Multi-Platform Pyreon

Status: PMTC (Pyreon Multi-Target Compiler) is experimental — demo-quality, self-rated 66/100, not production-ready. The 15-primitive canonical vocabulary spans all three targets — every primitive has a real web DOM runtime AND emits SwiftUI + Jetpack Compose. Be precise about what "validated" means, because the layers differ sharply in strength:

• (1) Per-PR gate — SYNTAX + STUB TYPE-CHECK. swiftc -parse (syntax, accepts unresolved types) runs on every emit; kotlinc type-resolves against Compose stubs; and swiftc -typecheck type-resolves against SwiftUI+PyreonRuntime stubs (swift-stubs.ts, the Swift sibling of kotlin-stubs.ts) — the stubs' generic constraints mirror the real SDK EXACTLY (animation<V: Equatable>, so a [Todo] value is correctly REJECTED) so this catches the type-corruption class that pure -parse waves through. Scope: the stub -typecheck now covers the 2 example apps + all 37 compiler fixtures — canonical primitives, common modifiers, i18n/machine/permissions/link/webview, the rx fixtures, the router-hook surface (PyreonRouter/useNavigate/useParams/matchPath/RouterProvider), the PyreonForm binding surface, and every @Observable fixture including the two LARGE showcase apps (showcase-finance/showcase-tasks, closed by M-gate.1f with PyreonAuth/PyreonDatabase/PyreonFetch + LazyVStack/Color/Font.custom/.foregroundColor/.task/navigationDestination stubs, all mirrored from runtime-swift). It also compiles an example's app-provided useNativeModule classes alongside the emit — the framework cannot stub a class the app owns, and a synthesized stand-in would mask a mismatched method name. The gate keeps earning its keep by SURFACING real emitter bugs that were then FIXED so the emitted native code actually compiles: rx .first/.last/.min/.average now infer Optional/Double bindings; a null-returning component emits EmptyView() not body { nil }; a top-level object-shape interface is SYNTHESIZED into a struct/data-class; and — found while closing the showcase apps — useDatabase's get/delete/find emitted Swift WITHOUT the argument labels the runtime declares (delete(_ collection:id:)), so those calls had never compiled. Argument labels are a TYPE-level concern, which is exactly why only a -typecheck gate could see it — see the silent-failure cliff below.

• (2) Real-toolchain BUILD of the four example apps — xcodebuild (full Swift compile) + gradle assembleDebug — via the native-device workflow (auto-runs on native-path PRs; conclusion advisory until the 14-green-nightly streak gate passes). This is the only place type-checking against the REAL SDK (not stubs) happens, and only for those apps.

• (3) Launch + interaction UI smokes — XCUITest (iOS Simulator) + Compose-instrumented-test (Android Emulator). Strong for the Tasks/Counter/Router examples (login validation, store mutation, typed-params nav, content-asserted fetch, Suspense/ErrorBoundary all assert real behaviour); launch-only for TodoMVC; absent for native-analytics.

Layers 2–3 now run automatically on every PR that touches the native pipeline (packages/native/** / examples/native-*/**, via the workflow's fail-closed changes job — the repo is public, so the macOS runner is free; the native-device label remains the manual opt-in for path-external changes) plus the nightly schedule. Their conclusion is still advisory (not in branch protection): the encoded promotion gate is bun scripts/check-native-device-streak.ts — 14 consecutive green nightlies — and the workflow shape is now required-check-safe by construction (job-level skips, no trigger-level path filter), so passing the streak gate makes promotion a pure branch-protection change. So "it compiles" on a PR means "the syntax parsed," not "it builds + runs on a device." Treat native output as demo-grade until the device gate is required.

The pitch

Write your app once. Run it on the web, iOS, and Android — each rendered with the platform's native primitives. (Web runs live today; iOS/Android are emitted as SwiftUI/Compose and build cleanly for the example apps — read "What runs on native" and the validation Status above for the honest scope, including which packages are web-only.)

// examples/native-todomvc-ios/src/TodoApp.tsx — single source, three targets
import { signal, computed } from '@pyreon/reactivity'
import { useStorage } from '@pyreon/storage'
import { Stack, Inline, Text, Field, Button } from '@pyreon/primitives'

export function TodoApp() {
  const todos = useStorage<Todo[]>('todos', [])
  const draft = signal('')

  return (
    <Stack gap="md">
      <Field
        value={draft}
        onChangeText={(text) => draft.set(text)}
        placeholder="What needs to be done?"
      />
      <For each={todos} by={(t) => t.id}>
        {(t) => (
          <Inline gap="sm">
            <Text>{t.text}</Text>
            <Button onPress={() => /* ... */}>Remove</Button>
          </Inline>
        )}
      </For>
    </Stack>
  )
}

This single file compiles to:

• Web via @pyreon/runtime-dom — <Stack> becomes <div style="display:flex;flex-direction:column">, <Field> becomes <input>, etc.

• iOS via PMTC → SwiftUI — <Stack> becomes VStack, <Field> becomes TextField("", text: $draft), etc.

• Android via PMTC → Jetpack Compose — <Stack> becomes Column, <Field> becomes TextField(value, onValueChange), etc.

Same source. Three idiomatic outputs (web rendered live; iOS/Android emitted as SwiftUI/Compose — see the validation reality in the Status note above).

What runs on native — and what's web-only (read this first)

PMTC compiles your component source — the 15 canonical primitives, signal/computed/effect, and a fixed set of hooks — to SwiftUI/Compose. It does NOT transpile npm packages to native. So a package runs on iOS/Android only if it is pure reactive logic with a Swift/Kotlin runtime port the compiler recognizes. Anything bound to the DOM, a <canvas>, the CSSOM, or a JS-only rendering vendor is web-only by architecture — no compiler setting changes that.

The fixed hook set is not a ceiling on platform capability. For anything the framework does not ship — Bluetooth, ARKit, a vendor SDK — useNativeModule lowers to a Swift/Kotlin class you provide, so adding a platform API is an app-level change rather than a framework PR.

✅ Runs on web + iOS + Android today (real runtime ports + compiler emit):

• @pyreon/reactivity (signal/computed/effect), @pyreon/primitives (the 15 canonical primitives)

• @pyreon/store, @pyreon/machine, @pyreon/state-tree, @pyreon/i18n, @pyreon/permissions

• @pyreon/form (validated forms — device-proven), @pyreon/storage (platform-storage backend)

• the native @pyreon/router port (useNavigate/useParams/useLoaderData, nested routes, beforeEnter)

• the subset of @pyreon/hooks with ports: useFetch, useOnline/useNetworkStatus, useAppState, useClipboard, useColorScheme

  • useFetch needs its response type on iOS: useFetch<Resp>(url). Without the generic the Swift emit decodes into Any, which cannot conform to Decodable, so the iOS build fails — Android compiles either way. The compiler now warns rather than letting you find out at build time.

  • useParams() must be destructured: const { id } = useParams() lowers per key. Binding the whole object (const p = useParams(); p.id) reads a property off a native dictionary/map and compiles on neither target.

❌ Web-only by design (a hard DOM/canvas/vendor dependency the compiler has no path for — these will NOT be native-rendered as SwiftUI/Compose; PMTC can't compile echarts/CodeMirror/elkjs/etc. But several CAN be hosted in a <WebView> — see the bridge escape hatch right after this table):

🌉 Escape hatch — host a web-only component in a <WebView> (the bridge). The "❌" packages can't be native-rendered, but a <WebView> embeds a real browser engine (WKWebView on iOS, Android WebView), so the web component runs inside it — echarts' canvas, flow's elkjs+SVG, CodeMirror, a document preview. The bridge is bidirectional:

• Forward — data={metrics()} is pushed into the page as window.__pyreonData (+ a pyreondata event) so the hosted component updates live, no reload.

• Reverse — the page calls window.pyreonPostMessage(payload) → your native onMessage={(m) => …} closure.

// examples/native-analytics — a chart hosted natively, both directions:
<WebView html={CHART_HTML} data={metrics()} onMessage={(m) => selected.set(m)} />

This is the right answer for charts / diagrams / code editors / doc previews (you wouldn't reimplement echarts in SwiftUI anyway). Caveats: it's a hosted web view (web look-and-feel, not native widgets), pays WebView boot + bundle weight (echarts ~1 MB), and is best for self-contained panes — not your whole app. The bridge runtime (PyreonWebView.swift/.kt) is real + swift build-clean and native-analytics demonstrates it, but a chart-in-WebView device test isn't in the nightly gate yet (the mechanism is proven; the on-device run isn't CI-gated). For your core app UI (nav/forms/lists/layout) use the native primitives; reach for the WebView only for the rich web-island pieces.

📦 Shipping a local viz bundle — <WebView src="…"> + the web/ staging step. Inline html= embeds a self-contained string; for a multi-file bundle (an index.html + its chart.js/chart.css), use <WebView src="chart.html"> and drop the files in a web/ directory at your project root. The scaffold's build scripts run pyreon-native stage-web, which copies that flat bundle into the exact app location each runtime resolves src against — iOS WebContent/ (an XcodeGen type: group → the files flatten to the app bundle's resource root, found by Bundle.main.url(forResource:)) and Android assets/ (served at file:///android_asset/). This keeps the whole bundle on-device (no remote fetch — the policy-safe path Apple 4.2 / Google's webview policy prefer) and lets chart.html's relative <script src="chart.js"> resolve. v1 is flat-only — a nested subdirectory is skipped with a build warning (the shipped Swift runtime resolves src by bare name with no subdirectory:; nested support is a runtime follow-up), and the on-device render rides the same nightly device rung as the rest of iOS — the staging layout itself is unit-locked, but "the WKWebView actually paints the staged file" is device-gated, not CI-proven.

🟡 Logic could port, but no native runtime exists yet: @pyreon/rx, @pyreon/validate, @pyreon/validation, @pyreon/url-state, @pyreon/toast (the store is pure-logic; the <Toaster> renderer is DOM), and @pyreon/sync's engine-neutral core (the Yjs engine + IndexedDB/WebSocket transports are web/Node-only).

The supported-TypeScript-surface ceiling (and the silent-failure cliff)

PMTC compiles a deliberately narrow, declarative subset of TypeScript in component bodies: signal/computed/effect declarations, typed props, the canonical-primitive JSX, <For>/<Show>, if/ternary control flow, array/string method calls, and two type-alias shapes (string-literal union → enum, object literal → struct/data class). Inside that lane it emits real, idiomatic SwiftUI/Compose.

