jz 0.5.1 → 0.6.0

package/module/math.js

@@ -2,9 +2,9 @@
  * Math module - Math.sin, Math.cos, Math.sqrt, Math.PI, etc.
  *
  * Module API:
- * - ctx.core.emit['math.X'] = emitter(deps, args => WasmNode) - emitters with declarative stdlib deps
- * - ctx.core.stdlib['math.X'] = '(func ...)' - WAT function definitions
- * - ctx.core.stdlibDeps['math.X'] = ['dep'] - direct stdlib→stdlib edges (expanded transitively)
+ * - reg('math.X', deps, args => WasmNode) — emit handler + declarative stdlib deps
+ * - wat('math.X', `(func …)`) — WAT stdlib via bridge
+ * - deps({ 'math.X': ['dep'] }) - WAT stdlib→stdlib deps (expanded transitively)
  *
  * Prepare resolves Math.sin(x) → ['()', 'math.sin', x]
  * Compile looks up ctx.core.emit['math.sin'] and calls it.
@@ -13,17 +13,26 @@
  */
 
 import { typed, asF64, asI32, toI32, toNumF64, temp, arrayLoop, isLit, litVal, isPureIR } from '../src/ir.js'
-import { emit } from '../src/emit.js'
-import { emitter } from '../src/ctx.js'
-import { repOf } from '../src/analyze.js'
+import { emit, emitter, reg, deps, dual, tag, wat, hostImport } from '../src/bridge.js'
+import { inc, declGlobal } from '../src/ctx.js'
+import { repOf } from '../src/reps.js'
 
