))`, unchanged at 1 MiB across 1,206 lexes. No accumulation, nothing to design around. **capnweb's serializer was re-opened and refused a second time.** Ticket 21 had rejected it on grounds the closed value domain removed, so ticket 9 re-opened it honestly — and then refused it on arithmetic: **14,418 B gzipped against a 221 B hand-rolled reviver**, landing on _every_ page including static ones, since capnweb otherwise reaches a page only when that page does RPC. **capnweb is adopted as a specification, not as a dependency.** **`wrangler` stops being types-only.** Ticket 11 added it as a peer for `InferEnv`; ticket 16 made `ursprung dev` shell out to it, so it is a **runtime CLI dependency of the dev loop** as well as a type source. **Build-time enforcement of what a signal may hold is gone**, because ticket 34 left the build with a lexer and it can never see what a signal holds. Enforcement moves to a `Serializable` natural signature in `ursprung check` plus an authoritative retention-path throw at serialize time. ## The two axioms that were challenged and held Day two recorded three findings that contradicted axioms set at charting. Two of them came due yesterday, and **both were upheld** — on evidence, not preference. ### No virtual DOM Ticket 19 **reproduced Qwik's premise exactly**: the DOM's _element model_ cannot express a component boundary. Attribute schemes fail on text-only, empty and multi-root components; positional paths break silently on the first mutation. Then it showed the conclusion does not transfer, **because comment nodes are not part of that model.** A component boundary is **a pair of HTML comments — `` … ``** — and a resumed document locates a boundary and replaces its subtree with server-sent markup **with no component code on the page at all**. That was proven structurally: the runtime has no imports. Ranges beat points, because a point's extent has to end at "the next marker or the parent's end", which is wrong for a component that renders nothing and for a fragment wrapping a component. A fifth scheme found on the way — `adaptive`, emitting a closer only where ambiguous — matched `range` at **half the bytes** and was **refused on diagnosis**: when the absence of a closer is _meaningful_, a build bug fails silently instead of raising. Ranges are **self-describing, zero side-channel**, which lets a navigation payload or a streamed fragment carry its own boundaries. Effect disposal falls out as a containment test at no marginal cost. Comment markers are **not** foster-parented inside ``. The cost is stated with its denominator, because otherwise the number means nothing: **21.9 raw bytes per component, +6.9% raw and +47.9% gzipped** on a realistic page. The axiom now carries a limit rather than an open challenge. Ticket 19 verified locate, replace, insert and dispose; it did **not** verify reconciling two component trees, which is where a VDOM historically earns its keep. That is ticket 32, and it carries the same instruction: **if a diffing structure turns out to be unavoidable, amend the axiom rather than quietly violating it.** ### Signals Ticket 18 did not defend the signals decision. **It replaced its justification.** The original ground was _"signals will one day be native"_, and that is not supportable: the proposal is **Stage 1**, `spec.emu` is a 313-byte stub declaring `stage: 0` with zero clauses, `tc39.es/proposal-signals` 404s, there is **no browser-engine champion** and no plenary appearance since 2024-06-13, the package has been **dormant for 18 months**, its own README says **"Do not use this in production"**, and **Angular, Vue, Preact and Solid all ship their own**. The replacement is structural rather than aspirational: **the build has no parser.** Tickets 34 and 36 removed every one, and what reads a module is a lexer whose entire output is a specifier list. Compiler-derived reactivity — Svelte-style dependency analysis, compiled subscriptions — is therefore not something Ursprung declines to build, it is something Ursprung is **incapable** of building. Read-time dependency discovery is the only kind left. **Standardisation is demoted to a free hedge**: if the proposal ships, ~800 lines get deleted. **The polyfill is vendored, not depended on**, and the decisive argument is local: **there is no bundler, therefore no tree shaking, therefore whatever a package exports is shipped.** Vendoring is the only tree shaking this framework has. Nine divergences from the proposal were triaged three-fixed, three-deleted, two-deliberately-kept, under the rule _observable behaviour matches; the implementation need not_. `Signal.*` is never exported — the public surface is `signal()`, `computed()` and `effect()` — so a future swap is invisible to apps. Two consequences with teeth. **On the server, a signal must be created inside a request**, which collapses both server hazards into one rule and **fails at deploy rather than at a user's request**, because Cloudflare runs module scope once at startup. And **`ursprung` gains two realm entry points chosen by `exports` condition, which must be type-identical**, because `ursprung check` is one tsc program that resolves only one of them. Recorded as **ADR-0001, the project's first** — with a review _trigger_ rather than a review date, and an explicit note that upstream polyfill activity is **not** one of the triggers, because the code is ours now and a new release is a diff to read. ## The format that shrank because the domain closed Ticket 9 settled the resumability serialization format, and the headline is that **closing the value domain is what deletes the machinery.** A signal may hold **`JSON + bigint + Date` and nothing else**, so there are no user-authored cycles and no shared identity — which removes allocate-then-inflate, the back-reference table, and delta-encoded reference ids in one go. Qwik needs all three because it accepted an open domain first. The domain is a **strict subset of capnweb's** (`S9 ⊂ S21`), deliberately, so _"anything the server sent you fits in a signal"_ is true by construction rather than by coincidence. `Error` is the one exclusion, because its **message** leaks internals into a cacheable public document — and the spike corrected the original reasoning on the way: the **stack does not cross the wire**. Encoding is JSON in capnweb's `[typeCode, ...args]` shape inside an inert `