Outside that lane the failure mode is sometimes silent, sometimes a named warning — the silent-drop surface is shrinking. Map/Set/Date and generics in logic are dropped or mis-emitted — some still with no warning. Imperative control-flow statements at a component-body top level — for / for…of / for…in / while / do…while / switch / try / throw, and an imperative if (a mutating body) — now fail with a NAMED warning pointing at the escape hatch (move the logic into a helper function that takes parameters — loops + switch DO lower there — compute the value with array methods .map/.reduce/.filter, or render conditionally/iteratively in JSX via <Show> / <For>), no longer a silent drop that left the render on stale values. An EARLY-RETURN conditional render — if (cond) return <JSX> (optionally else return <JSX>, and chained: if (a) return <A>; if (b) return <B>; return <C>) — now LOWERS: it folds to a ternary the emitter emits as a native result-builder conditional view — a SwiftUI @ViewBuilder if cond { A } else { B } (NOT the ? : operator, which swiftc rejects between DIFFERENT view types: sizeClass == "regular" ? HStack {…} : VStack {…} fails "result values in '? :' expression have mismatching types HStack<Text> and VStack<Text>" — the adaptive Stack↔Inline case), Compose if (cond) A else B. Both a DIRECT view-branch ternary (sizeClass() === 'regular' ? <Inline> : <Stack> — the size-class-driven adaptive-layout idiom) and the folded early-return lower this way, verified to swiftc -typecheck + kotlinc (imperative control flow can't sit in a SwiftUI var body result builder — only a conditional VIEW can — so this is the one control-flow shape that lowers directly in a component body). A VALUE ternary (cond ? "a" : "b") stays a ? : expression. Pre-fix the if was dropped (rendered only the fallthrough branch), or — with an else-return and no fallthrough — the WHOLE component was skipped. instanceof/in already warn. (Destructuring of locals is now lane-precise: flat const {a} = obj / const [a, b] = xs LOWER; nested / rest / default patterns — const {a:{b}} = o, const {a, ...r}, const [a, ...r], const {a = 1} — now fail with a NAMED warning pointing at the escape hatch, no longer a silent drop. try/throw and class/new already emit a named "unsupported" warning. A top-level pure-logic helper function — function dbl(x: number): number { return x * 2 }, the L1 "shared pure logic" layer — is now EMITTED at file scope as a native func / fun (reusing the same function emitter store methods use), where before it was silently misclassified as a component and mis-emitted as a broken struct dbl: View (its value params dropped, the body referencing an unbound name → a cryptic cannot find 'x' in scope). A call computed(() => dbl(21)) infers the helper's return type (a helperReturns registry threaded into the computeds pre-inference), so the Swift computed annotates Int, not Any (String(dbl(21)) typechecks). The classifier is narrow — only a function that takes value parameters and returns no JSX (a function OF ITS INPUTS) is a helper, so a no-param function C() { …; return out } component-returning-a-value emits unchanged, and a component (return resolves to JSX — directly or through a cond ? <A/> : <B/> / && root) is never misread. BOTH declaration forms are recognized: a function dbl(){} AND a top-level ARROW-CONST const dbl = (x: number) => x * 2 (tryHelperFnFromArrowConst routes the arrow-const into the same helperFns path — before, it fell through to a mis-scoped private let dbl = { x in … } closure + Any, a silent uncompilable mis-emit); a JSX-returning arrow-const (a component), a no-param arrow-const, and a plain module const (const APP = '1.0') fall through unchanged. A helper with NO return-type annotation (function dbl(x: number) { return x * 2 }) also emits — its return type is INFERRED from the body (refineHelperReturns seeds the params + walks for the first return, reusing the same inferReturnType util the emitters use for un-annotated function signatures), so the emit signature AND the call-site helperReturns registry both get the real type. A helper with a FRACTIONAL body (function scale(x: number) { return x * 1.5 }, x / 2, Math.sqrt(x)) also emits: emitSwiftFunction seeds the helper's Int params into the coercion ctx (the same _activeInferCtx.locals the element-callback coercion uses) so x * 1.5 → Double(x) * 1.5, and refineHelperReturns refines the number return Int → Double so the signature matches (Kotlin auto-promotes Int×Double, needing only the Double return). Division lowers to Double(x) / Double(2) (JS-fractional 4.5, not integer 4). ONE shape keeps a NAMED warning (deferred, never a broken emit): a GENERIC helper (function first<T>(…) — the native IR can't represent <T>); a body whose type still can't be inferred (an untyped param the return depends on) also warns + is dropped. The emit + call-site inference are proven on both real toolchains (swiftc -typecheck + kotlinc). A JSX spread on a canonical primitive — <Stack {...cfg()}> — now fails with a NAMED warning too (its layout props were silently dropped before: a runtime prop-bag can't apply to a static SwiftUI view / Compose composable, so pass props explicitly — <Stack gap="md">); a spread on a USER component still expands against its declared props.) (Recently CLOSED — these now LOWER and are no longer dropped/mis-emitted: template literals `Hi ${name}` → native interpolation; for…of/while/switch + reassignment (t = t + x, +=) + multi-declarator; optional chaining a?.b (member); object destructuring of locals const {a} = obj (body-local + hook-result, e.g. const { data } = useFetch(url)); untyped object / array-of-object signals → a synthesized struct with a precise type annotation (signal({x:1}) / signal([{id:1,name:"a"}]) emitted a broken Any annotation on Swift, now an inferred struct/[Struct]); nested-array signals (signal([[1,2],[3,4]]) / [[[1]]] / [["a"]]) — inferTypeFromInitial had scalar-array + flat-object-array cases but NO array-OF-arrays case, so a grid/matrix signal degraded to Any (the value [[1,2],…] was valid Swift but the annotation failed swiftc, and a downstream grid()[0][1] then also degraded). Now recurses into the element ([[Int]] / [[[Int]]] / [[String]]); a fractional leaf flags EVERY nested integer literal float ([[1.5],[2]] → [[Double]] = [[1.5], [2.0]]) so Kotlin's List<List<Double>> accepts it. Swift-only in effect — Kotlin infers List<List<…>> on its own; .sort + chained array-method element inference ([...xs()].sort(cmp).slice(0, n) — .sort was inferred Any, breaking the chained .slice); .flatMap result-type inference — .flatMap already EMITTED natively (arr.flatMap({…}) on both targets), but inferType had no case, so the computed typed Any on Swift and a chained .length failed. Now arr.flatMap(x => [E]) flattens one level → [E] (the callback's ARRAY body type ITSELF, unlike .map(x => [E]) → [[E]]). This surfaced a sibling array-literal homogeneity bug: n * 2 returned { number, float: false } while a bare n returned { number }, so [n, n * 2] was treated as heterogeneous → unknown → [Any]; fixed to follow the "float ONLY when true" convention (a non-float result is now byte-identical to { number }), which ALSO fixes any array literal mixing a value + arithmetic ([a, a * 2]). (Swift-only — Kotlin infers on its own. A String()-constructor body + the Int * Double coercion are separate pre-existing gaps.) seedless .reduce(fn) — JS's no-initial-value reduce (arr.reduce((a, b) => …) seeds the accumulator with the FIRST element) had no Swift lowering: the bare arr.reduce({…}) bound to reduce(into:) ("missing argument for parameter 'into'"). Now lowers to arr.dropFirst().reduce(arr[0], fn) (Kotlin's 1-arg .reduce {…} already matches JS). The lowering names the receiver TWICE, so it only fires when the receiver is RE-READABLE (identifier / signal / store read, via isReReadableExpr); a receiver with a chained method call (filter(…)) that would re-run work emits a NAMED build-failing warning + defers. The seedless result type is the array's ELEMENT type (JS seeds with arr[0]), so the max idiom (a, b) => a > b ? a : b infers correctly instead of Any. object-literal → declared-struct inference — a computed RETURNING an object literal (() => ({ x, y })), or a .reduce with an OBJECT accumulator (reduce((a, b) => ({ sum, count }), { sum: 0, count: 0 })), now infers the DECLARED struct whose field-set matches the literal (mirroring the emit's _structFieldsToName first-wins lookup) instead of degrading the computed to Any — so a downstream out().sum / out().x field read resolves (was Swift "value of type 'Any' has no member 'sum'"). A literal matching NO declared type X = { … } stays Any (the emit synthesizes an anonymous struct for it, but inference can't name that without the emitter's per-run registry — a follow-up). field access on an inline object literal (({ count: nums().length, label: s() }).count) — a partial answer to the naming follow-up just above: an INLINE object literal infers unknown (there is deliberately no general object-literal inference case — a nameless literal has no struct NAME for an annotation), so ({…}).count degraded the computed to Any — the emit (__Obj0(…)).count is a valid Int EXPRESSION, but the Any annotation fails swiftc once consumed (String(out())). Now the FIELD'S type resolves directly from the literal's own fields (unwrapping the paren node ({…}) wraps it in) → out: Int / String / Double, WITHOUT the struct name (a field access needs the field type, not the name). A SPREAD literal ({ ...base, y }) bails (stays Any); the intermediate-const form (const o = {…}; return o.count) is a separate multi-statement-dataflow gap. field access on a TERNARY of two object literals ((cond ? { v: 1 } : { v: 2 }).v) — completes the field-access-on-object-producing-expr class the inline-object-literal case opened: a TERNARY operand also infers unknown, so (cond ? {v:1} : {v:2}).v degraded the computed to Any even though both branches synthesize the SAME struct (emit (cond ? __Obj0(v:1) : __Obj0(v:2)).v, a valid Int expression). Now the field's type resolves from a branch (paren-unwrapping each), gated on the field existing in BOTH branches → out: Int / String; a MIXED ternary (cond ? {v:1} : {w:2}, different fields) bails to Any. intermediate-const object-field access (const o = { count: nums().length }; return o.count inside a computed) — the COMMON const o = compute(); return o.field shape. findFirstReturnExpr already seeds a computed body's locals into the infer ctx, but an object-literal initializer infers unknown (the no-nameless-object-annotation rule), so o was seeded unknown and o.count couldn't resolve → the computed degraded to Any. Now object-literal consts are ALSO recorded in ctx.objectLocals (name → the literal), and the member case resolves o.count from the recorded fields → out: Int / String — WITHOUT changing o's own type in locals, so a bare return o (whole local object) is UNCHANGED (still Any, not a new tuple emit). Only a FIELD access newly resolves. Object.keys() / Object.values() over a known object shape — keys → a static key array (["a","b"] / listOf("a","b")), NOW including a DECLARED-struct arg (Object.keys(p()) where p: signal<P> — the typeRef resolution gap that made the dominant real shape degrade-warn); values → a static member-access array ([p.a, p.b] / listOf(p.a, p.b), field order = declaration order) gated on ALL field types identical (JS's mixed values array has no native analog) + a re-readable receiver (named once per field). .entries, mixed shapes, chained receivers, and non-struct args keep the typed-empty degrade + NAMED warning — never the old silently-uncompilable Object.keys(...); Int×Double coercion inside an element-callback body — Swift has no implicit Int→Double conversion, so nums().map(x => x * 1.5) (an Int-array param × a fractional literal) emitted a bare x * 1.5 INSIDE the closure → "cannot convert value of type 'Int' to expected argument type 'Double'". Component-scope coercion (n * 1.5 → Double(n) * 1.5) already worked, but an element-callback PARAM is neither a signal nor a const, so inside the closure it inferred unknown and the coercion never fired. Now the emit binds a .map/.filter/.forEach/.find/.some/.every/.flatMap callback's first param to the receiver's element type while emitting the closure, so +/-/* on that param coerce (Double(x) * 1.5 / - 0.5 / + 0.5) — including inside a filter predicate; nested/sibling closures each see their own element type (restore via try/finally). (Swift-only — Kotlin auto-promotes Int×Double arithmetic. A .flatMap whose body is itself a .map still infers Any — a separate flatMap-of-map inference gap.); 2-param index-callback Int×Double + mixed-comparison coercion — completes the element-callback coercion family: the 2-param (el, idx) form lowers through the separate enumerated() path, which skipped the element-type scoping, so .map((x, i) => x * 1.5) emitted the bare x * 1.5 ("cannot convert value of type 'Int' to expected argument type 'Double'"); the indexed-closure emit now binds the element param to the receiver's element type + the index param to Int (try/finally-restored). Fixing that exposed the sibling gap: Swift also requires same-type operands for COMPARISONS (x * 1.5 > i failed at Double > Int), so the comparison emit now coerces the Int side when exactly one operand is Double — non-literal operands only (Swift self-types integer literals in a Double context, so > 2 stays bare and every existing emit is byte-identical); enum/string/bool comparisons untouched (numericFloatness returns 'other'). (Swift-only — Kotlin allows Int↔Double comparison + auto-promotes arithmetic.) ternary empty-array branch unification — cond ? [x] : [] (the conditional filter-map idiom, esp. as a .flatMap body: nums().flatMap(x => x > 1 ? [x] : [])) and the mirrored cond ? [] : [x] degraded the whole ternary to Any (a bare [] carries no element type → unknown → the branch-kind equality check failed), breaking any typed consumer. The ternary inference now unifies to the array branch's type when the OTHER branch's expr is an untyped empty array LITERAL (never on a mere unknown type); both emits already compiled under the unified annotation (Swift types [] bidirectionally from context; Kotlin's listOf() is List<Nothing>, a subtype) — an inference-only fix; mixed-type ternaries (cond ? 1 : "x") still degrade honestly. Date.now() — emitted VERBATIM on both targets (a clean-parse silent mis-emit: Swift "cannot call value of non-function type 'Date'" — Foundation's Date resolves as a TYPE; Kotlin "unresolved reference 'Date'"). Now lowers to epoch-ms as a DOUBLE (Swift Date().timeIntervalSince1970 * 1000 / Kotlin System.currentTimeMillis().toDouble()) — Double because ms-since-epoch (~1.7e12) OVERFLOWS Kotlin's 32-bit Int (PMTC's number→Int default) and Double is exact below 2^53; the float inference composes with the Int×Double coercion (Date.now() - start()) and the Math return-type work (Math.floor(Date.now()/1000) → Int). Other Date.* statics (Date.parse, …) → a NAMED build-failing warning; new Date() already warns via the class/new path. KNOWN pre-existing limit (general, tracked): .set(<float expr>) into an Int-typed signal (signal<number>(0) + start.set(Date.now()) — the stopwatch shape; same for price.set(1.5)) fails LOUD with a native type error — write-site float-WIDENING of the signal's declared type is the next inference gap. <For by={(x) => x}> identity keying — the by resolver only understood the member shape ((i) => i.id) and SILENTLY fell back to .id for everything else, so an identity key over a plain string list (<For each={names} by={(n) => n}> — the tags/subjects shape) emitted the uncompilable ForEach(names, id: \.id) / key = { it.id } with zero warnings. Identity now lowers to id: \.self (String/Int are Hashable) / key = { it }; member keys unchanged; any OTHER by-shape (computed keys) emits a NAMED build-failing warning — never silent. Surfaced by the StatsPage device example (the realistic-app discovery pattern, 6th instance); the example now rides the auto-firing device gate in examples/native-tasks exercising this sprint's vocabulary end-to-end on real toolchains (Object.keys/values over a declared struct, seeded reduce, Double division, the filter-map idiom, 2-param indexed Int×Double + mixed comparison). the idiom-sweep canary + its five first-run finds — a permanent regression corpus (native-idiom-sweep.test.ts) asserts common JS idioms are LOUD-OR-TYPECHECKS on BOTH targets (a warning-free emit must pass real swiftc -typecheck AND kotlinc — a Swift-clean / Kotlin-broken emit, or vice-versa, can no longer slip through; a future silent MIS-emit regression fails the canary before shipping). It also locks a statement-context corpus (handler bodies): a ||= b / a &&= b / a ??= b stay LOUD because a naive parse-time desugar to a = a || b is UNSOUND — JS ||/&& are truthiness ops over any type, but Swift/Kotlin ||/&& are Bool-only, so n &&= 5 → n = n && 5 fails both toolchains (a faithful lowering needs type-aware emitter gating, not a parse desugar). Its first run caught five SILENT fails, all fixed: arr.join(sep) on a NON-String array (Swift joined(separator:) is [String]-only — now element-type-aware, mapping String.init first); arr.lastIndexOf(x) (now lastIndex(of:) ?? -1, indexOf's mirror; non-array receivers warn NAMED); arr.flat() (the emit existed but no inference case → Any — now array-of-array → inner array); Number.isInteger(x) (raw emit failed BOTH targets — now Int arg → true, Double → the parenthesized remainder check, unknown → NAMED warning; infers boolean); Math.max(...arr)/Math.min(...arr) (the SPREAD form bypassed the fixed-arity mapping on BOTH targets — now the collection max()/min() with JS's empty-array sentinel analog: Int.min/Int.max for Int arrays, ±infinity for Double). Kotlin default WebSocket transport (OkHttp) — PyreonWebSocket's Android side previously required a HOST-SUPPLIED transport (connect(register: (WebSocketHandlers) -> WebSocketSender) — the documented URLSession-vs-no-JDK-socket asymmetry). The runtime now ships PyreonWebSocketOkHttp.kt: a connect(url: String) extension wiring a shared OkHttp client's WebSocketListener into the container's pure state machine (onOpen/onMessage/onFailure/onClosed → the reactive fields) + a WebSocketSender (send/close(1000)), matching Swift's connect(to: URL) one-for-one. The extension is the ONLY runtime source importing okhttp3 (the core container stays dependency-free per its design contract); every Android example's gradle gains the okhttp dep (the srcDir compiles all runtime sources); the per-service kotlinc verify gains an okhttp3 stub set mirroring the real 4.x surface EXACTLY (typecheck-only — the semantic proof is the device build with real OkHttp, which the auto-firing device gate runs on the PR). UNBLOCKS: the compiler flipping Android's ws.connect() from the named transport warning to the faithful connect(url) emit, and lifecycle AUTO-START on both targets (a synthesized onMount(connect)) — both land with the onMount-lowering PR once it merges. !x/!!x truthiness + isNaN + two loud-guard conversions (idiom-sweep batch 2 — five more SILENT fails): !x on a non-Boolean is JS truthiness negation ("type 'Int' cannot be used as a boolean") and !!x was doubly broken (juxtaposed unary is a Swift parse error) — both now lower by the arg's inferred type on BOTH targets (number → == 0/!