 export default (ctx) => {
-  // Direct stdlib→stdlib call edges. resolveIncludes() expands these transitively,
-  // so each emitter declares only the single stdlib it directly calls.
-  Object.assign(ctx.core.stdlibDeps, {
-    'math.sin': ['math.isFinite'],
-    'math.cos': ['math.isFinite'],
+  // Math.random seeding. DEFAULT: entropy-seeded once from the host on first use (crypto under
+  // host:'js', `random_get` under WASI), so randomness "just works" and isn't silently reproducible.
+  // `randomSeed: <n>` picks a fixed seed for a reproducible sequence; `true` forces entropy explicitly.
+  // Either way jz emits the randomness syscall only when `Math.random` is actually used.
+  const rngEntropy = ctx.transform.randomSeed === undefined || ctx.transform.randomSeed === true
+  const rngSeedConst = typeof ctx.transform.randomSeed === 'number'
+    ? ((ctx.transform.randomSeed >>> 0) || 1)   // xorshift dies on 0 → floor at 1
+    : 12345
+  deps({
+    'math.sin': ['math.sin_core'],
+    'math.cos': ['math.cos_core'],
+    'math.sin_core': [],
+    'math.cos_core': [],
     'math.tan': ['math.sin', 'math.cos'],
+    'math.exp': ['math.exp2'],
     'math.expm1': ['math.exp'],
     'math.log2': ['math.log'],
     'math.log1p': ['math.log'],
@@ -40,7 +49,6 @@ export default (ctx) => {
     'math.cbrt': ['math.isFinite', 'math.pow'],
     'math.sumPrecise': ['__ptr_offset', '__len', '__alloc'],
   })
-
   // Helpers: all math ops take f64 and return f64. Args go through ToNumber
   // (toNumF64) — ECMA Math methods coerce each argument, so null→0, undefined→NaN.
   const f = (op, a) => typed([op, toNumF64(a, emit(a))], 'f64')
@@ -56,13 +64,72 @@ export default (ctx) => {
     }
     return false
   }
+  // Emit-time fast path: call $math.sin_core/$math.cos_core directly instead of
+  // the $math.sin/$math.cos wrappers. sin_core still runs the isFinite guard
+  // (finite operands can overflow to ±Inf); this only saves a call indirection.
+  const isBoundName = (name) =>
+    typeof name === 'string' && (
+      repOf(name) != null ||
+      ctx.func.locals?.has(name) ||
+      ctx.func.current?.params?.some(p => p.name === name)
+    )
+  const isSinCoreFastPath = (src) => {
+    if (src == null) return false
+    if (typeof src === 'number') return Number.isFinite(src)
+    if (typeof src === 'string') return isIntCertain(src) || isBoundName(src)
+    if (!Array.isArray(src)) return false
+    const op = src[0]
+    // jz source literals: [null, n], [, bool], [null, null]
+    if (op == null && src.length === 2) {
+      const v = src[1]
+      if (typeof v === 'number') return Number.isFinite(v)
+      if (typeof v === 'boolean') return true
+      if (v == null) return true
+    }
+    if (op === 'literal') {
+      const v = src[1]
+      return typeof v === 'number' && Number.isFinite(v)
+    }
+    if (op === 'global' || op === 'param') return true
+    if (op === '()') {
+      const fn = src[1]
+      if (fn === 'math.PI' || fn === 'math.E' || fn === 'math.LN2' || fn === 'math.SQRT2') return true
+      if (typeof fn === 'string' && fn.startsWith('math.')) {
+        const finiteOps = new Set([
+          'math.abs', 'math.floor', 'math.ceil', 'math.trunc', 'math.round', 'math.sqrt',
+          'math.sin', 'math.cos', 'math.tan', 'math.imul', 'math.clz32',
+        ])
+        if (finiteOps.has(fn)) return src.slice(2).every(isSinCoreFastPath)
+      }
+    }
+    if (op === '|' && src.length === 3) return isSinCoreFastPath(src[1]) && isSinCoreFastPath(src[2])
+    if (op === '%' || op === '&' || op === '^' || op === '<<' || op === '>>' || op === '>>>') {
+      return src.slice(1).every(isSinCoreFastPath)
+    }
+    if (op === '+' || op === '-' || op === '*' || op === '**') {
+      return src.slice(1).every(isSinCoreFastPath)
+    }
+    if (op === '/') return isSinCoreFastPath(src[1]) && isSinCoreFastPath(src[2])
+    if (op === '?:' && src.length === 4) return isSinCoreFastPath(src[2]) && isSinCoreFastPath(src[3])
+    if ((op === '&&' || op === '||') && src.length === 3) return isSinCoreFastPath(src[1]) && isSinCoreFastPath(src[2])
+    if (op === '!' && src.length === 2) return isSinCoreFastPath(src[1])
+    if (op === '[]' && src.length === 3) {
+      // Chord tables / semitone indices — index is int-shaped, element is a small int.
+      return isSinCoreFastPath(src[2]) || isIntCertain(src[2])
+    }
+    if (op === '.' && src.length === 3 && src[2] === 'length') {
+      return typeof src[1] === 'string' || isSinCoreFastPath(src[1])
+    }
+    if (op === '.' && src.length === 3) return isIntCertain(src)
+    return false
+  }
   const fInt = (op, a) => isIntCertain(a) ? asF64(emit(a)) : f(op, a)
   // ECMA Math methods perform ToNumber on each argument. toNumF64 short-circuits
   // for known-number nodes, and routes everything else through __to_num so null→0,
   // undefined→NaN, and strings get parsed. Without this, raw NaN-boxed pointers
   // (null/undefined/strings) would propagate through math.log etc. and surface
   // as the original null/undefined sentinel after decode.
-  const call = (name, ...args) => typed(['call', `$${name}`, ...args.map(a => toNumF64(a, emit(a)))], 'f64')
+  const fn = (name, ...args) => typed(['call', `$${name}`, ...args.map(a => toNumF64(a, emit(a)))], 'f64')
 
   // Canonicalize a possibly-NaN f64 result. A wasm arithmetic op that mints a
   // fresh NaN (f64.sqrt of a negative, f64.min/max with a NaN operand) leaves
@@ -80,6 +147,25 @@ export default (ctx) => {
         ['f64.ne', ['local.get', `$${t}`], ['local.get', `$${t}`]]]], 'f64')
   }
 
+  // sqrt(x) needs no NaN-canon when its argument is provably ≥ 0 with no spurious NaN:
+  // the result is then a normal non-negative f64 (or +0 / +inf, or a propagated input
+  // NaN that is already canonical), never a freshly-minted ±NaN. A sum of pure squares
+  // — the vector-length idiom sqrt(x*x + y*y + …) — is exactly that: every term is ≥ +0
+  // (even ±0·±0 = +0, ±inf² = +inf), the sum never cancels to inf−inf, and any NaN can
+  // only come from a NaN input (already the canonical number-NaN). So a distance /
+  // normalize loop sheds the per-sqrt select + f64.ne + local on its critical path.
+  // `pure` excludes call/load/tee, so two structurally-equal operands of a `*` really
+  // are the same value (a genuine square), not two side-effecting evaluations.
+  const pureF64 = (n) => !Array.isArray(n) ? true :
+    (n[0] === 'local.get' || n[0] === 'global.get' || n[0] === 'f64.const' || n[0] === 'i32.const' || n[0] === 'i64.const') ? true :
+    (n[0] === 'f64.add' || n[0] === 'f64.sub' || n[0] === 'f64.mul' || n[0] === 'f64.div' || n[0] === 'f64.neg' || n[0] === 'f64.abs' || n[0] === 'f64.convert_i32_s' || n[0] === 'f64.convert_i32_u') ? n.slice(1).every(pureF64) : false
+  const sameIR = (a, b) => Array.isArray(a) !== Array.isArray(b) ? false : !Array.isArray(a) ? a === b : (a.length === b.length && a.every((x, i) => sameIR(x, b[i])))
+  const nonNegF64 = (n) => !Array.isArray(n) ? false :
+    n[0] === 'f64.mul' ? (pureF64(n[1]) && sameIR(n[1], n[2])) :     // x·x ≥ 0
+    n[0] === 'f64.add' ? (nonNegF64(n[1]) && nonNegF64(n[2])) :      // (≥0) + (≥0) ≥ 0
+    n[0] === 'f64.const' ? (typeof n[1] === 'number' && n[1] >= 0) : false
+  const sqrtIR = (a) => { const ir = f('f64.sqrt', a); return nonNegF64(ir[1]) ? ir : canon(ir) }
+
   // Constants
   ctx.core.emit['math.PI'] = () => typed(['f64.const', Math.PI], 'f64')
   ctx.core.emit['math.E'] = () => typed(['f64.const', Math.E], 'f64')
@@ -105,7 +191,7 @@ export default (ctx) => {
   // Built-in WASM ops. sqrt/min/max mint a fresh NaN (sqrt of a negative, min/max
   // with a NaN operand) whose sign is platform-nondeterministic — `canon` folds it
   // back to the canonical pattern. abs/floor/ceil/trunc never produce a new NaN.
-  ctx.core.emit['math.sqrt'] = a => canon(f('f64.sqrt', a))
+  ctx.core.emit['math.sqrt'] = a => sqrtIR(a)
   ctx.core.emit['math.abs'] = a => f('f64.abs', a)
   ctx.core.emit['math.floor'] = a => fInt('f64.floor', a)
   ctx.core.emit['math.ceil'] = a => fInt('f64.ceil', a)
@@ -151,36 +237,42 @@ export default (ctx) => {
   ctx.core.emit['math.fround'] = a => typed(['f64.promote_f32', ['f32.demote_f64', toNumF64(a, emit(a))]], 'f64')
 
   // Sign
-  ctx.core.emit['math.sign'] = emitter(['math.sign'], a => call('math.sign', a))
+  reg('math.sign', ['math.sign'], a => fn('math.sign', a))
 