= 0, string → isEmpty/non-empty, boolean verbatim, optional → nil/null check; unknown stays raw-loud). isNaN(x): Int arg → statically false, Double → the native .isNaN/.isNaN(), infers boolean. String .at(i) (the ARRAY lowering emitted uncompilable garbage on a String — Swift String indices aren't Int; Kotlin getOrNull yields Char?) → gated to arrays + a NAMED warning on strings. A multi-statement .sort comparator (silently dropped via the block-body sentinel; Swift's < 0 Bool conversion can't wrap a block) → NAMED warning both targets, expression bodies unchanged. onMount(fn) lifecycle lowering — THE documented lifecycle escape hatch ("call .start()/.connect() from an onMount") was a SILENT drop: the component-body walker only handled declarations + return, so onMount(() => ws.connect()) compiled clean with zero warnings and did nothing on device (the worst class: silent + documented). Now lowers to SwiftUI .onAppear { … } on the stable-identity ZStack host (the fetch-arc .task trap applies to .onAppear too) / Compose LaunchedEffect(Unit) { … }. Companions: ws.connect() (0-arg TS surface) threads the useWebSocket(url) decl's url into Swift's connect(to: URL(string: …)!); on Kotlin it emits a NAMED warning + the loud raw call (the runtime's connect(register:) needs a HOST-SUPPLIED transport — the default-OkHttp-transport is the tracked follow-up, after which compiler AUTO-START becomes a synthesized onMount). A returned cleanup fn → NAMED warning (mount body still emitted; unmount cleanup is v2). await hook.method() in an async event handler (M4.5) — an async-RESULT service call awaited inside onPress={async () => { const ok = await bio.authenticate('…'); status.set(ok ? 'ok' : 'denied') }} now LOWERS (before, any await in a component was a named "use useFetch" warning that DROPPED the call): a synchronous SwiftUI/Compose action slot can't await, so the handler body is wrapped in a native async scope — SwiftUI Button { Task { let ok = await bio.authenticate('…'); … } }, Compose onClick = { pyreonAsyncScope.launch { val ok = bio.authenticate('…'); … } } with a composable-top val pyreonAsyncScope = rememberCoroutineScope() hoisted once (a Kotlin suspend call carries NO await keyword — the coroutine provides the context) — so the post-await statements run when the async result resolves. A SYNC handler is untouched (no Task/launch wrap). The first async-result consumer is useBiometrics() (recognized natively → @State private var bio = PyreonBiometrics() / remember { PyreonBiometrics() }; iOS LAContext biometrics-gate runtime ships, and the @pyreon/hooks web hook now ships too (useBiometrics() — web feature-detects PublicKeyCredential + resolves false, a real WebAuthn assertion needs a server challenge; the Android BiometricPrompt/FragmentActivity runtime is a follow-up, the v1 scaffold resolves false). Proven on both real toolchains — swiftc -typecheck (the Task { await … } shape) + kotlinc (the scope.launch { … } shape) — AND now DEVICE-PROVEN (M3.5): the shared counter's Unlock button awaits the gate; on an UNENROLLED Simulator/emulator it resolves false with NO prompt (biometrics-only policy, canEvaluatePolicy guard), so the observable outcome flips Lock: idle → Lock: denied, and the iOS XCUITest + Android Compose test assert that flip — proving the async scope RUNS at runtime (the post-await re-render fired), not just compiles. Any OTHER bare component-body statement (bare effect(...), stray calls) now warns NAMED — the whole silent-expression-statement class is closed. component-body top-level reassignment (let a = 1; a = 5; / a += 2 / a++ at the component top level) — completes the above: a reassignment is an ExpressionStatement whose expression is an AssignmentExpression/UpdateExpression (NOT a CallExpression), so it fell past the bare-call warn branch into the intentional no-op drop meant for harmless void x discards → a REAL mutating reassignment was silently dropped (the render used the initial value). Now a NAMED warning: a component body emits declarations + the return JSX, not setup-time statements — components run ONCE (compute the final value directly with const x = …, or use a signal). Harmless void x / bare-ref / unary / logical discards stay silent (only genuine reassignments warn — the documented over-eager-regression guard holds). component props via a NAMED local type — type CardProps = { qty: number; label?: string } + function Card(props: CardProps) (the DOMINANT component shape) parsed to EMPTY props with NO warning: the emitted component declared no stored properties (Swift) / parameters (Kotlin) while its body referenced them bare and call sites passed args — uncompilable on BOTH targets (only the INLINE annotation props: { … } and destructured-inline shapes worked). A pre-pass now registers every local object-shape type alias and the props annotation resolves it regardless of declaration order (destructured NAMED refs too); an UNRESOLVABLE ref (imported type, interface) fails with a NAMED warning instead of the silent garbage emit. FOUR sibling gaps closed in the same class: (1) optional fields dropped ? everywhere — label?: string now parses as the union-with-undefined convention, so structs/data-classes emit var label: String? = nil / var label: String? = null and component props var label: String? = nil / label: String? = null, the explicit default making the memberwise/named-arg parameter OMITTABLE (<Card qty={2}/> compiles; before, the field emitted REQUIRED and every omitting site failed); (2) Swift call-site arg order — JSX attrs emitted in AUTHOR order, but Swift's memberwise init hard-errors out of DECLARATION order (argument 'qty' must precede argument 'label') — args now re-sort against the target's declared props (the old spread spec had CODIFIED an uncompilable order); (3) props never seeded inference — a computed over props.qty (member form) or a bare destructured prop annotated Any on Swift, breaking String(total())/arithmetic; both read shapes now resolve the declared prop type; (4) bare optional in Text rendered Optional(x) (Swift debug description) / literal null (Kotlin) where JSX renders EMPTY — an optional-typed interpolation now emits \((x).map { "\($0)" } ?? "") / ${x ?: ""} (a ??-collapsed read keeps the plain byte-shape). Whole-app emit proven by real swiftc -typecheck (multi-component + omitted optionals + out-of-order attrs). Known follow-ups: an optional inside a TEMPLATE literal still renders the raw value; a FUNCTION-typed field still derives Codable/@Serializable (uncompilable — the optional-callback-prop arc). function-typed struct fields — a declared type carrying a callback (type RowActions = { label: string; onDone: () => void }) emitted struct RowActions: Codable — closures aren't Codable, a HARD swiftc error ("does not conform to protocol 'Decodable'") — and @Serializable data class — the kotlinx serialization plugin rejects function properties on the REAL Compose build while the kotlinc validate STUBS mask it (a device-gate-only red). The conformance is now GATED: any field containing a function (directly, in a union branch, or as an array element — typeContainsFunction) drops : Codable / @Serializable (named structs AND synthesized anon-object data classes); function-free structs are byte-identical. Faithful, not a limitation: a type carrying functions can't JSON-round-trip in any language. TWO siblings in the same class: (() => void) | undefined hit the parser's unknown default (TSParenthesizedType was unhandled) → the whole union silently degraded to Any? (compiles for assignment, uncompilable the moment the callback is CALLED) — parens now unwrap; and with that fixed, the optional-function emit needed PARENS on both targets — a bare () -> Void? / (Int) -> Unit? is a function RETURNING an optional, not an optional function; now (() -> Void)? / ((Int) -> Unit)?. Gated + parenthesized shapes proven by real swiftc -typecheck (construction + cb?() invocation). The fn?.() optional-CALL lowering builds on this optional-type preservation and lands as its follow-up (see below). sweep batch 4 — dynamic a11y/testids, enum switches, JSON, destructured callback params — four finds, three fully SILENT. (1) A template-literal (or ANY dynamic) value in data-testid / accessibilityLabel — the a11y/e2e-critical shape inside For rows (data-testid={``row-${i}``}) — was silently DROPPED on both targets (the modifier emits read static-only): both native slots accept dynamic string exprs, so it now lowers — Swift .accessibilityIdentifier("row-\(n)") / .accessibilityLabel("item \(n)"), Compose .testTag("row-${n}") / semantics { contentDescription = … }; templates splice natively, any other expr string-interpolates (JS coercion); static values byte-identical incl. the container .contain gate. (2) A switch over an ENUM-typed subject emitted raw STRING case labels — case "busy": against a Status enum is a swiftc type error, "busy" -> in a when a kotlinc incompatible-types error — with zero warnings; case labels now map through the existing active-enum context (case .busy: / Status.busy ->; labels only — case BODIES keep string literals; non-enum switches byte-identical). (3) JSON.parse/JSON.stringify emitted VERBATIM — JSON doesn't exist natively, an unresolved reference with no warning — now a NAMED build-failing warning (a Codable/kotlinx serialization bridge is the tracked follow-up, gated on the function-field conformance PR). (4) A DESTRUCTURED callback parameter (.map(([k, v]) => k) / ({ id }) =>) emitted a closure over UNBOUND names — now a NAMED warning (take a plain param, read fields/indices), and the tuple-type annotation warning now names the fix (use an object type) instead of the generic "Unknown type annotation". Dynamic-attr + enum-switch shapes proven by real swiftc -typecheck. regex literals (/pat/flags) — a JS construct with no native form (Swift uses Regex/.firstMatch(of:), Kotlin Regex(...); neither has /…/ literal syntax + .match/.test/regex-.replace), so s.match(/x/) / /x/.test(s) / .replace(/x/g, …) emitted the raw /…/ VERBATIM on both targets — uncompilable, with ZERO warnings (a silent-drop the opening list above didn't even name). Faithful regex lowering (flags, capture groups, differing match APIs) is a tracked follow-up; for now the silent mis-emit is a NAMED warning + a safe "" fallback (never the uncompilable verbatim regex). A STRING-arg .replace("x", …) is a separate path — unaffected. computed object keys ({ [k]: v }) — a computed-key property carries the key EXPRESSION, not a static name, but the object parser matched an identifier-keyed computed prop via p.key?.name and used the VARIABLE NAME as the struct field: { [k]: 1 } (k a var) emitted __Obj0(k: 1) and a downstream o.a / o[k] read missed — a clean-PARSE mis-emit (swiftc -parse accepts it; only -typecheck / device catches the wrong field). A native struct/data-class needs static field names, so a computed key has no faithful lowering → now a NAMED warning (never the silent wrong field); static keys + object spreads are untouched. call-argument spreads (f(...args) / o.h(...args)) — DISTINCT from the array-literal / object spreads above (those lower): a spread ARGUMENT in a call reached the expr emitter's case 'spread' fallthrough and degraded to the bare argument, silently passing the whole array/list as ONE scalar arg (f(xs)) — uncompilable, since Swift/Kotlin calls take a fixed argument list (no variadic spread), with ZERO warnings. Now a NAMED warning at the EMITTER — the correct layer, because the parser can't distinguish a call-arg spread from an array-ELEMENT spread (both are SpreadElement), so the disambiguating context only exists downstream where every faithful consumer (array-concat, object partial-update, Math.max/Math.min) has already extracted its spread. Those consumers are untouched. dynamic disabled on <Field> / <Toggle> — the disable-a-form-control-during-submit shape (disabled={busy()}). <Button> was fixed (Round-1 audit) to use the shared swiftDisabledModifier/kotlinEnabledArg helper, but <Field> + <Toggle> still read disabled via readStaticAttr (static-only) — so a DYNAMIC value was SILENTLY DROPPED on both targets (the control stayed enabled/interactive regardless). All three now route through the shared helper: disabled is a runtime boolean (not a compile-time token), so a dynamic value lowers directly — Swift .disabled(busy), Compose enabled = !busy — no warning. Static byte-identical (.disabled(true) / enabled = false). Proven by real swiftc -typecheck + kotlinc. dynamic placeholder on <Field> — a reactive hint (placeholder={hint()}, or placeholder={terse() ? "Search" : "Search products by name…"}) read placeholder STATIC-only (the same readStaticAttr/readStaticAttrKotlin class as disabled/color), so a dynamic value was SILENTLY DROPPED — Swift fell back to "", Compose omitted the placeholder = arg entirely (the field rendered no hint). UNLIKE the compile-time token props (gap/color/align/level), a placeholder is a RUNTIME String — SwiftUI's TextField(_:text:) accepts a LocalizedStringKey (literal) OR a StringProtocol (runtime String), and Compose's Text(text: String) takes any runtime String — so like the Image dims, a dynamic value lowers DIRECTLY with no warning (both a signal read and a two-literal ternary): Swift TextField(hint, text: $draft) / TextField(terse ? "Search" : "…", text: $draft), Compose placeholder = { Text(hint) } / placeholder = { Text(if (terse) "Search" else "…") }. Static byte-identical (TextField("name", …) / Text("name")). Proven by real swiftc -typecheck + kotlinc. dynamic kind on <Field> (show/hide-password toggle) — kind="password" renders a masked field (Swift SecureField, Compose visualTransformation = PasswordVisualTransformation()); the DYNAMIC toggle kind={reveal() ? "text" : "password"} read kind STATIC-only, so it SILENTLY fell back to the PLAIN field on both targets — the password rendered in CLEARTEXT regardless of the toggle (a SECURITY silent-drop, worse than a dropped modifier). On Swift kind switches the VIEW TYPE (SecureField vs TextField are distinct types, so a bare ternary of the two won't typecheck), so the branches are erased through AnyView into ONE well-typed conditional the modifier chain still binds to: (reveal ? AnyView(TextField(ph, text: $draft)) : AnyView(SecureField(ph, text: $draft))); Compose keeps ONE TextField and toggles the parameter: visualTransformation = if (reveal) VisualTransformation.None else PasswordVisualTransformation(). A ternary of two literal kinds where one branch is "password" lowers; a fully-dynamic (non-ternary) kind → a NAMED warning + a plain-field fallback. Static byte-identical. This ALSO closed a latent STATIC-password device-build bug: no example had used kind="password", so PasswordVisualTransformation (in androidx.compose.ui.text.input, outside the unconditional import set) shipped UNIMPORTED — masked by the kotlinc validate stub (validate-green / gradle-red); the CLI now conditionally imports it + VisualTransformation, and the stub mirrors the real base-type surface. Proven by real swiftc -typecheck + kotlinc. dynamic styling-attr values (gap/padding(X/Y)/background/radius) — a non-static value in a styling attr SILENTLY dropped the WHOLE modifier on both targets with zero warnings (gap={dense() ? "sm" : "lg"} emitted a bare VStack { / Column { — the binary-density idiom just lost its spacing). Styling tokens resolve at COMPILE time (the numeric forms are token INDICES on the 4px scale, not pixels — a runtime number can't map), so the faithful dynamic form is a ternary of two literal tokens: both branches compile-resolve and the condition emits natively — Swift VStack(spacing: (dense ? 8 : 16)) / .padding((dense ? 4 : 12)) / .background((dense ? Color(…) : Color(…))), Compose Arrangement.spacedBy((if (dense) 8 else 16).dp) etc. (shared IR classification + per-emitter emit; conditions run through the optional-truthiness-aware condition helpers). Any OTHER dynamic value (a signal read, an arbitrary expression) now fails with a NAMED per-attr warning pointing at the two supported forms — never the silent drop. Static values are byte-identical. Ternary shapes proven by real swiftc -typecheck (spacing/padding/background/cornerRadius). dynamic Icon color/size styling — the state-driven icon (<Icon color={active() ? "primary" : "muted"}>) read its color/size STATIC-only, so a dynamic value SILENTLY dropped the modifier (no .foregroundColor/.imageScale on Swift, no tint/.size on Compose, zero warnings) — the same class as the gap/padding gap above, on the Icon primitive. Now routed through the same swiftStylingValue/kotlinStylingValue ternary-of-two-literal-tokens machinery: static byte-identical, a ternary of two literal tokens → a native conditional (Swift .foregroundColor((on ? Color(…) : Color(…))) / .imageScale((on ? .large : .small)), Compose tint = (if (on) … else …) / .size((if (on) 24.dp else 16.dp))), any other dynamic value → a NAMED per-attr warning (never the silent drop). Ternary shapes proven by real swiftc -typecheck + kotlinc. dynamic Image width/height (runtime pixels) — Image dims are RAW pixels, not compile-time tokens, so UNLIKE the token props (gap/color/align) a dynamic dim isn't a "ternary of tokens" — it's a runtime numeric. Pre-fix the emit read them STATIC-only, so ANY dynamic dim (a ternary OR a signal read) SILENTLY dropped the .frame / .width modifier. Now ANY dynamic value lowers to the runtime numeric — Swift .frame(width: CGFloat(<expr>)) (SwiftUI's .frame(width:) takes CGFloat?; an Int/Double runtime value or an Int-literal ternary needs the explicit init), Compose Modifier.width((<expr>).dp) (the .dp extension applies to Int/Double) — NO warning (a pixel dim IS a runtime value, so width={size()} lowers rather than warning; the token props still warn on a fully-dynamic value because a compile-time token can't map to one). Static byte-identical. Ternary + signal shapes proven by real swiftc -typecheck + kotlinc. dynamic Stack/Layer align — cross-axis alignment lives in the container CONSTRUCTOR arg (VStack(alignment:) / Column(horizontalAlignment =) / ZStack(alignment:) / Box(contentAlignment =)), not the modifier chain, and it read STATIC-only — so a dynamic value (align={rtl() ? "end" : "start"}) SILENTLY dropped the alignment (a bare VStack { / Column {, zero warnings). Now routed through the same swiftStylingValue/kotlinStylingValue ternary-of-two-literal-tokens machinery: static byte-identical, a ternary → a native conditional INSIDE the constructor arg (Swift VStack(alignment: (rtl ? .trailing : .leading)) / ZStack(alignment: (rtl ? .topLeading : .bottomTrailing)), Compose Column(horizontalAlignment = (if (rtl) Alignment.End else Alignment.Start)) / Box(contentAlignment = (if (rtl) Alignment.TopStart else Alignment.BottomEnd))), any other dynamic value → a NAMED warning (never the silent drop). Ternary shapes proven by real swiftc -typecheck + kotlinc. dynamic Heading level — the level maps to a font/typography size; a DYNAMIC level (level={compact() ? 3 : 1}) SILENTLY DEFAULTED to level 1 (typeof levelRaw === 'number' ? … : 1) — a silent MIS-emit (the heading rendered largeTitle regardless, not a drop). The level is a compile-time token (a font-map index), so the faithful dynamic form is a ternary of two literal levels — each branch resolves to its font/style — Swift .font((compact ? .title2 : .largeTitle)).bold(), Compose style = (if (compact) MaterialTheme.typography.h6 else MaterialTheme.typography.h4); a fully-dynamic level warns NAMED + falls back to largeTitle/h4 (the map isn't runtime-indexable). Routed through the same swiftStylingValue/kotlinStylingValue machinery. Static byte-identical. Ternary shapes proven by real swiftc -typecheck + kotlinc. dynamic Heading color — a state-driven heading (color={err() ? "danger" : "text"}) read color STATIC-only, so a dynamic value SILENTLY dropped the .foregroundColor (Swift) / color = (Compose) — the same readStaticAttr class as Icon color. Now routed through swiftStylingValue/kotlinStylingValue: static byte-identical, a ternary of two literal tokens → a native conditional (Swift .foregroundColor((err ? Color(…) : Color(…))), Compose color = (if (err) Color(…) else Color(…))), any other dynamic value → a NAMED warning (never the silent drop). Ternary shapes proven by real swiftc -typecheck + kotlinc. break/continue, labeled loops, and the comma-operator handler body — break and continue statements warn-DROPPED, a SEMANTIC mis-emit (the emitted loop ran EVERY iteration where JS would exit/skip — a for…of with if (x === 3) break summed ALL elements, not the JS prefix); a LABELED loop (outer: for … break outer, the standard nested-scan idiom) dropped the WHOLE handler body ("Unsupported statement: LabeledStatement"). Both targets support all of it natively, so it now lowers faithfully: Swift outer: for … { break outer / continue outer }, Kotlin outer@ for … { break@outer / continue@outer }, plain break/continue verbatim; the switch-case fall-through strip now removes only UNLABELED breaks (a break outer inside a case exits the enclosing loop — real semantics, previously stripped). A labeled NON-loop statement warns by name. And the comma-operator ARROW body — onPress={() => (a.set(1), b.set(2))}, the compact multi-write handler — emitted a ("") junk body dropping BOTH writes; in statement position the sequence value is discarded, so each sub-expression now lowers to its own statement (arrow bodies AND block statement position; VALUE-position sequences still warn). Labeled/plain loop-control proven by real swiftc -typecheck. C-style for + do…while — completing the loop vocabulary: EVERY ForStatement and DoWhileStatement warn-dropped the WHOLE loop (the do-while residue was semantically wrong — post-loop reads saw initial values). The canonical COUNT-loop (for (let i = 0; i < n; i++), <=, or i += k with a positive literal step) now lowers to a native RANGE — Swift for i in 0..<n / 1...n / stride(from:to:by:), Kotlin for (i in 0 until n) / 1..n / step k — chosen over a while-desugar because ranges keep break/continue semantics intact (the desugar skips the update on continue → infinite loop). Non-canonical shapes (decrement, non-literal step, a counter REASSIGNED in the body — Swift's range binding is immutable, checked by an AST walk) warn by NAME with the rewrite hint. do…while maps directly: Swift repeat { } while / Kotlin do { } while ( ). Range + repeat shapes proven by real swiftc -typecheck. String(x) / Boolean(x) coercion constructors as values — String(x)'s emit was already valid on both targets, but with no inference case the RESULT typed Any, degrading any typed consumer (["v", String(n())] → [Any] → a chained .length failed Swift); it now infers string (the array stays [String]). Boolean(x) was doubly broken — no inference AND no emit mapping (the raw Boolean(n) fails BOTH targets: Swift "cannot find 'Boolean' in scope", Kotlin "unresolved reference") — and now lowers by the arg's inferred type, JS-exact for scalars: bool → identity, number → != 0, string → non-empty, optional number/string → inner-value check ((x ?? 0) != 0 — JS Boolean(undefined)=Boolean(0)=false), other optionals → presence (!= nil/!= null, matching the optional-truthiness condition lowering). An unresolvable arg type keeps the raw emit + a NAMED build-failing warning (never a silent drop). NaN edge documented (no native analog — same simplification the parseInt ?? 0 mapping makes). write-site float widening — JS has ONE number type; PMTC splits Int/Double from the declared generic + initializer, so signal<number>(0) declared Int even when every WRITE was fractional: price.set(1.5) / price.update(v => v + 0.5) emitted an Int @State/mutableStateOf(0) receiving a Double (a loud native type error on both targets). widenFloatSignals now walks the whole component IR (a generic structural walk — no node-kind enumeration, no missable shapes) for .set/.update writes against Int-typed number signals; a float-inferring written value widens the DECLARED type to Double and an integer-literal initializer emits 0.0 (Kotlin's mutableStateOf(0) carries no annotation — the initializer IS the type there); runs to fixpoint (b.set(1.5); a.set(b()) widens both). Int-written signals stay Int; a write the pass can't prove float keeps the loud native error (fail-safe, never silent truncation). Composes with the Date.now() lowering: signal<number>(0) + start.set(Date.now()) — the stopwatch shape — widens once both land. optional index a?.[i] (safe-index) — the optional COMPUTED link was chain-bailed to the "" fallback (a chained find(...)?.tags?.[0] ?? "none" even COMPILED with a semantically wrong value — always ""). JS returns undefined out-of-bounds (and nil-propagates an optional receiver), so it now lowers to the guarded idioms: Swift (a.indices.contains(i) ? a[i] : nil) — both operands named twice so both must be re-readable (scalar literals now count) + the receiver non-optional; other Swift shapes emit the nil-propagating subscript + a NAMED warning (OOB traps, the warning says so). Kotlin getOrNull(i) composes on EVERY shape (single-eval — chained receivers need no guard; optional-link receivers get ?.getOrNull via the syntactic exprHasOptionalLink check, since the type layer doesn't wrap optional-member results in a union). The optional form infers element | undefined so ?? fallback collapses (#1957); a bare a?.[0] annotates Int?. fn?.() (optional call) now lowers too (see below). Map/Set collection vocabulary — new Map/Set were warn-dropped to the "" sentinel, breaking the accumulator + dedup idioms. Now lowered end-to-end: new Map<K,V>() → [K: V]()/mutableMapOf, new Set<T>()/new Set(arr) → Set<T>()/Set(arr) · mutableSetOf/.toMutableSet(); m.set(k,v) → m[k] = v (the SIGNAL-set rewrite is now gated to ONE argument — a 2-arg .set is never a signal write); m.get(k) → m[k] (V? both targets — faithful to JS's V | undefined, infers the union so ?? fallback collapses); has/delete/add/clear/.size map per target (Swift [k] != nil/removeValue/insert/removeAll/.count; Kotlin containsKey/remove/add/clear/.size). Swift collection LOCALS force var (subscript-assign + insert are mutating; the reassignment tracker only sees =). Bare new Map() (no generics) + other new X() keep NAMED warnings. SIBLING FIND (8th instance): the dedup idiom exposed the PLAIN 1-param multi-statement callback silently dropping its body (the block-body sentinel — the indexed path was fixed in #1954, the plain path never was): Swift now emits the block; Kotlin emits return-free blocks and NAMED-warns on return-bearing bodies (labeled-return call-site wiring is the tracked follow-up — previously a SILENT drop). Kotlin labeled-return plain callbacks — a bare return inside a Kotlin lambda is prohibited (it targets the enclosing function); the labeled return@<method> form is required and only the CALL SITE knows the emitted method name. The indexed (2-param) path has had this since #1954; the PLAIN (1-param) path never did — a multi-statement predicate (filter(x => { if (cond) return false; return x > 0 })) silently dropped its body. Now wired at the plain call sites — filter/map/forEach/flatMap/find/findLast (same-name labels) + some→any / every→all (their emitted Kotlin names) — via emitKotlinPlainCallback, reusing the indexed path's labeled-body machinery with a 1-param head; expression bodies + indexed callbacks unchanged. SIBLING FIND (9th): the forEach ACCUMULATE idiom (let acc = 0; nums().forEach(x => { acc = acc + x })) exposed the mutability tracker missing reassignments inside CALLBACK arrows (it walked statement bodies, never expression trees) — the outer let acc stayed immutable ("'val' cannot be reassigned"); a generic structural walk now finds every nested arrow's statements (over-marking → a harmless never-mutated-var warning). the Swift optional-chain propagation bug (idiom-sweep batch 3, + three companions): the member emit propagated ?. down chains (a?.b.c → a?.b?.c) — correct for Kotlin (which REQUIRES it) but WRONG for Swift: after the first ?. Swift auto-propagates, and a redundant ?. on a chain-unwrapped NON-optional field is an ERROR — find(...)?.name?.length ?? 0 (the master-detail field-length shape) was a SILENT fail, and the codified p?.addr?.city spec was itself broken, masked by an emit-shape-only assertion with NO Swift compile proof (the missing-rung lesson: every codified emit needs a compile proof on BOTH targets). Swift now emits ?. only on the first optional link (genuinely-optional mid-chain fields keep it via the declared-union check); Kotlin unchanged. Companions: switch-in-computed now infers its case-return type (findFirstReturnExpr never entered switch/loop bodies → Any); num.toString() → Swift String(x); a DESTRUCTURED callback param (({name}) => …) warns NAMED (was silently filtered → unbound names; the binding prelude is a tracked follow-up). optional call f?.() (invoke-if-present) — the last optional-chaining shape to lower (member a?.b and index a?.[i] already did). Invoking a function-typed prop/field only when non-nil — props.onDone?.(), props.fmt?.(5) — used to warn-fall-back (the "index/call" explicit-guard diagnostic); it now lowers faithfully: Swift f?(args) (the optional-function field type (() -> Void)? already parenthesizes, so f?() is valid) / Kotlin f?.invoke(args) (a nullable functional type is invoked via ?.invoke — a bare f() is a type error). The callee is a MEMBER (a function-typed prop/field), so the Swift emit short-circuited in the generic member-call branch BEFORE the bare-identifier tail — that branch dropped the ? (emitting onDone(), invoking a nil closure at runtime); fixed to mirror the tail's optional lowering. Kotlin routes both callee shapes through one tail (already correct — this locks the Swift parity). Handler + computed-value shapes proven by real swiftc -typecheck + kotlinc. Closes the optional-chaining class — no optional shape (member / index / call) warns anymore. fractional/Double math — signal<number>(9.99) now emits var price: Double = 9.99, not Int; **/bitwise operators; Math.* functions; ?? optional-collapse inference — the idiomatic optional-consumption shape opt ?? fallback (nums().at(-1) ?? 0, .find(…) ?? def, data() ?? []) now infers the NON-optional result type (Int, not Int?), so consuming it (String(out()), arithmetic, a typed position) no longer fails Swift's "value of optional type 'Int?' must be unwrapped". The bug was a JS ?? used on the two inferred types — but inferType never returns null, so it ALWAYS kept the LEFT's optional branch; the fix unwraps the left's optional (unwrapOptionalType) and falls back to the fallback's type. .at() itself already lowered — the SIBLING ?? inference was the real bug (root-caused from an .at() probe). Array.from / Array.isArray — Array.from(x) → Swift Array(x) / Kotlin x.toList() (shallow copy); Array.from(x, fn) → x.map(fn) (the map form, lowered via a shared .map-rewrite so the callback element-type inference is reused); Array.isArray(x) → the literal true (a typed source IS statically an array). Both had NO native Array member — the generic emit was uncompilable (no member from / Element could not be inferred, verified). The Array.from({ length: n }, (_, i) => expr) numeric-RANGE form NOW lowers → Swift (0..<n).map { i in expr } / Kotlin (0 until n).map { i -> expr } (index-map form). The always-undefined element param can't map to a native value, so a callback that REFERENCES it (guarded by exprReferencesIdent), the 1-arg { length: n } form (no callback), and block-body callbacks stay a NAMED build-failing warning (never a silent drop); Math.* computed RETURN-TYPE — a computed RETURNING a Math.* call had no inference case, so it typed Any on Swift (private var pageCount: Any { ceil(…) }) and a downstream String(pageCount()) / arithmetic / page() < pageCount() failed ("no exact matches in call to initializer" / "cannot convert 'Any' to 'Int'"). Now inferred by JS semantics: ceil/floor/round/trunc are INTEGER-VALUED (page counts / indices) → Int, and the Swift emit wraps the Double free-function result Int(ceil(Double(x))) so it stays an Int usable in page < pageCount and prints "4" not "4.0" (inferring Double — matching the old ceil(Double(x)) emit — was a HALF-fix, Int < Double then failed); sqrt/pow + the trig/log/exp free functions → Double; abs preserves the arg's numeric type; min/max return the args' common type. inferMathCall (infer-type.ts) is the shared inference; Kotlin's derivedStateOf infers on its own (and allows Int↔Double comparison) so it needed no change — a realistic paginated DATA-TABLE now compiles end-to-end both targets. as cast + typed-empty array — TS type-operators in expression position (x as T, x satisfies T, x!) now PARSE (unwrap to the inner expression) instead of hitting the parser's unsupportedExpr fallback, which emitted the literal string "" (so [] as number[] — the idiomatic typed-empty seed for a .reduce with an ARRAY accumulator — mis-emitted as "": Swift failed loudly, Kotlin FALSE-PASSED via "" + listOf(x) coercing to a String). An EMPTY array carries no element type, so [] as T[] threads the cast's element type onto the array IR → a typed-empty emit ([Int]() / emptyList<Int>()) + inference [T], so an array-accumulator reduce (reduce((a, b) => [...a, …], [] as T[])) compiles + types correctly on both targets. array-literal spreads — [...a, ...b] / [...a, 9] / [9, ...a] / [...a, 9, ...b] (any spread count / position; merge-arrays + the add-to-list idiom) now lower to a parenthesised native concat — Swift (a + b) / (a + [9]), Kotlin (a + b) / (a + listOf(9)), with the () so a method on the literal binds to the whole concat ([...a, ...b].map { … } → (a + b).map { … }). Pre-fix only a single LEADING spread emitted, MULTI-spread wrapped the 2nd arg (a + [b] / a + listOf(b) → a type error), and nothing was parenthesised ([...a, 9].length → a + [9].count, binding .count to the tail); the add-to-list set([...items(), x]) happened to work (single leading spread, bare arg) so the bug hid until a 2nd spread or a chained method; negative-index .slice — arr.slice(-m) (last m) / arr.slice(0, -n) (drop last n) now lower to the native count-from-the-end methods (Swift suffix(m) / dropLast(n), Kotlin takeLast(m) / dropLast(n)); the existing front-counting slice (dropFirst/prefix · drop/take) explicitly bailed on a unary-minus arg so the raw .slice(-1) survived → "no member 'slice'". The COMBOS now lower too: slice(s, -n) (positive-literal start) → Swift dropFirst(s).dropLast(n) / Kotlin drop(s).dropLast(n), and slice(-m, -n) → suffix(m).dropLast(n) / takeLast(m).dropLast(n) (the native ops clamp like JS, so no bounds guard is needed). A NON-literal (variable) start with a negative end can't be proven front-anchored, so it still falls through — an honest follow-up.) 2-param array-method callbacks — arr.map((el, idx) => …) / arr.forEach((el, idx) => …) (the index form, ubiquitous for enumeration) lowered the bare 2-arg closure ({ x, i in … } / { x, i -> … }) which both targets reject (Swift "expects 1 argument, but 2 were used"; Kotlin "cannot infer type parameter R"); now lowers to the index-aware native variant — Swift enumerated().map { (idx, el) in … } / .forEach, Kotlin mapIndexed { idx, el -> … } / forEachIndexed, both index-FIRST so the params bind swapped from JS's (el, idx); 1-param callbacks fall through unchanged; the array PREDICATE methods with a 2-param index callback — .filter/.some/.every/.findIndex ((el, idx) => …) — completing the family #1934 began: a 2-param arrow is still ONE argument, so the pre-existing if (args.length === 1) 1-arg branch fired first and emitted the bare 2-param closure both compilers reject ("expects 1 argument, but 2 were used" / Kotlin "argument type mismatch"); the shared indexedArrayCallback gate is now checked BEFORE the 1-arg branch and lowers to the index-aware native form — .filter → Swift enumerated().filter{ (i,x) in … }.map{ $0.element } / Kotlin filterIndexed{ i,x -> … }; .some → Swift enumerated().contains(where:{ (i,x) in … }) / Kotlin withIndex().any{ (i,x) -> … }; .every → Swift enumerated().allSatisfy{ (i,x) in … } / Kotlin withIndex().all{ (i,x) -> … }; .findIndex → Swift enumerated().first(where:{ (i,x) in … })?.offset ?? -1 / Kotlin withIndex().firstOrNull{ (i,x) -> … }?.index ?: -1 (all index-FIRST, params swapped from JS, matching the map/forEach convention; 1-param forms unchanged); a component-level value-const (const pageSize = 2 / const steps = ["a","b","c"]) referenced from a COMPUTED or a HANDLER (Swift-only — Kotlin's const val, derivedStateOf computed and handler lambdas share the Composable body). Two sub-gaps: (a) TYPE — value-consts were never seeded into the inference ctx, so a computed referencing one inferred Any and Swift emitted private var x: Any { … }, breaking a downstream String(x()) ("no exact matches in call to initializer"); most visible for a FLOAT const (const factor = 2.5) where the binary otherwise types Int but the emit is Double(…) * (2.5) → "cannot convert Double to Int". Fixed by a persistent valueConsts map in InferenceCtx (populated in buildInferenceCtx, checked in inferType's identifier case). (b) SCOPE — a value-const emits as a body-local let in the ViewBuilder, which a private func handler AND an inline handler closure both sit OUTSIDE (if step < steps.length → "cannot find 'steps' in scope"); fixed by inlining value-consts into handler bodies (the same inline the computed path already uses). A REASSIGNED binding (let nextId = 1; nextId++) is excluded from inlining (collectMutatedComponentVars) — substituting its initial value would emit the broken (1) += 1 — so it stays body-local, read by name. A realistic WIZARD (array-const in a computed + two handlers) now compiles both targets. NOTE a realistic paginated DATA-TABLE additionally needs Math.ceil(…) to infer a numeric type instead of Any (a separate Math.*-return-type gap — deferred); JS truthiness on an OPTIONAL value in a condition — const t = todos.find(…) used as t ? a : b / {t && <X/>} / <Show when={t}> / if (t) / while (t), AND the NEGATED !t form — now lowers to an explicit t != nil / !t → t == nil (Swift) and t != null / !t → t == null (Kotlin) at every condition site (the bare optional, and !optional itself, were rejected as a non-Bool condition — a clean-parse but uncompilable silent mis-emit). The if/while STATEMENT positions lower any optional condition — inline (if (todos.find(…))), component-level (if (optionalComputed())), AND handler-LOCAL (const t = …find(…); if (t)): a handler / function body's const/let types are now seeded into the infer ctx during statement emission (seedHandlerLocals, wired into all THREE statement-body emit paths — inline handlers, named const onTap arrow decls, AND computed bodies; the computed-body path was the last hole — a MULTI-computed app emits each computed against a SHARED infer ctx that retains the LAST computed's locals, so a summary computed's const found = todos().find(...) read as unknown and found ? stayed un-lowered). A companion fix lowers the find-then-field idiom opt ? opt.prop : else to optional-chaining on BOTH targets — Swift (opt?.prop ?? else), Kotlin (opt?.prop ?: else) (optionalMemberTernary, matching any cond structurally-equal to the then-branch's member object — a bare identifier, a computed-READ selected(), or a member chain). Neither target narrows the optional in a ternary then-branch the JS way: Swift rejects opt != nil ? opt.prop : … ("value of optional type 'T?' must be unwrapped"); Kotlin smart-casts a bare-val local but NOT a selected() read (a by remember { derivedStateOf } DELEGATED property — "smart cast … impossible … delegated property"), the dominant master-detail shape (const selected = computed(() => items().find(…))). A third companion resolves the member-ternary's TYPE: member access on selected()'s T | undefined union missed the field lookup, so detailQty = computed(() => selected() ? selected().qty : 0) inferred Any and String(detailQty()) failed ("no exact matches in call to initializer") — unwrapOptionalType unwraps the union to its non-nullish branch so the field type resolves (→ Int). With all of these, BOTH a realistic TodoMVC-shape CRUD app (signal<Todo[]>, draft / nextId, find-then-field summary, index-callback labels, add / toggle / remove handlers, Field / Button) AND a realistic master-detail app (a selected computed over .find, two computed-read detail-field ternaries, a total reduce, select / bump handlers) compile end-to-end through real swiftc -typecheck AND kotlinc (the first realistic multi-feature components proven on both targets). The optional-truthiness class is now fully closed — every condition site, both forms, all binding scopes. See the supported-surface tables below for the exact set.) The per-PR validation gate is swiftc -parse (syntax-only — it does not typecheck) + kotlinc against Compose stubs, so it cannot catch this class of type-level corruption; the full real-compiler build only runs for the example apps in the advisory native-device workflow. Treat native PMTC as demo-quality (the project self-rates it 66/100): write components in the restricted declarative style, keep data-structure manipulation in pure-logic helpers, and verify on a real Simulator/Emulator before trusting native output.