-  // Trig
-  ctx.core.emit['math.sin'] = emitter(['math.sin'], a => call('math.sin', a))
-  ctx.core.emit['math.cos'] = emitter(['math.cos'], a => call('math.cos', a))
-  ctx.core.emit['math.tan'] = emitter(['math.tan'], a => call('math.tan', a))
+  // Trig — isSinCoreFastPath skips the $math.sin/$math.cos wrapper call.
+  ctx.core.emit['math.sin'] = dual(
+    emitter(['math.sin'], a => fn('math.sin', a)),
+    emitter(['math.sin_core'], a => fn('math.sin_core', a)),
+    isSinCoreFastPath)
+  ctx.core.emit['math.cos'] = dual(
+    emitter(['math.cos'], a => fn('math.cos', a)),
+    emitter(['math.cos_core'], a => fn('math.cos_core', a)),
+    isSinCoreFastPath)
+  reg('math.tan', ['math.tan'], a => fn('math.tan', a))
 
   // Inverse trig
-  ctx.core.emit['math.asin'] = emitter(['math.asin'], a => call('math.asin', a))
-  ctx.core.emit['math.acos'] = emitter(['math.acos'], a => call('math.acos', a))
-  ctx.core.emit['math.atan'] = emitter(['math.atan'], a => call('math.atan', a))
-  ctx.core.emit['math.atan2'] = emitter(['math.atan2'], (a, b) => call('math.atan2', a, b))
+  reg('math.asin', ['math.asin'], a => fn('math.asin', a))
+  reg('math.acos', ['math.acos'], a => fn('math.acos', a))
+  reg('math.atan', ['math.atan'], a => fn('math.atan', a))
+  reg('math.atan2', ['math.atan2'], (a, b) => fn('math.atan2', a, b))
 
   // Hyperbolic
-  ctx.core.emit['math.sinh'] = emitter(['math.sinh'], a => call('math.sinh', a))
-  ctx.core.emit['math.cosh'] = emitter(['math.cosh'], a => call('math.cosh', a))
-  ctx.core.emit['math.tanh'] = emitter(['math.tanh'], a => call('math.tanh', a))
+  reg('math.sinh', ['math.sinh'], a => fn('math.sinh', a))
+  reg('math.cosh', ['math.cosh'], a => fn('math.cosh', a))
+  reg('math.tanh', ['math.tanh'], a => fn('math.tanh', a))
 
   // Inverse hyperbolic
-  ctx.core.emit['math.asinh'] = emitter(['math.asinh'], a => call('math.asinh', a))
-  ctx.core.emit['math.acosh'] = emitter(['math.acosh'], a => call('math.acosh', a))
-  ctx.core.emit['math.atanh'] = emitter(['math.atanh'], a => call('math.atanh', a))
+  reg('math.asinh', ['math.asinh'], a => fn('math.asinh', a))
+  reg('math.acosh', ['math.acosh'], a => fn('math.acosh', a))
+  reg('math.atanh', ['math.atanh'], a => fn('math.atanh', a))
 
   // Exponential and logarithmic
-  ctx.core.emit['math.exp'] = emitter(['math.exp'], a => call('math.exp', a))
-  ctx.core.emit['math.expm1'] = emitter(['math.expm1'], a => call('math.expm1', a))
-  ctx.core.emit['math.log'] = emitter(['math.log'], a => call('math.log', a))
-  ctx.core.emit['math.log2'] = emitter(['math.log2'], a => call('math.log2', a))
-  ctx.core.emit['math.log10'] = emitter(['math.log10'], a => call('math.log10', a))
-  ctx.core.emit['math.log1p'] = emitter(['math.log1p'], a => call('math.log1p', a))
+  reg('math.exp', ['math.exp'], a => fn('math.exp', a))
+  reg('math.expm1', ['math.expm1'], a => fn('math.expm1', a))
+  reg('math.log', ['math.log'], a => fn('math.log', a))
+  reg('math.log2', ['math.log2'], a => fn('math.log2', a))
+  reg('math.log10', ['math.log10'], a => fn('math.log10', a))
+  reg('math.log1p', ['math.log1p'], a => fn('math.log1p', a))
 