Architecture overview

Pyreon's multi-platform story is built on a four-layer model. Code in lower layers is reused unchanged across platforms; code in higher layers gets per-platform implementations behind a shared API.

Layer 4: <NativeIOS> / <NativeAndroid> / <Web>  (escape hatches, opt-in)
Layer 3b: @pyreon/elements                       (web-only rich primitives)
Layer 3a: @pyreon/primitives                     (canonical multi-platform primitives)
Layer 2: useStorage / useRouter / useFetch       (ServiceBackend pattern)
Layer 1: useDebounce / useToggle / ...           (pure-logic hooks, 100% shared)
Layer 0: signal / computed / effect              (reactive core, 100% shared)

Layer 0 — Reactive core (100% shared)

signal(), computed(), effect(), batch(), onCleanup() — these are the same on every platform. PMTC maps them to @State / @Observable on iOS and mutableStateOf / derivedStateOf on Android. On web they're native Pyreon.

Layer 1 — Pure-logic hooks (100% shared)

Custom hooks composed entirely of signals + business logic. useDebounce, useToggle, usePrevious, useControllableState — no DOM, no platform APIs. They work identically on every target.

Layer 2 — Platform-abstracted services

Services with a shared API surface + per-platform implementation. Established by @pyreon/storage:

// Same code on all three platforms
import { useStorage } from '@pyreon/storage'
const todos = useStorage<Todo[]>('todos', [])

Behind the scenes:

• Web: backed by localStorage via @pyreon/storage

• iOS: backed by UserDefaults via @PyreonAppStorage (from @pyreon/native-runtime-swift)

• Android: backed by an in-memory or DataStore backend via rememberPyreonStorage (from @pyreon/native-runtime-kotlin)

The PMTC compiler rewrites useStorage<T>('key', default) to the platform-native one-liner on iOS / Android. On web it stays as the standard @pyreon/storage call.

Same pattern extends to: @pyreon/router (iOS NavigationStack + Android NavHost runtimes are Phase C), network fetching, permissions, lifecycle hooks.

Layer 3 — UI primitives (the architectural fork)

Two separate primitive layers serve different needs:

Layer 3a: @pyreon/primitives — canonical multi-platform

The cross-platform vocabulary. 15 semantic primitives designed for fundamentally the easiest DX across all three targets:

Designed for cross-platform from scratch. Semantic names (<Stack> not <View> / <VStack> / <div>). One canonical event name per concept (onPress everywhere). Tokens-first styling (padding={4} resolves via theme).

Layer 3b: @pyreon/elements — web-only rich

The existing web primitive layer (Element, Text, List, Overlay, Portal). Built on rocketstyle + styler + unistyle — rich responsive props, extendCss, full DOM-coupled styling. Stays as-is. Web-only.

Cross-platform apps use @pyreon/primitives. Web-only apps that need rocketstyle's rich features use @pyreon/elements. The two coexist — no naming collision because imports are explicit.

Layer 4 — Platform escape hatches

Two different escape hatches, for two different problems. Reach for the right one:

1. Per-platform BRANCHING — <NativeIOS> / <NativeAndroid> / <Web>. When one target needs a different arrangement of canonical primitives than another:

<NativeIOS>
  {/* Rendered on iOS; Compose + web targets render nothing here */}
  <Stack gap="md"><Text>iOS-specific layout</Text></Stack>
</NativeIOS>

Note what this is not: the children are ordinary PMTC JSX compiled through the normal canonical-primitive path, so this is a platform conditional — it does not let you write raw SwiftUI or Kotlin inline. For that, use the second hatch.

2. User-defined NATIVE MODULES — useNativeModule. When you need a platform capability the framework does not ship (Bluetooth, ARKit, a payments or analytics SDK, any vendor library). PMTC lowers the call to an instance of a class you provide, and passes member calls through verbatim:

import { defineNativeModule, useNativeModule } from '@pyreon/primitives'

type Bluetooth = {
  isSupported(): boolean
  connect(id: string): Promise<boolean>
}

// The WEB implementation — native targets never run this.
defineNativeModule<Bluetooth>('Bluetooth', {
  isSupported: () => 'bluetooth' in navigator,
  connect: async (id) => { /* Web Bluetooth */ return true },
})

function Pairing() {
  const bt = useNativeModule<Bluetooth>('Bluetooth')
  //  iOS     → @State private var bt = Bluetooth()
  //  Android → val btCtx = LocalContext.current
  //            val bt = remember { Bluetooth(btCtx) }
  return <Button onPress={() => { void bt.connect('cuff') }}>Connect</Button>
}

You supply the platform halves as ordinary app code:

// ios/Bluetooth.swift — a no-argument initialiser; @Observable if its state drives the view
final class Bluetooth {
  func isSupported() -> Bool { true }
  func connect(_ id: String) async -> Bool { /* CoreBluetooth */ true }
}

// Android — a SINGLE Context parameter (the emit injects LocalContext.current).
// Declare it in the package the generated sources use (your `--kotlin-package`),
// so the emit's unqualified reference resolves.
class Bluetooth(private val context: Context) {
  fun isSupported(): Boolean = true
  suspend fun connect(id: String): Boolean = true
}

The contract is small and checked by the platform compiler, not by PMTC: the class name matches the string, the initialiser matches the shape above, and method names/arities match what the shared source calls. A mismatch is a normal swiftc / kotlinc error in your own build. await bt.connect(id) works with no extra machinery — an awaited module method rides the same async lowering the built-in services use.

Because the module name is emitted verbatim as a native type name and PMTC resolves one file at a time, it must be a string literal at the call site (not an imported constant) and a valid identifier; anything else is a named compiler warning and the declaration is skipped rather than mis-emitted.

Why this matters. Before this hatch existed, platform services were recognised by hard-coded hook name inside the compiler, so every new capability required a framework PR — an app that needed Bluetooth simply had no path. useNativeModule is what makes the platform layer extensible by applications. Device-proven: the counter reference app renders a value produced by an app-provided DeviceInfo class on both iOS and Android.

Canonical primitive vocabulary (Layer 3a)

Layout

Content

Interaction

Input

Event model

One canonical event name per concept; the compiler maps it to the platform-native handler:

Hover events are deferred (mobile platforms don't have hover).

Handlers may be multi-statement — onPress={() => { a.set(1); b.set(2) }} emits every statement (including if blocks) into the native closure on both targets. (Earlier the compiler silently kept only the first statement.) A single-expression handler (onPress={() => a.set(1)}) keeps the compact one-line form.

Style system (v1)

Tokens-first. No raw pixels in cross-platform code.

No responsive props in v1. Web has media queries, iOS has size classes, Android has configuration changes — unifying these as per-primitive responsive props is a multi-week design problem deferred to a future arc. Apps that need responsive web layouts use @pyreon/elements directly (it has full responsive prop support). The one adaptive primitive that HAS landed is the useSizeClass() READ hook (M2.2) — a single 'compact' | 'regular' signal that lowers to iOS @Environment(\.horizontalSizeClass) / Android LocalConfiguration width / web matchMedia — so shared code can branch on width today; the size-class-driven layout primitive (Stack↔Inline) is the M2.2b follow-up.

No animation primitives in v1. Same reasoning.

Escape hatch. <NativeIOS style={...}> / <Web className="..."> for per-platform overrides when the canonical style system doesn't reach.

Inline style={{ … }} on a canonical primitive

A raw inline style — <Stack style={{ padding: 16, backgroundColor: '#2563eb', borderRadius: 8 }}> — now lowers to native modifiers on both targets (it was silently dropped before — no reader, no warning). This is the CSS-in-JS connector core: one flat CSS-in-JS object → SwiftUI .modifier() / Compose Modifier.x() from a single source, applied at the same cross-cutting seam as the token props above, so it reaches every primitive.

Colors accept #hex / #rgb / #rrggbbaa / rgb() / rgba() (alpha carried through); dimensions accept a unitless number or "Npx". The chain order is idiomatic per target (SwiftUI .padding().background().cornerRadius(); Compose .clip().background().padding() — clip-first so the fill is rounded).

Web-only declarations are stripped with no warning — cursor, userSelect, pointerEvents, transition, outline, appearance, … are a correct no-op on a touch target. Any other property is dropped with a NAMED warning (margin, boxShadow, border, flex, transform, positioning, font*, … — nothing vanishes silently), pointing at the canonical prop or a Layer-4 adapter.

Reactive style={cond ? {A} : {B}} lowers too. A ternary of two object literals lowers each shared property to a reactive conditional-value modifier — .background(active ? A : B) on iOS, .background(if (active) A else B) on Android. Because cond reads a signal (→ SwiftUI @State / Compose mutableStateOf), the style flips on state change with no extra machinery — the exact reactive mechanism the canonical-prop ternary emit already ships and the counter example device-proves (tap → re-render). This is the plan's make-or-break "dynamic resolution → reactive emit", proven on a primitive's own inline style before it rides on rocketstyle's dimension props.

• A property present in only ONE branch (asymmetric), or a padding box that isn't a single all-sides value in both branches, can't fold into one conditional value — it's emitted from the first branch statically with a named warning (never a silent drop). Give both branches the same property with two literal values for a reactive flip.

• Still out of scope (warned, not silently dropped): a non-ternary dynamic style (style={obj}, cond && {A}, nested ternary) and any non-literal field value (style={{ padding: n() }}). Reactive resolution beyond a two-literal ternary is the tracked follow-up.

Both the static and dynamic emits are toolchain-validated (swiftc -typecheck against the real SwiftUI SDK + kotlinc against the Compose stubs) — the dual-toolchain gate is what caught a Kotlin float-suffix bug (.alpha((if (c) 1 else 0.7)f) → invalid; fixed to 1f/0.7f per branch).

styled(Prim) components lower too

The connector reaches component-level styling, not just inline style. A styled() wrapping a canonical primitive —

const Card = styled(Stack)`
  background: #2563eb;
  padding: 16px;
  border-radius: 8px;
`
// <Card><Text>Hi</Text></Card>

— lowers each <Card> use-site to <Stack> with the captured CSS injected as a style, so the whole inline-style connector lowers it unchanged: VStack{…}.padding(16).background(…).cornerRadius(8) on iOS, Column(Modifier.clip(…).background(…).padding(…)) on Android. kebab-case CSS is normalized to the connector's keys (border-radius → borderRadius); use-site children/props (gap, onPress, …) are preserved.

styled(Prim) is a real styler pattern (styled's tag is string | ComponentFn) — it's this component-lowering foundation that rocketstyle's dimension resolution builds on. Scope (v1): only styled(<canonical primitive>) lowers — styled('div') / styled(NonPrimitive) warn (no native primitive); a template interpolation that is a theme token (${(p) => p.theme.color.primary}) is resolved to its value, while any other (a runtime expression) warns + drops; a use-site inline style on a styled component is a v1 gap. Both emits are toolchain-validated.

rocketstyle multi-dimensional resolution

The architecture is per-package native frontends over a shared backend: each ui-system package owns a module that lowers its constructs to a style-object IR, which the connector (the shared backend) lowers to native — mirroring how the runtimes compose. The rocketstyle-native frontend is the first non-styler one.

A rocketstyle component over a canonical-primitive base —

const Btn = rocketstyle()({ name: 'Btn', component: Stack })
  .theme(() => ({ padding: '8px 16px', borderRadius: '8px' }))
  .states({ primary: { backgroundColor: '#2563eb' }, danger: { backgroundColor: '#dc2626' } })
  .sizes({ medium: { padding: '12px' }, large: { padding: '16px' } })
// <Btn state="primary" size="large">

— resolves at compile time: at each use-site the frontend reads the state/size/variant attrs, merges base ∪ matched-dims into ONE style object (the rocketstyle cascade — dims override base), and reuses the styled rewrite (→ <Stack style={merged}> → connector). So <Btn state="primary" size="large"> → VStack{}.padding(16).background(…blue).cornerRadius(8) (size=large's padding overrode the base). This is what makes user-authored multiplatform components real: build your own on ui-system over the primitive bases and it lowers — primitives are the compiler's internal native target, not your authoring constraint.

Reactive dimension flips

A dimension prop can be dynamic — driven by a signal at runtime:

function App() {
  const active = signal(false)
  return <Btn state={active() ? 'primary' : 'danger'} size="large">…</Btn>
}

<Btn state={active() ? 'primary' : 'danger'}> resolves both branches (base ∪ static-dims ∪ each state's set) into a ternary style value handed to the connector's reactive path — so each property that differs across the branches lowers to a conditional-value modifier: SwiftUI .background((active) ? Color(…blue) : Color(…red)), Compose .background(if (active) Color(0xFF2563EB) else Color(0xFFDC2626)). Any static dimension (size="large") + the theme merge into both branches. This is what makes a native rocketstyle component reactive rather than static-only — the runtime state flip re-styles in place, no remount.

Scope: a canonical-primitive base (component: Stack); static string dimensions (the useBooleans: false default) merge into one object; one dynamic dimension (a ternary of two DECLARED dimension values) lowers to the reactive flip. Declaration values may be literals OR theme tokens (backgroundColor: t.color.primary). Remaining follow-up: ≥2 simultaneous dynamic dimensions (a switch over the dimension-set product) → warns + falls back to the first branch; a ternary whose branches aren't both declared dimension values → warns + drops. Both emits are toolchain-validated.

Dark mode

Dark mode needs no new mechanism — it composes from useColorScheme() (which lowers to SwiftUI @Environment(\.colorScheme) / Compose isSystemInDarkTheme() as a reactive "dark"/"light" string) + the reactive dimension flip above:

const Card = rocketstyle()({ name: 'Card', component: Stack })
  .states({ onLight: { backgroundColor: '#ffffff' }, onDark: { backgroundColor: '#111827' } })
function App() {
  const scheme = useColorScheme()
  return <Card state={scheme === 'dark' ? 'onDark' : 'onLight'}>…</Card>
}

→ SwiftUI .background((scheme == "dark") ? Color(…#111827) : Color(…#ffffff)) with @Environment(\.colorScheme) injected; Compose .background(if (scheme == "dark") Color(0xFF111827) else Color(0xFFFFFFFF)). The system color-scheme flip re-styles in place.

Responsive (size class)

CSS pixel breakpoints ([xs, sm, md, lg]) don't map to native, but native has a 2-bucket width class — SwiftUI @Environment(\.horizontalSizeClass) (compact/regular), Compose LocalConfiguration.current.screenWidthDp. useSizeClass() lowers both to a reactive "compact"/"regular" string, so — same composition as dark mode — a dimension flip gives a real mobile-vs-expanded responsive layout that re-flows on rotation / split-screen:

const Panel = rocketstyle()({ name: 'Panel', component: Stack })
  .states({ mobile: { padding: '12px' }, expanded: { padding: '32px' } })
function App() {
  const size = useSizeClass()
  return <Panel state={size === 'regular' ? 'expanded' : 'mobile'}>…</Panel>
}

→ SwiftUI .padding((size == "regular") ? 32 : 12); Compose .padding(if (size == "regular") 32 else 12). It's two buckets, not per-pixel breakpoints — matching how native size classes actually work.

Theme-token resolution

The mainline styler/rocketstyle value is a theme token, not a literal — background: ${(t) => t.color.primary}, .states({ primary: { backgroundColor: t.color.primary } }). The theme-native frontend resolves such a token at compile time to a concrete value the connector lowers (#hex colors → Color(.sRGB, …) / Color(0xFF…); spacing / radius → numbers).

Resolution is against your app's own theme, declared with defineTheme:

const theme = defineTheme({
  color:   { primary: '#ff3b30', danger: '#dc2626' },
  spacing: { md: 20, lg: 24 },
  radius:  { sm: 6 },
})

const Card = styled(Stack)`
  background: ${(t) => t.color.primary};   /* → your #ff3b30, not a guess */
  padding: ${(t) => t.spacing.md};          /* → 20 */
`

defineTheme({ … }) is a compile-time declaration — the compiler parses its literal tokens and drops the declaration from the native output (there is no native defineTheme; the runtime helper is identity on web). The parsed theme is merged over the bundled defaults per entry, so overriding only color.primary keeps every other default token, and a zero-config app (no defineTheme) resolves standard tokens against the defaults (which mirror @pyreon/ui-theme + the primitive defaults). Both the styler (p) => p.theme.color.primary (props) and rocketstyle (t) => t.color.primary (theme-directly) shapes resolve; group aliases (colors/space/borderRadius), flat and nested paths are accepted.

Scope (v1): the color / spacing / radius groups with literal leaf values (a native theme must be static — a runtime-computed token can't be baked). An unknown token (t.color.doesNotExist) or a non-token interpolation warns + drops.

Device-proven (first for the styling track). The gated counter app carries arocketstyle()({ component: Text }) badge whose state flips with a signal; the iOS XCUITest and the Android Compose test both assert it renders and re-renders on the flip. Note precisely what that covers: neither harness can read a COLOUR, and a missing colour still compiles — so colour presence is locked by the emit test (native-text-reactive-color-parity), while the device tests cover what a unit test cannot, that the lowering survives into a real app and reacts. Building the Android app additionally proves the ComposeText(color = …) constructor arg is type-correct.