   // Power. Constant-integer-exponent `Math.pow(x,n)` / `x ** n` (|n| ≤ POW_FOLD_MAX)
   // lower to inline square-and-multiply instead of a $math.pow call. The fold is
@@ -190,7 +282,7 @@ export default (ctx) => {
   // both the exact bits and the result sign (negative iff x<0 ∧ n odd, which is exactly
   // its `neg_base`). A program whose only pow use is folded then never pulls the
   // math.pow/exp/log stdlib. `**`'s exponent is parsed as a bare number (incl. negatives).
-  const POW_FOLD_MAX = 8
+  const POW_FOLD_MAX = 16
   const get = name => ['local.get', `$${name}`]
   const constInt = b => {
     const v = typeof b === 'number' ? b
@@ -226,18 +318,52 @@ export default (ctx) => {
     const inner = typed(['block', ['result', 'f64'], ...stmts, result], 'f64')
     return (minted && !neverNaN(a, baseIR)) ? canon(inner) : inner
   }
-  const powCall = emitter(['math.pow'], (a, b) => call('math.pow', a, b))
-  ctx.core.emit['math.pow'] = (a, b) => {
+  const constNum = b => typeof b === 'number' ? b
+    : (Array.isArray(b) && b.length === 2 && b[0] == null && typeof b[1] === 'number') ? b[1]
+    : null
+  // `x ** 0.5` folds to f64.sqrt instead of the exp/log $math.pow call — saves the
+  // whole pow/exp/log stdlib (the headline `dist` example drops from ~1.0kB to 70B)
+  // and runs at hardware-sqrt speed. f64.sqrt is correctly-rounded, so for every
+  // normal input it is bit-identical to V8's `Math.pow(x, 0.5)`, and it agrees with
+  // jz's own `Math.sqrt(x)` by construction (mirrors the math.sqrt emit: always
+  // canon, since a negative finite base yields a NaN whose sign needs canonicalizing).
+  // Two exotic inputs follow sqrt rather than Math.pow semantics — a deliberate
+  // trade in the same class as jz's other boundary divergences: `(-0) ** 0.5` is -0
+  // (Math.pow: +0; and -0 === 0), `(-Infinity) ** 0.5` is NaN (Math.pow: +Infinity).
+  // `** -0.5` is intentionally NOT folded: 1/sqrt double-rounds and loses the last
+  // ULP vs Math.pow's single rounding, so it keeps the exact $math.pow path.
+  const powCall = emitter(['math.pow'], (a, b) => fn('math.pow', a, b))
+  // base-2 power → dedicated $math.exp2(y) (skips exp's ×ln2 / ÷ln2 round-trip)
+  const exp2Call = emitter(['math.exp2'], (exp) => typed(['call', '$math.exp2', toNumF64(exp, emit(exp))], 'f64'))
+  // Shared pow/** lowering.
+  const emitPow = (a, b, allowExpPos) => {
     const n = constInt(b)
-    return n !== null && Math.abs(n) <= POW_FOLD_MAX ? foldPow(a, n) : powCall(a, b)
+    if (n !== null && Math.abs(n) <= POW_FOLD_MAX) return foldPow(a, n)
+    if (constNum(b) === 0.5) { const ir = typed(['f64.sqrt', toNumF64(a, emit(a))], 'f64'); return nonNegF64(ir[1]) ? ir : canon(ir) }
+    // Both args are compile-time constants: evaluate now, emit f64.const.
+    // Catches pow(2, -2/12) where the arithmetic folds emit f64.const for both sides.
+    const ca = constNum(a), cb = constNum(b)
+    if (ca !== null && cb !== null) return typed(['f64.const', Math.pow(ca, cb)], 'f64')
+    // IR-level fold: peek at emitted IR for both args — e.g. -2/12 emits f64.const -0.1666.
+    // We emit, check, and if not foldable, the emitted IR is used by the fallthrough paths.
+    const irA = toNumF64(a, emit(a)), irB = toNumF64(b, emit(b))
+    if (isLit(irA) && isLit(irB)) return typed(['f64.const', Math.pow(litVal(irA), litVal(irB))], 'f64')
+    // base 2 → dedicated 2^y (exp2 is exact for integer y, and skips exp's ×ln2/÷ln2).
+    // Every other literal base keeps $math.pow: `exp(y·ln base)` would lose ulps and,
+    // worse, miss Math.pow's integer-exponent semantics — e.g. `16 ** flen` with a
+    // runtime-integer flen must reproduce the exact square-and-multiply value (2⁵² for
+    // flen=13), which only $math.pow's integer fast path delivers.
+    if (allowExpPos && isLit(irA) && litVal(irA) === 2 && n === null)
+      return (inc('math.exp2'), typed(['call', '$math.exp2', irB], 'f64'))
+    return (inc('math.pow'), typed(['call', '$math.pow', irA, irB], 'f64'))
   }
-  ctx.core.emit['math.pow'].deps = powCall.deps   // metadata parity (tabulation/analysis)
-  ctx.core.emit['**'] = ctx.core.emit['math.pow']
-  ctx.core.emit['math.cbrt'] = emitter(['math.cbrt'], a => call('math.cbrt', a))
-  ctx.core.emit['math.hypot'] = emitter(['math.hypot'], (a, b, ...rest) => {
+  ctx.core.emit['math.pow'] = tag((a, b) => emitPow(a, b, true), powCall.deps)
+  ctx.core.emit['**'] = tag((a, b) => emitPow(a, b, true), powCall.deps)
+  reg('math.cbrt', ['math.cbrt'], a => fn('math.cbrt', a))
+  reg('math.hypot', ['math.hypot'], (a, b, ...rest) => {
     if (a === undefined) return typed(['f64.const', 0], 'f64')
     if (b === undefined) return f('f64.abs', a)
-    let r = call('math.hypot', a, b)
+    let r = fn('math.hypot', a, b)
     // ToNumber every rest arg too (matches min/max) — an object arg's valueOf
     // must run and may throw, which Math.hypot propagates.
     for (const x of rest) r = typed(['call', '$math.hypot', r, toNumF64(x, emit(x))], 'f64')
@@ -246,7 +372,7 @@ export default (ctx) => {
 
   // Math.sumPrecise(iterable) — exact, correctly-rounded summation (ECMA-262).
   // jz models the array case; the WAT routine sums via a fixed-point accumulator.
-  ctx.core.emit['math.sumPrecise'] = emitter(['math.sumPrecise'], arr =>
+  reg('math.sumPrecise', ['math.sumPrecise'], arr =>
     typed(['call', '$math.sumPrecise', ['i64.reinterpret_f64', asF64(emit(arr))]], 'f64'))
 
   // Integer/bit operations: return i32 directly. Consumers `asF64`-rebox at
@@ -258,99 +384,170 @@ export default (ctx) => {
   ctx.core.emit['math.imul'] = (a, b) => typed(['i32.mul', toI32(emit(a)), toI32(emit(b))], 'i32')
 
   // Random
-  ctx.core.emit['math.random'] = emitter(['math.random'], () => typed(['call', '$math.random'], 'f64'))
+  reg('math.random', ['math.random'], () => {
+    // Entropy mode: pull the host randomness syscall on demand (only when
+    // Math.random is actually used) — env.rngSeed (JS host) or WASI random_get.
+    if (rngEntropy) {
+      if (ctx.transform.host === 'wasi')
+        hostImport('wasi_snapshot_preview1', 'random_get', ['func', '$__random_get', ['param', 'i32'], ['param', 'i32'], ['result', 'i32']])
+      else
+        hostImport('env', 'rngSeed', ['func', '$__env_rng_seed', ['result', 'i32']])
+    }
+    return typed(['call', '$math.random'], 'f64')
+  })
 
   // ============================================
   // WAT stdlib implementations
   // ============================================
 