Building ui-system components — coverage

You build your own components on the ui-system styling frontends (styled + rocketstyle + theme tokens) — over the canonical primitives or over @pyreon/elements' Element (the base the 67 @pyreon/ui-components use). <Element> maps to the canonical <Stack> (direction/alignX/alignY/gap translated), so a rocketstyle()({ component: Element }) component — the ui-components authoring pattern — lowers to SwiftUI and Compose. (PMTC compiles your source; an imported pre-built @pyreon/ui-components component doesn't lower — re-author the pattern.) The Element / PyreonUI / Container / Row / Col alias lowerings are import-source-aware: a tag is intercepted only when it is imported from its expected @pyreon package (Element←@pyreon/elements, Container/Row/Col←@pyreon/coolgrid, PyreonUI←@pyreon/ui-core), so a same-named user component (import { Row } from './my-components') is never mis-lowered as a coolgrid Row. All examples are real swiftc/kotlinc-validated end-to-end:

• Interactive — rocketstyle()({ component: Button }) with theme tokens + a reactive state={sig() ? 'a' : 'b'} flip + disabled + onPress + size.

• Container — rocketstyle()({ component: Stack }) (or Element) with theme tokens + dark mode (state={scheme === 'dark' ? 'onDark' : 'onLight'}).

The ❌ rows are architectural boundaries, not tracked bugs — they're bound to the DOM / CSSOM / a pointer model that doesn't exist natively. The ✅ rows are what makes user-authored ui-system components run on iOS + Android + web from one source.

Per-platform import resolution

The DX-critical question: how does import { Stack } from '@pyreon/primitives' resolve on each target?

• Web: @pyreon/primitives is a real npm package with real implementations. Stack is a ComponentFn that renders DOM. Standard module resolution.

• iOS / Android (via PMTC): The PMTC compiler INTERCEPTS JSX with <Stack> etc. at compile time and emits platform-native code BEFORE the runtime is involved. The import is type-anchor only — the JSX never calls into @pyreon/primitives's runtime.

The same source file works on all three targets. The compiler-side handling for each target is different but the developer doesn't see it.

Migration

@pyreon/primitives is a NEW package. Adding it breaks nothing.

Existing PMTC source using SwiftUI-flavored names (<VStack>, <HStack>, <TextField>) continues to work via the existing per-target emit. The TodoMVC migration to canonical vocabulary is Phase E — a deliberate, additive port. After migration is proven, deprecation warnings land on SwiftUI-flavored tags. Removal happens in a major-version bump LATER.

Current state + roadmap

The 5-phase implementation roadmap:

Foundation rollout (A–E):

ONE TodoApp.tsx source → THREE example apps (web, iOS, Android), all typecheck-clean.

Beyond the foundation — toward production-grade

The vocabulary is multiplatform; the road to shipping real production apps continues:

Loader auto-emit — v1 landed (zero-param, expression-body). A route's loader: () => <expr> now compiles to a runtimePyreonRouteLoader host wrapping the route's component: its .task(SwiftUI) / LaunchedEffect (Compose) fires the loader ONCE on the route's appear and stores the result via router.setLoaderData(<active path>, …), where the already-shipped useLoaderData<T>() reads it (keyed by the runtime active path, so it matches for both literal and:param routes; the home route keys by its literal path = currentPathat launch). The store is guarded (loaderData[path] == nil) so re-renders never re-run it. Realistic native loader bodies are signal / store reads (() => globalCache()) and sync expressions(() => buildInitialData()); both emit correctly.

Still deferred (each leaves loader undefined + WARNS, so the route renders with no loader rather than mis-emitting): a param-usingloader ((ctx) => fetchUser(ctx.params.id) — ctx has no value source in the load closure yet), a block-body loader (() => { … } — needs statement emit), and a truly-async body (async () => await fetch(…) — no await lowering / typed-decode generic, the useFetch<T>territory). ctx.params threading + async-loader decode are the next arc; until then, fetch-style loaders populate loaderData from native code (or via a useFetch container) as before.

Production capability matrix (weighted, rung-labeled)

This table is the denominator for every "N% of usages" claim. Real mobile/tablet apps decompose into the weighted usage categories below (weights = editorial judgment of a typical app's surface, documented so they can be argued with). Each row states its verification rung —R1 emit · R2 swiftc/kotlinc typecheck · R3 unit/spec ·R4 local simulator/emulator build+run · R5 the nightly device gate (XCUITest + Compose connectedCheck) — and the fraction of the category with an ASSERTED behavior at R4+ (exercised-but-never-asserted counts as 0; strictness is the point). Coverage numbers are recomputed whenever a row changes; do not edit the totals without recomputing.

Weighted totals (2026-07-08 baseline; M2.3 + M3.1 + M3.2 + M3.2b + M3.3 + M2.2 + a11y-label + M2.2b + M2.4-i18n + M2.5-colorScheme + M2.6-machine + M2.7-animations + M2.8-transitiongroup + M3.5-biometrics applied): device-proven (R4+) coverage≈ 57% (64.5 / 113 — +1.5 for the Forms row (0.5 → 0.75): useFieldArray built end-to-end and device-proven on both platforms (runtimes + lowering + keyed-For integration; the accessor-unwrap seam makes a paren-keeping emit uncompilable), on top of +1.75 for the Storage row (0.45 → 0.7): useSecureStorage device-proven on both platforms (iOS Keychain terminate+relaunch survival; Android cold-instance decrypt + encrypted-at-rest — closing a sub-capability that was three-quarters missing: warn-drop emit, no real Android backend, no web half), on top of +2.0 for the Core-UI row closure (0.8 → 1.0, now COMPLETE: all 15 canonical primitives behaviourally device-asserted, Layer/Spacer/Heading by geometry on both platforms): the four canonical primitives this table had named as gaps (Modal/Toggle/Scroll/Link) are now device-asserted, and TWO of them were blocked by emit bugs that only a real simulator could surface — <Modal> anchored .sheet to an EmptyView() host so it never presented on iOS, and <Link> dropped data-testid before it became .accessibilityIdentifier, making the element unqueryable and therefore unassertable. Both emitted valid, typechecking Swift the whole time, which is precisely the class R1–R3 cannot see. Prior increments follow: — the Platform-APIs row's +1.0 each for haptics/share/link/notifs, +1.0 for useSizeClass() (M2.2), +0.45 for the accessibilityLabel device assertion, +1.0 for the adaptive Stack↔Inline layout (M2.2b), +0.9 for the createI18n translation device assertion (M2.4), +1.2 for the <Transition show> animation device assertion (M2.7), +0.6 for the <TransitionGroup> animated-list device assertion (M2.8), plus +1.0 for the useBiometrics() async-lowering device assertion (M3.5) (the Platform-APIs row moved 0.4 → 0.5 as the M4.5 await hook.method() lowering — the keystone for the whole async-platform-API tier — was proven to RUN on-device, not just compile, via a biometric gate whose deterministic denied path flips an observable text; the biometric SUCCESS path + Android real BiometricPrompt stay follow-ups) and +1.0 for the useImagePicker() device assertion (M3.4) (the row moved 0.5 → 0.6 as the first CAMERA/photo-library capability landed with a behavioural iOS round trip and a real — not scaffolded — Android launcher emit) and +1.0 for the useFilePicker() device assertion (M3.8): the row moves 0.6 → 0.7 as the files/documents capability lands (the document sibling — UIDocumentPicker / SAF OpenDocument), same behavioural iOS round trip + real Android launcher emit, leaving the pick-a-real-file path, the Android round trip, and file SAVING (M3.8b) as disclosed follow-ups. useColorScheme (M2.5) + createMachine (M2.6) landed device-proofs that deepened already-recorded rows so neither moved the fraction. The denominator grew, and the headline fell: ≈53% → ≈51%. A Styling & design system row (weight 6) was added because it had been missing entirely — the whole styled / elements / coolgrid / attrs / rocketstyle / theme-token surface lowers to both targets and is documented as supported, yet had no row and so sat outside the denominator. Its fraction is 0.2: the rocketstyle-over-Text pattern with a reactive dimension flip IS device-asserted on both platforms at the RE-RENDER level (the counter's StatusBadge flips Badge:ok → Badge:warn; the rendered COLOUR is not asserted — the harnesses cannot read it, and those tests are explicit about that), while styled() / Element / coolgrid / attrs / defineTheme tokens / typography have no native example at all. (Drafted as 0.0 and corrected on review — the "nothing is proven" reading was true of the branch it was written on and false of main, where the badge had already landed. Overstating a gap is as wrong as hiding one.) Omitting a track a real app leans on heavily made every percentage here flattering rather than true, and this table is supposed to BE the denominator. The drop is a correction to the measurement, not a regression in the product. Before that, Then +1.05 for useDatabase (Storage 0.3 → 0.45) — the first device assertion for structured storage, and the one that required THREE emitter/runtime fixes before the test could be written at all: get/delete/find had no argument labels (never compiled), db.insert({ id, fields }) lowered the record to an anonymous tuple (so the only WRITE never compiled either), and the default backend was in-memory (so nothing survived a relaunch). "No gated app renders FROM the database" turned out to have a cause rather than being an absence of effort. Before that, +0.5 for useAuth (Auth 0.3 → 0.4), from the first REAL-APP device gate — examples/native-finance-ios, which asserts a COMPOSED flow (useForm validation → the useAuth transition → a store-backed route guard → a computed balance → a keyed <For> mutation) rather than one hook at a time. That gate immediately earned its place by catching a locale-formatting bug four demo apps could not: Text("\(balance)") selects SwiftUI's LocalizedStringKey overload, which formats numbers per locale, so a 2700 balance rendered "2 700" and disagreed with the web and Android builds of the same source — invisible below 1000, hence invisible to Count: 0); plus +0.9 for the useGeolocation device assertion (Maps/geolocation 0.0 → 0.3 — the row's first non-zero score; the fraction stays low because useMap has no DEVICE test at all and the Android half is compile-only, so 0.3 buys iOS geolocation and nothing else. useMap's web half now exists and its moveTo/removeMarker now compile on Swift, which moves it from unbuildable to unproven-on-device — that is a real change but NOT a fraction change, because this number measures device proof. The headline does not move off ≈52% at this precision, which is the honest reading: one hook's device proof is a small share of a 113-point denominator). compile-proven (R2+) upper bound ≈ 75% — i.e. roughly three-quarters of the weighted surface already exists and typechecks, but only about a third is proven to behave on a device.The production goal is 70–90% at R4+; the gap between the two numbers is, precisely, the roadmap: assert what exists (storage, auth, services), then build what doesn't (platform APIs, animations, adaptive layout, gestures beyond tap). The self-rating moves only on device evidence — this table is where that evidence is ledgered.

Per-hook device-behavior audit (M1.2)

The service-hook layer at per-hook granularity — same rung vocabulary, same strictness (a hook that runs inside a green device test but has no assertion on its behavior earns nothing). "In device app" names which of the nightly-built apps (todomvc / counter / router-demo / tasks, plusfinance — the first REAL app in the gate rather than a demo) uses the hook at all.

Assertion queue (the follow-up PR order; each moves a row to R5 and re-scores the matrix). useStorage (todomvc terminate+relaunch) anduseAuth (the finance sign-in flow) are now done; remaining:useDatabase behaviour is now done too (iOS — the counter's terminate+relaunch record assertion). Remaining: 1. useDatabase on ANDROID, 2. useWebSocket, 3. useClipboard. The three manual-.start() hooks (geolocation/push/payments) are gated on M3.9 auto-start parity first.

useDatabase deserves a note. Two things blocked it, and the second was worse than "unproven":

1. Its emitted Swift did not compile — the get/delete/find calls were missing the argument labels the runtime declares (see the type gate).

2. It did not persist. Both runtimes' PyreonDatabase defaulted to an in-memory backend, and the emit constructed exactly that default — so an app that inserted records and relaunched found them gone, with no warning, no error, and nothing failing. The entire reasonuseDatabase exists over useStorage is structured data that outlives the process, so an ephemeral default was not a conservative starting point; it was silent data loss wearing the word "default".

Both are now fixed. FileDatabaseBackend (one JSON file per collection, atomic writes, Application Support on iOS / filesDir on Android) is the default on both platforms, and the Kotlin emit threadsLocalContext.current into the constructor because Android cannot resolve app-private storage without a Context. The spelling is deliberately asymmetric — Swift's no-arg initialiser is the persistent one, Kotlin has no no-arg form at all — and the on-disk bytes are identical, locked by a cross-language format test that asserts the same string from Swift'sJSONSerialization and Kotlin's hand-written codec.

Deliberately Foundation/JVM-only, no SQLite: a record is an id plus string fields, and a SQLite module map differs between Apple platforms and Linux — the exact toolchain split that has broken this runtime's CI before. Apps that outgrow the file store inject Room / SQLDelight / Core Data through the same constructor.

A third defect surfaced while writing the device test, and it explains the other two: db.insert(collection, { id, fields }) lowered the record to an anonymous tuple, so the only way to WRITE to the store did not compile. With no writes possible, nothing downstream was reachable — "no gated app renders FROM the database" had a cause. Fixed in both backends.

With all three closed, the capability is device-proven on iOS: the counter taps Save Note, asserts the rendered count advanced, then terminates and relaunches and asserts the record survived. Still owed: the same assertion on Android (its emit and persistence are unit-proven, but nothing has run it on an emulator).

M2.3 — gestures (long-press) SHIPPED. <Press onLongPress={fn}> now lowers on native (the type + web 500ms-polyfill already existed; only the native emit was missing). Swift uses a SIMULTANEOUS LongPressGesture, not .onLongPressGesture — a bare long-press modifier on a Buttondoes NOT fire (the button's tap recognizer swallows it; found on a real Simulator, invisible to swiftc). Android uses combinedClickable( onClick, onLongClick). Device-proven: examples/native-counter-ios's XCUITest holds a <Press> reset zone >=0.5s and asserts the counter resets (R4 local pass); the Android longClick() sibling is proven by the device run. First gesture beyond tap. Deferred: onSwipe / drag (M2.3b).

Native routing

createRouter({ routes }) compiles to native dispatch — SwiftUINavigationStack + .navigationDestination(for:) on iOS, a Composewhen (router.currentPath) block on Android. One route table, both targets.

const router = createRouter({
  routes: [
    { path: '/',            component: Home },
    { path: '/users/:id',   component: User },          // path param
    { path: '/files/:rest*', component: Files },         // splat / catch-all
    { path: '/old',         redirect: '/users/1' },      // redirect (alias)
    { path: '/admin',       component: Admin, beforeEnter: () => isAuthed() }, // guard
    { path: '/app',         component: AppLayout, children: [   // nested layout
      { path: 'dashboard',  component: Dashboard },
      { path: 'settings',   component: Settings },
    ] },
    { path: '*',            component: NotFound },        // wildcard 404
  ],
  beforeEach: [requireAuth],                              // global guards run before every nav
  afterEach: [logAnalytics],                              // global hooks fire after every nav
})
return <RouterProvider router={router}><RouterView /></RouterProvider>

Inside a route component, read path params via destructuring:

function User() {
  const { id } = useParams<{ id: string }>()   // → id reads the active route's param
  return <Text>{id}</Text>
}

Path matching (mirrors @pyreon/router's match.ts, verified by the native router runtime's own swift test / kotlinc smoke):

Leading/trailing slashes are tolerated (/about/ matches /about).

Redirects are compile-time aliases: { path: '/old', redirect: '/new' }makes the /old dispatch branch render /new's component directly (no runtime push). Chains (/a → /b → /c) resolve transitively; cyclic / dangling redirects are dropped to the no-match fallback.

Wildcard 404: a * / (.*) route's component becomes the dispatchelse-branch — the canonical not-found page for any unmatched path.

Guards (beforeEnter: () => <boolExpr>) wrap the matched component in an inline conditional checked at navigation time; on failure the branch renders the wildcard catch-all (if present) or a denial placeholder.

Nested routes (children: [...]) compile to a flattened full-path dispatch where each leaf is wrapped in its layout chain via a content slot: a layout component (a route parent) is emitted with a@ViewBuilder content closure (SwiftUI) / content: @Composable () -> Unit(Compose), and its <RouterView /> becomes that slot. So /app/dashboardrenders AppLayout { Dashboard() }; the layout's own /app index rendersAppLayout { EmptyView() }. Three-plus levels nest outermost-first (AppLayout { TeamLayout { Members() } }). Flat route tables keep the original dispatch unchanged.

useParams() destructuring — const { id } = useParams() (and{ id: userId } aliasing) binds each field to the active router's param map: a computed private var id: String { useParams(router:)["id"] ?? "" }on SwiftUI (computed, not stored — it reads @Environment), val id = useParams()["id"] ?: "" on Compose.

Typed params prop — a route component may instead declareprops: { params: { id: string } } (the web router's prop-injection shape). PMTC synthesizes a named type per component — UserPage →struct UserPageParam: Codable (SwiftUI) / data class UserPageParam(Compose) — and the dispatcher constructs it from the matched path segments: UserPage(params: UserPageParam(id: params["id"] ?? "")) /UserPage(params = UserPageParam(id = params["id"] ?: "")). number /boolean fields coerce from the string segments with safe defaults (Int(...) ?? 0, == "true"). If the params shape structurally matches a struct you declared yourself (type RouteParams = { id: string }), your name is reused instead of synthesizing. Components without aparams prop are dispatched with no arguments.

Loader data — PyreonRouter exposes a loaderData store +useLoaderData<T>(); a route's loaded data is keyed by the active path and read back, typed, by the current route. A route'sloader: () => <expr> is now auto-emitted (v1): the compiler wraps the route in a runtime PyreonRouteLoader host that fires the loader once on appear (.task / LaunchedEffect) and calls setLoaderData, souseLoaderData<T>() reads it with no manual wiring. v1 covers zero-param, expression-body loaders (signal/store reads + sync expressions); param-using, block-body, and truly-async loaders WARN and emit unloaded — see the loader-auto-emit note in the roadmap.

Global guards (beforeEach / afterEach) — pass arrays of identifier-referenced guard/hook functions on the createRouter({ ... })config. The parser extracts the identifiers (inline arrow bodies + non- array forms are silently dropped — a documented follow-up); the emit configures the router via a Swift closure-init / Kotlin apply { } block. At runtime, push / replace wrap the navigation in the guard chain — any guard returning false blocks the navigation, then every afterEachhook fires after a successful commit:

const requireAuth = (path: string) => isAuthed() || path === '/login'
const logAnalytics = (path: string) => trackPageView(path)

const router = createRouter({
  routes,
  beforeEach: [requireAuth],   // any → false blocks the nav
  afterEach: [logAnalytics],   // all fire after successful commit
})

Falls back to bare init when no guards are configured (back-compat — existing apps need no changes).

Throw-redirect pattern (router.redirect(path)) — the native equivalent of web's throw redirect("/login") from a loader/guard, without the guard-return-type redesign. Inside a beforeEach,router.redirect(path) queues a replace AND returns false-equivalent short-circuit semantics; an internal _inGuard re-entry flag prevents the redirect's own navigation from infinite-recursing through the same guard chain:

router.beforeEachGuards.append { path in
    if !isAuthed() && path != "/login" {
        router.redirect("/login")  // queues replace, re-entry-safe
        return false               // blocks the original push
    }
    return true
}

Same shape on Kotlin (router.beforeEachGuards.add { path -> … }). The runtime addition is ~30 LOC per target; no compiler changes.

Status: path matching, redirects, wildcard 404, per-route guards,nested routes, useParams destructuring, the loader-data runtime, global beforeEach / afterEach guards (#1108), therouter.redirect() throw-pattern (#1109), and the typedparams prop (synthesized per-component struct/data class + dispatcher construction from the matched segments) are all landed. Loader auto-emit and a typed useParams<T>() hook generic are planned.

Bundled images — the asset pipeline

One assets/ directory next to your shared src/ carries the app's images; the pyreon-native assets build step materializes it per target:

All 15 primitives compile + render on a REAL Android build, not just the kotlinc-validate subset: the emit's androidx symbols that live outside the star-imported packages (Color, RoundedCornerShape,verticalScroll/rememberScrollState for <Scroll>, Dialog for<Modal>, Coil's AsyncImage for remote <Image>) each get a content-keyed conditional import — the kotlinc stubs would otherwise MASK a missing import (green validate, red gradle assembleDebug).

The src dispatch is canonical across targets: http(s)://… is remote (AsyncImage/Coil/<img>), a BARE name (logo.png) is a bundled asset, and a path-style src (/img/x.png) is web-only — the compiler warns and native falls through to the remote emit (visible failure, never silent). fit maps toscaledToFill/scaledToFit (SwiftUI) and ContentScale.Crop/Fit/ FillBounds/None (Compose); the web default cover holds everywhere.

Asset-name collisions after Android sanitization (my-logo.png vsmy_logo.png → both my_logo) abort the build loudly.

Icons — the canonical name map

<Icon name="star"> uses ONE semantic name everywhere: iOS maps it to an SF Symbol (Image(systemName: "star.fill")), Android to a COMPILE-TIME Material reference (Icons.Filled.Star — hosts need only the small material-icons-core artifact, never -extended), and web to the app sprite's symbol id. The curated ~37-glyph map lives incanonical-primitives.ts (ICON_MAP: navigation, actions, status). An UNMAPPED name warns at compile time and stays visible: iOS passes it through raw (direct SF ids keep working), Android renders thewarning placeholder glyph — never a silent blank.

Custom fonts

Drop .ttf/.otf files in the same assets/ dir; the assets step copies them per target (iOS bundle + UIAppFonts; Android res/font; web public/fonts). <Text font="Brand"> / <Heading font="Brand">renders the bundled family. The load-bearing detail iOS gets wrong by default: Font.custom needs the font's POSTSCRIPT NAME (its internalname-table id), NOT the filename — a filename-keyed Font.customsilently falls back to the system font on-device. The CLI reads the PostScript name from the sfnt table (no dependency) and bakes it into the emit, so <Text font="Brand"> → Font.custom("Trattatello", …)even when the file is Brand.ttf. Android resolves res/font at runtime via pyreonFont(name) (a missing font throws loudly).

Native data & services

Data hooks compile to native via per-service runtime ports behind the shared TS API (the PyreonStorage pattern — each service has a Swift + Kotlin runtime the emitted code drives):

• Platform prerequisites for networked apps (both device-CI findings): Android needs <uses-permission android:name="android.permission.INTERNET" /> in the manifest — without it socket creation fails with the opaqueSocketException: socket failed: EPERM — plus a network-security-config exception if the endpoint is plain http (scope it to loopback/dev hosts only). iOS needs an ATS exception for non-HTTPS endpoints (NSAllowsLocalNetworking for loopback/dev). The create-multiplatform scaffold ships the INTERNET permission by default.

• useFetch<T>('/url') → a PyreonFetch<T> reactive container ({ data, error, isPending, refetch }). The compiler emits a mount-time.task { } (SwiftUI) / LaunchedEffect (Compose) that runs the request through the container's begin → resolve | reject state machine and decodes into T. Field reads (x.data, x.isPending) are @Observableproperties on iOS, Compose MutableState on Android.

• useForm → a PyreonForm container (per-field values / errors / touched + submit state). const form = useForm({ initialValues }) emits@State PyreonForm(initialValues:[...]) (SwiftUI) / remember { PyreonForm(mapOf(...)) } (Compose); MutableState field reads append.value on Compose (except the derived isValid getter).v2 (form-binding arc) — device-proven. useForm({ initialValues, validators, onSubmit }) lowers fully: per-field validators emit as native closures ('' = valid), <Field value={form.values.x}> binds through the runtime (form.binding("x") on SwiftUI — a realBinding<String> whose setter re-validates after an error; a value/onValueChange pair through setValue on Compose), per-field dict access subscripts with typed defaults (form.errors.x →form.errors["x"] ?? ""), and submit() gates on validateAllbefore invoking onSubmit. The web-parity names (setFieldValue,handleSubmit) exist on both runtime ports. SwiftUI nuance handled by the emit: an onSubmit capturing instance members (navigate, store writes) attaches via .onAppear { form.onSubmit = … } — a @State property initializer runs before self exists. The tasks showcase's login is the canonical validated form; its device smokes assert the ERROR path before the happy path. Open: block-body + async validators, schema validation (@pyreon/validationreachability), <Form>/<Submit> wrappers.

• usePermissions → a PyreonPermissions container (RBACcan/cannot/all/any with "x.*" wildcards). const can = usePermissions([...]) seeds the grant set; reads are method calls (no.value rewrite).

• useOnline → a PyreonNetworkStatus container with a reactiveisOnline flag (real NWPathMonitor on iOS; the Compose side takes the app's connectivity callback). Because the WEB useOnline() returns an ACCESSOR (() => boolean), ONE shared source reads it as net() — the emit lowers that accessor call to the container's isOnline read (net.isOnline on SwiftUI, net.isOnline.value on Compose). The directnet.isOnline member read also compiles on native (a native-only shape with no web equivalent), so both idioms work — prefer net() for cross-platform parity.

• useAppState → a PyreonAppState container with a reactive phaseString ("active" | "inactive" | "background"), driven byUIApplication lifecycle notifications on iOS and an app-injectedProcessLifecycleOwner source on Android (the same injected-source shapePyreonNetworkStatus/PyreonStorage use to stay Android-SDK-free for the kotlinc stub gate). The WEB useAppState() returns an ACCESSOR (() => 'active' | 'inactive' | 'background'), so ONE shared source reads it as state() — the emit lowers that accessor call to state.phase (SwiftUI@Observable) / state.phase.value (Compose MutableState). The directstate.phase member read also works. Use it to pause a live poll while backgrounded or dim UI while inactive — identical source on all three targets.

• useClipboard → a PyreonClipboard container with a copy(text)method + a reactive copied: Bool flag that auto-resets to false ~2s after each copy (matches the web @pyreon/hooks contract). WrapsUIPasteboard.general.string on iOS (cross-platform UIKit/AppKit — #1096 split out the macOS NSPasteboard path so the Swift runtime builds on both Apple platforms) and the system ClipboardManager on Android. Reads are plain method calls + a plain Bool/Boolean field — no .value rewrite. Kotlin emit is a two-line shape — val cbCtx = LocalContext.current hoisted out of the remember { … } lambda (the lambda is non-Composable; LocalContext.current can't be read inside it) + val cb = remember { PyreonClipboard(cbCtx) }. The Swift container's deinit now cancels the in-flight reset Task (#1107 — Class I leak fix) so a view that disappears mid-copy doesn't leak a pending 2-second timer. BOTH the single-binding shape const cb = useClipboard() AND the destructure form const { copy, copied } = useClipboard() compile (PR1 — destructure lowers to a synthetic single-binding container + per-field aliases; see "Binding idioms").

• useColorScheme() → returns "light" | "dark" reactively from the platform's preferred-color-scheme channel. No runtime port needed — both SwiftUI (@Environment(\.colorScheme)) and Compose (isSystemInDarkTheme()) ship the primitive directly, so PMTC emit is a thin per-target wrapper: Swift injects @Environment(\.colorScheme) private var pyreonColorScheme on the View struct + a computed private var <name>: String { pyreonColorScheme == .dark ? "dark" : "light" }; Kotlin emits val <name> = if (isSystemInDarkTheme()) "dark" else "light" inline. Same "light" | "dark" string contract the web hook uses — scheme === 'dark' works identically across all three targets (#1103).

• Data/services hooks (Phase 5 — #1689): seven more @pyreon/* hooks now compile to their runtime containers, instantiated + read exactly likeuseOnline (Swift @State private var x = PyreonX() with bare@Observable reads; Kotlin val x = remember { PyreonX() } with .valueon the MutableState fields, bare on Bool getters + methods):useAuth<User>() → PyreonAuth<User> (status / user / error +isAuthenticated); useDatabase() → PyreonDatabase (insert / get / all / find / delete / count); useGeolocation() → PyreonGeolocation (lat / lon / accuracy / isAuthorized); useMap() → PyreonMapState (camera / markers / selectedMarkerId); useWebSocket('wss://…') → PyreonWebSocket(lastMessage / messages / isConnected; URL must be a string literal);usePush() → PyreonPushNotifications (token / lastNotification / isAuthorized); usePayments() → PyreonPayments (products / ownedProductIds / purchasing). This unblocks writing the finance, analytical, maps, and realtime/uber archetypes from one .tsx — verified at the compile rung: an archetype component using all seven emits typecheck-clean Swift (swiftc) and Kotlin (kotlinc).Two honest limits:(1) WebSocket lifecycle auto-start is EMITTED — useWebSocket(url)now auto-connects on mount on BOTH targets, matching the web hook: the compiler synthesizes a mount-time ws.connect() (Swift .onAppear on the stable host / Kotlin LaunchedEffect(Unit)), url-threaded to the faithfulconnect(to: URL(string: …)!) (Swift) / connect("wss://…") via the@pyreon/native-runtime-kotlin OkHttp transport extension (fun PyreonWebSocket.connect(url: String), Kotlin). An EXPLICITonMount(() => ws.connect()) is respected — the auto-connect is skipped when the component already calls .connect(), so there's no double-connect. Both auto-connect shapes are proven by real swiftc -typecheck + kotlinc.geolocation.start() / push.start() still LOWER only via the EXPLICITonMount(() => …) escape hatch (their Kotlin containers need an app-injected source — FusedLocationProvider — that the compiler can't synthesize, so their zero-call auto-start stays a host-wiring follow-up). (2) useSecureStorage() now lowers on both targets (the v1 warn-drop is closed): Swift PyreonSecureStorage() (Keychain default), KotlinPyreonSecureStorage(ctx) (AndroidKeyStore AES-GCM viaKeystoreSecureBackend — the stated blocker, an app-injected backend requirement, was resolved the same way FileDatabaseBackend was). Calls emit KEY-FIRST with Swift labels (write(key:value:)), so a crossed positional call is uncompilable rather than a silent wrong-key write. As with every native service, the runtime behavior is compile-verified, not device-proven — the nightly device gate + example apps are the runtime-proof layer.

Status: useForm (v2 — validated forms, device-proven via the tasks showcase's error-path smoke), useFetch (device-proven — the networked Quotes screen), usePermissions,useOnline, useClipboard, and useColorScheme are landed(runtime port + compiler emit — useColorScheme is emit-only because the platform primitive is enough). useFetch's open item is a device-scope NETWORK proof (the UITest gates don't run a backend yet). useValidation reachability planned.

The supported TypeScript surface

PMTC compiles a deliberate SUBSET of TypeScript — the shapes the canonical examples exercise, enumerated here so you know where the boundary is BEFORE the compiler tells you. Outside the subset, the contract is: a warning naming the construct + either a conservative passthrough (the native compiler then errors loudly at the site) or a whole-decl bail — never silent misbehavior. pyreon-native build prints every warning; treat any warning as "this construct is outside v1."

Declarations (component body)

Binding idioms — one requirement, one convenience that now lowers:

• Hook results → single-binding OR destructure (both work now). Bind the hook's result to one name and read fields off it (const q = useFetch<T>(url); q.data() / q.isPending), OR destructure it directly (const { data, isPending } = useFetch<T>(url)). The destructure formlowers (PR1) to a synthetic single-binding container (const __pyHookN = useFetch<T>(url)) + one field alias per key, so each local rewrites to __pyHookN.<field> at its use sites — producingbyte-identical native output to the single-binding form on both targets, with the call form preserved (accessor data() vs plainisPending). Covers useFetch / useForm / useClipboard /useStorage / usePermissions / useOnline / useColorScheme /useNetworkStatus / the seven Phase-5 data/services hooks. useParamskeeps its own per-key lowering (router section). Still warn-drop: a rest element (const { data, ...rest } = …) or a nested pattern (const { user: { id } } = …) — v1 lowers all-simple destructures only; and useLoaderData's destructure (its read returns an opaque T with no field shape to alias).

• Destructured function/arrow params → lowered. A helper with a destructured param — const dist = ({ x, y }: Point): number => x + y,const apply = ({ id }: T) => { remove(id) } — synthesizes a positional param __pN (typed from the pattern's annotation — a named type resolves to the declared struct) + prepends let x = __pN.x per key, so the body references x/y as written. Works for void handlers and functions with an explicit return-type annotation. A value-returning destructured function WITHOUT a return annotation should annotate it (({ x }: P): number => x) — an unannotated value return infers Unit on Kotlin (a separate, pre-existing return-inference limit, not specific to destructuring). Rest / nested patterns warn + stay un-destructured.

• Store reads → inline OR aliased (both work). Read store state inline through the hook (useApp().store.tasks()) OR bind the hook to a local first (const app = useApp(); app.store.tasks()) — the aliaslowers to a useApp() call at every use site, producing byte-identical native output to the inline form. (Aliasing previously failed the native build with Unresolved reference 'app'; it now compiles.)

• Static attrs → literal OR module-level const. A native-mapped static attribute (<Image src=…>, <WebView src=… /> / html=, font, background, …) accepts an inline string literal OR a module-levelconst string/number/boolean binding referenced by name:

  const CHART_URL = "https://x.example/c.png"
  // both emit AsyncImage(url:) / Coil AsyncImage(model=) identically:
  <Image src="https://x.example/c.png" alt="chart" />
  <Image src={CHART_URL} alt="chart" />

  Copy

  A let (mutable) binding, a non-literal init (const x = f()), or a component-scope / unknown identifier is NOT resolved — it falls through to the normal "needs static" emit path. (Component-scope const + transitive const B = A resolution are tracked follow-ups.)

Expressions

Types

Statements (function/computed bodies): const/let (incl.multi-declarator const a = 1, b = 2 → split into one decl each),return, if/else, for…of / while / switch (→ native for in /while / switch·when), and reassignment — t = t + x, += -= *= /= %= (a reassigned local is emitted var, not let/val, automatically), so an imperative loop body can accumulate. C-style for (let i = 0; …; …) is NOT in v1 — use for…of over an array or while. Array destructuring of locals (const [a, b] = xs) is NOT in v1 — the binding is silently dropped, leaving later references undefined; index explicitly (xs()[0]) instead. (Object destructuring const {a, b} = obj DOES lower — body-local, component-param, and hook-result forms, e.g. const { data } = useFetch(url).)<For> remains the idiom for RENDERING a list; these loops are for in-body data work.

JSX: the 15 canonical primitives, <For each by>, <Show when>,<Suspense fallback>, <ErrorBoundary fallback>, <KeepAlive when>,<Transition show>, <Modal open>, <RouterProvider>/<RouterView>/<Link>. data-testid flows to accessibilityIdentifier / testTag(containers gain the queryability semantic automatically). The cross-platform a11y vocabulary on @pyreon/primitives lowers the same way on every target — write the neutral prop once:accessibilityLabel="…" → web aria-label / SwiftUI.accessibilityLabel(…) (the VoiceOver name) / Compose.semantics { contentDescription = … } (the TalkBack name); andaccessibilityHidden → web aria-hidden="true" / SwiftUI.accessibilityHidden(true) / Compose .clearAndSetSemantics { }(clears the node + subtree from the a11y tree — clearAndSetSemantics is stable in the targeted Compose 1.7 BOM, vs the experimentalinvisibleToUser()); and accessibilityRole ("button" / "image" /"header" — the values that map 1/* zero-content: unhandled mdast node "textDirective" */ to every target's role model) → webrole="button"/"img"/"heading" / SwiftUI.accessibilityAddTraits(.isButton / .isImage / .isHeader) / Compose.semantics { role = Role.Button / Role.Image } (and heading() for headers). Component children must be JSX or value expressions (auto-wrapped in Text).

Module scope: let/const primitives (non-reactive on native), type aliases, the recognized factory calls. Module-scope signal() is NOT lowered — declare signals inside components or stores.

Consuming compiler diagnostics

The parser warnings introduced by Round-1 (#1094 — Icon / Image /Link missing required props) and Round-2 (#1099 — Press withoutonPress, native Link prefetch={…}, Stack/Inline/Layer align="<typo>") flow through the same result.warnings channel as every other parse warning. Read them programmatically from the compiler:

import { transform } from '@pyreon/native-compiler'

const { code, warnings } = transform(source, { target: 'swift' })
for (const w of warnings) console.warn(w)

The shipped surface today is the pyreon-native build CLI, which aggregates warnings per file and prints them to stderr as[pyreon-native] N warning(s): after each build. There is no Vite-plugin / LSP / editor-diagnostic surfacer yet — that's an explicit Phase 6 DX follow-up. The package is @pyreon/native-compiler(private / workspace-only); consumers using transform() directly are the path until a public published API lands.

WebView host — embedding web-only-rich viz (charts / flow)

Some libraries are structurally web-only — @pyreon/charts (ECharts),@pyreon/flow (elkjs), @pyreon/code (CodeMirror), @pyreon/document(pdfmake) all wrap a browser-runtime engine and cannot compile to SwiftUI / Compose. The multiplatform answer is a hybrid: a substantial native shell (the canonical primitives) with the heavy viz hosted in a<WebView> — WKWebView on iOS, Android WebView, an <iframe> on web.

<NativeIOS><WebView html={CHART_HTML} /></NativeIOS>
<NativeAndroid><WebView html={CHART_HTML} /></NativeAndroid>
<Web>{/* render the chart inline — it's already web */}</Web>

<WebView> takes html (inline page — loadHTMLString / srcdoc) ORsrc (a LOCAL bundled asset — preferred, policy-safe — or a remote URL). For App Store / Play Store review, prefer bundled local assets so the viz is app content, not remote code.

The two-way data bridge

A hosted chart isn't a static screenshot — it stays live in BOTH directions, over a unified JS API the runtime wires per platform:

Forward — native → page (data). Pass a signal; the runtime JSON-encodes it and PUSHES it into the already-loaded page aswindow.__pyreonData, firing a pyreondata event, WITHOUT reloading — so the chart updates in place (no flicker, zoom/animation preserved). Adata-only change never reloads; only an html/src change does.

const metrics = signal<Metric[]>([…])
<WebView html={CHART_HTML} data={metrics()} />

// In the hosted page:
function render() { const d = window.__pyreonData || []; /* draw d */ }
window.addEventListener('pyreondata', render); render()

Reverse — page → native (onMessage). The page callswindow.pyreonPostMessage("payload"); the string is delivered to the native onMessage callback, so a tap inside the chart drives a native signal. (iOS WKScriptMessageHandler; Android a main-thread-marshalled@JavascriptInterface; web the parent defines window.pyreonPostMessageon the iframe.) The payload is a plain string — JSON-stringify structured data and parse it in the handler.

const selected = signal('')
<WebView html={CHART_HTML} data={metrics()} onMessage={(m) => selected.set(m)} />
<Text>Selected: {selected()}</Text>

// In the hosted page — a tapped bar reports back:
bar.addEventListener('click', () => window.pyreonPostMessage(region))

Together these make a webview-hosted chart a first-class interactive member of the native app: live native data flows in, user events flow back out. examples/native-analytics is the canonical end-to-end proof (native data table + aggregation + an interactive WebView chart from ONE.tsx). Web caveat: both bridges are same-origin / srcdoc only — a cross-origin remote src can't be reached from the parent frame (the native targets reach remote content via evaluateJavaScript / the script handler).

DX surfaces on native (honest scope)

The "one source" promise extends to WRITING the source, not just shipping it. Pyreon ships several developer-experience surfaces; which of them work on the native targets is a structural question — some are pre-emit (source-level) and target-agnostic, others depend on the Pyreon runtime that PMTC erases when emitting Swift/Kotlin.

Works on native source (✅ — same DX as web)

These analyze your .tsx source BEFORE PMTC emits anything, so they are target-agnostic by construction.

• Reactivity Lens (analyzeReactivity from @pyreon/compiler). Returns the same structural reactivity facts (reactive /reactive-prop / static-text / hoisted-static) and footgun findings (props-destructured, signal-write-as-call, …) on a PMTC source file as it does on a web-only source. Verified end-to- end against a <Stack>/<Button>/<Text> Counter fixture: the Lens correctly flags const { x } = props as footgun and the signal reads inside {count()} as reactive, identical to the output it produces for the same shape in a web component.

• @pyreon/lint rules + pyreon doctor. Every rule runs on the source AST; none of them load the runtime. pyreon/no-window-in- ssr, pyreon/signal-write-as-call, pyreon/props-destructured,pyreon/no-iterate-children-without-resolve, the islands audit, the SSG audit, the test-environment audit — all surface the same findings on a PMTC source file. The pyreon/no-window-in-ssrrule is actually MORE valuable on native sources (the emit target literally has no window), but the surface is the same.

• Static type checking + audit-types. tsc --noEmit and the typed-but-unimplemented gate care only about TypeScript types, so they work identically across targets.

• MCP tools (validate, get_api, get_pattern,get_anti_patterns, get_changelog, audit_test_environment,audit_islands). All operate on source / repo metadata, not the runtime. An AI agent driving a native source through validategets the same anti-pattern catalog as it would for a web file.

• pyreon-native check editor-ready diagnostics. The fast authoring-loop command exposes an in-memory checkSource(code, fileName, opts) core — no disk read, so it checks an unsavededitor buffer, the case a linter plugin needs — and attaches aposition ({ line, column }) to transform + type-check-error findings, parsed from the file:line:col the toolchains embed. Socheck --json is consumable by an editor linter integration (null-ls, a generic-linter extension) or a CI annotation matcher, with precise squiggles on the errors that carry a location. Honest scope:unsupported-subset WARNINGS are still position-less (rendered as afile-level diagnostic — threading source spans through the ~110 compiler warn sites is a tracked follow-up).

• pyreon-native check --lsp — a stdio LSP server. Publishes those findings as live editor diagnostics on open / change (atextDocument/didOpen / didChange → publishDiagnostics loop over the in-memory checkSource core), so the authoring loop no longer needs a manual CLI run — squiggles appear as you type in any LSP-speaking editor (VS Code, Neovim, …). The JSON-RPC framing is hand-rolled, mirroring @pyreon/lint --lsp (no vscode-languageserverdependency); the pure core (finding→diagnostic mapper + message handler) is unit-tested without the transport, and the end-to-end server is proven by spawning it and round-tripping a realinitialize + didOpen. Honest v1 scope: diagnostics only (no inlay hints — native has no reactivity-lens surface to project), a synchronous re-check on change (debounce is a follow-up — the in-memory transform is fast), and no swiftc -typecheck on the keystroke path (slow / macOS-only). Warnings stay file-level until the compiler warn sites carry spans.

Web-only by structural design (❌ — not coming to native)

These surfaces depend on the Pyreon RUNTIME (signal registry, effect graph, devtools hook). PMTC erases that runtime when it emits to SwiftUI @State / Compose mutableStateOf — there is no Pyreon-side data structure to introspect on a native target; SwiftUI's _GraphInputs and Compose's SlotTable own the reactive graph end-to-end. This is structural-infeasibility, not engineering effort.

• LPIH (Live Program Inlay Hints — fire counts / re-run counters at the source line). Requires the dev-mode@pyreon/reactivity registry (activateReactiveDevtools +getFireSummaries) to be alive in the running app. On native builds the entire reactivity package is tree-shaken — thesignal(0) call you wrote is emitted as @State var count = 0, there is no Pyreon-side wrapper to count fires. Use on web during development; the inlay hints don't reach a running iOS / Android build, by design.

• Devtools panel (the Chrome extension underpackages/tools/devtools). Connects towindow.__PYREON_DEVTOOLS__ (a hook attached by@pyreon/runtime-dom's installDevTools()) to walk the component tree, highlight nodes, watch signals fire. On a native build there is no window, no __PYREON_DEVTOOLS__, and no Pyreon component tree — SwiftUI and Compose own the view hierarchy. For native runtime debugging use Xcode's View Hierarchy Debugger (iOS) and Android Studio's Layout Inspector (Android); they're the native equivalents of the Pyreon devtools panel and they work on the emitted view tree directly.

• Pyreon HMR + @pyreon/vite-plugin signal-preserving HMR. Web-only by construction (Vite is a web dev server). iOS uses Xcode's incremental compile + Simulator hot-reload; Android uses Gradle's incremental build + Compose's LiveLiterals /recomposeHighlighter. These are platform-native HMR equivalents — there is no shared Pyreon HMR surface across targets.

Partial — works for the source-level part, runtime part is on the platform

• pyreon-native build warnings (the silent-drop diagnostic surface from PRs #1235 / #1441). Pre-emit warnings about droppeduseLoaderData() reads, dropped <Suspense fallback> props, etc. ARE shown — they're emitted at compile time, surfaced viatransform() result.warnings and the CLI's[pyreon-native] N warning(s) stderr aggregation. The actual runtime-state debugging is per-target (Xcode + Android Studio above).

If you adopt PMTC for a real production app, the practical workflow is: write + debug source-level concerns on web (Lens, devtools, HMR, lint) where the iteration loop is fastest; verify + debug native-runtime concerns on the device with the platform's own tooling. Same .tsx, two debugging surfaces.

Verifiable today (compile contract)

• Web: @pyreon/runtime-dom renders any Pyreon JSX. Full ecosystem available.

• iOS: pyreon-native build --target=ios --source=./src --out=./generated produces typecheck-clean Swift (verified via swiftc -parse in the native-validate CI). The opt-in native-device workflow additionally runs xcodegen + xcodebuild to compile the full example app on a real Xcode/Simulator SDK, then xcodebuild test boots the iPhone 15 Simulator + runs PyreonTodoMVCUITests to assert accessibilityIdentifier("todo-app") renders within 30s.

• Android: pyreon-native build --target=android --source=./src --out=./generated produces typecheck-clean Kotlin (verified via kotlinc + Compose stubs). The same opt-in native-device workflow runs gradle assembleDebug against the real Android toolchain, then boots a Pixel-6 emulator (API 33, google_apis, x86_64, via reactivecircus/android-emulator-runner) + runs gradle connectedCheck which executes TodoAppInstrumentedTest's composeRule.onNodeWithTag("todo-app").assertIsDisplayed().

TodoMVC reference walkthrough (locally verified, June 2026)

The examples/native-todomvc-{web,ios,android} apps form the canonical proof of the single-source contract. The shared TodoApp source (examples/native-todomvc-ios/src/TodoApp.tsx) renders on all three targets without modification.

Web (a real running app in the browser):

cd examples/native-todomvc-web
bun run build      # 88 modules → 35 KB JS bundle, 13 KB gzipped
bun run dev        # http://localhost:5173/

Then in a browser: type, hit Enter, toggle, filter All/Active/Completed, click Clear completed. Zero console errors. Web fully working.

iOS Swift emit:

bash examples/native-todomvc-ios/scripts/build.sh
# → examples/native-todomvc-ios/generated/TodoApp.swift

The emitted file opens with the import preamble (import SwiftUI / PyreonRuntime / PyreonRouter) and emits idiomatic SwiftUI: @PyreonAppStorage("pyreon-todomvc:todos") for persistence, @State for local signals, VStack(spacing: 8) / HStack for layout, TextField(..., text: $draft) with .onSubmit { addTodo() }, ForEach keyed by id, Button(action:). The data-testid="todo-app" JSX attribute becomes .accessibilityIdentifier("todo-app") so the same string works on the iOS UI test.

Verify it compiles against the real SwiftUI SDK:

swiftc -typecheck \
  -target arm64-apple-macos14.0 \
  packages/native/runtime-swift/Sources/PyreonRuntime/*.swift \
  packages/native/router-swift/Sources/PyreonRouter/*.swift \
  examples/native-todomvc-ios/generated/TodoApp.swift
# → exit 0 (zero errors)

Android Kotlin emit:

bash examples/native-todomvc-android/scripts/build.sh
# → examples/native-todomvc-android/app/src/main/kotlin/com/pyreon/generated/TodoApp.kt

The emitted file opens with package com.pyreon.generated, the Compose import preamble (androidx.compose.runtime.* / material.* / kotlinx.serialization.Serializable / com.pyreon.runtime.*), and emits idiomatic Compose: var todos by rememberPyreonStorage<List<Todo>>(...), var filter by remember { mutableStateOf(Filter.all) }, val visible by remember { derivedStateOf { ... } }, Column(verticalArrangement = Arrangement.spacedBy(8.dp), modifier = Modifier.testTag("todo-app")), TextField with KeyboardOptions(imeAction = ImeAction.Done) + KeyboardActions(onDone = { addTodo() }), LazyColumn { items(visible, key = { it.id }) { ... } }, Button(onClick = ...).

Verify against the framework's validateKotlin (same Compose stub set the validate-kotlin.test.ts gate uses):

bun -e "
  import('./packages/native/compiler/src/validate.ts').then(async (m) => {
    const { readFileSync } = await import('node:fs')
    const src = readFileSync('examples/native-todomvc-android/app/src/main/kotlin/com/pyreon/generated/TodoApp.kt', 'utf8')
    // Strip the package + wildcard imports (the stub set is in default package).
    const stripped = src.split('\n')
      .filter(l => !l.startsWith('package ') && !l.startsWith('import androidx') && !l.startsWith('import kotlinx') && !l.startsWith('import com.pyreon'))
      .join('\n')
    console.log(JSON.stringify(m.validateKotlin(stripped), null, 2))
  })
"
# → { "ok": true }

One source. Three targets. Verified locally on macOS 14 with Xcode 15 + JDK 21 + Kotlin 2.x.

The runtime packages exist, with one reactive container per data/service hook:

• @pyreon/native-runtime-swift — @PyreonAppStorage + PyreonStorage, PyreonFetch<T>, PyreonForm, PyreonPermissions, PyreonNetworkStatus (@Observable containers)

• @pyreon/native-runtime-kotlin — rememberPyreonStorage + the same PyreonFetch / PyreonForm / PyreonPermissions / PyreonNetworkStatus / PyreonClipboard containers (Compose MutableState); PR #1104 closed the last untested service by adding the Kotlin PyreonClipboard test suite, bringing every container to parity test coverage

• @pyreon/native-router-{swift,kotlin} — PyreonRouter (path stack, matchPath, params, loaderData) + useNavigate / useParams / useLoaderData hooks

Reference

• Compiler source: packages/native/compiler/src/ — emit-swift.ts / emit-kotlin.ts per-target emit; canonical-primitives.ts shared name maps + token resolution

• Native runtime packages: packages/native/runtime-swift/, packages/native/runtime-kotlin/

• Web runtime: packages/core/primitives/src/web/ — all 15 canonical primitives

• Example apps: examples/native-todomvc-{ios,android,web}/ + examples/native-router-demo-{ios,web}/ — native-router-demo-ios ships a full XcodeGen host shell (#1105) so bash scripts/build.sh produces a buildable Xcode project, not a source-only stub. examples/native-todomvc-web/README.md was also corrected (#1106) so it no longer references a fictional src/TodoApp.tsx — the one-source contract (Phase E3) keeps the shared TodoApp source in examples/native-todomvc-ios/src/.

• Real-device build gate: .github/workflows/native-device.yml (opt-in via the native-device label / dispatch)

• CLAUDE.md "PMTC Multi-Target Architecture" section — agent-context summary of the layered model + roadmap