-  ctx.core.stdlib['math.sign'] = `(func $math.sign (param $x f64) (result f64)
+  wat('math.sign', `(func $math.sign (param $x f64) (result f64)
     ;; sign(NaN) = NaN, sign(±0) = ±0 — both pass x through unchanged.
     (if (f64.ne (local.get $x) (local.get $x)) (then (return (local.get $x))))
     (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
     (if (result f64) (f64.gt (local.get $x) (f64.const 0.0))
       (then (f64.const 1.0))
-      (else (f64.const -1.0))))`
-
-  ctx.core.stdlib['math.sin'] = `(func $math.sin (param $x f64) (result f64)
-    (local $n i32) (local $r f64) (local $x2 f64) (local $sign f64)
-    ;; NaN/±Infinity → NaN (avoid trapping i32.trunc on infinities).
-    (if (i32.eqz (call $math.isFinite (local.get $x))) (then (return (f64.const nan))))
-    (local.set $sign (f64.const 1.0))
-    (local.set $n (i32.trunc_f64_s (f64.floor (f64.div (local.get $x) (f64.const ${Math.PI})))))
-    (local.set $r (f64.sub (local.get $x) (f64.mul (f64.convert_i32_s (local.get $n)) (f64.const ${Math.PI}))))
-    (if (i32.and (local.get $n) (i32.const 1)) (then (local.set $sign (f64.const -1.0))))
-    (if (f64.gt (local.get $r) (f64.const ${Math.PI / 2})) (then (local.set $r (f64.sub (f64.const ${Math.PI}) (local.get $r)))))
-    (if (f64.lt (local.get $r) (f64.const 0.0)) (then
-      (local.set $r (f64.neg (local.get $r)))
-      (local.set $sign (f64.neg (local.get $sign)))))
-    (local.set $x2 (f64.mul (local.get $r) (local.get $r)))
-    (f64.mul (local.get $sign) (f64.mul (local.get $r) (f64.sub (f64.const 1.0) (f64.mul (local.get $x2)
-      (f64.sub (f64.const 0.16666666666666666) (f64.mul (local.get $x2)
-        (f64.sub (f64.const 0.008333333333333333) (f64.mul (local.get $x2)
-          (f64.sub (f64.const 0.0001984126984126984) (f64.mul (local.get $x2)
-            (f64.sub (f64.const 0.0000027557319223985893) (f64.mul (local.get $x2)
-              (f64.const 2.505210838544172e-8))))))))))))))`
-
-  ctx.core.stdlib['math.cos'] = `(func $math.cos (param $x f64) (result f64)
-    (local $n i32) (local $r f64) (local $x2 f64) (local $sign f64)
-    ;; NaN/±Infinity → NaN (avoid trapping i32.trunc on infinities).
-    (if (i32.eqz (call $math.isFinite (local.get $x))) (then (return (f64.const nan))))
-    (local.set $sign (f64.const 1.0))
-    (local.set $n (i32.trunc_f64_s (f64.floor (f64.div (local.get $x) (f64.const ${Math.PI})))))
-    (local.set $r (f64.sub (local.get $x) (f64.mul (f64.convert_i32_s (local.get $n)) (f64.const ${Math.PI}))))
-    (if (i32.and (local.get $n) (i32.const 1)) (then (local.set $sign (f64.const -1.0))))
-    (if (f64.gt (local.get $r) (f64.const ${Math.PI / 2})) (then
-      (local.set $r (f64.sub (f64.const ${Math.PI}) (local.get $r)))
-      (local.set $sign (f64.neg (local.get $sign)))))
-    (if (f64.lt (local.get $r) (f64.const 0.0)) (then (local.set $r (f64.neg (local.get $r)))))
-    (local.set $x2 (f64.mul (local.get $r) (local.get $r)))
-    (f64.mul (local.get $sign) (f64.sub (f64.const 1.0) (f64.mul (local.get $x2)
-      (f64.sub (f64.const 0.5) (f64.mul (local.get $x2)
-        (f64.sub (f64.const 0.041666666666666664) (f64.mul (local.get $x2)
-          (f64.sub (f64.const 0.001388888888888889) (f64.mul (local.get $x2)
-            (f64.sub (f64.const 0.0000248015873015873) (f64.mul (local.get $x2)
-              (f64.const 2.7557319223985893e-7)))))))))))))`
-
-  ctx.core.stdlib['math.tan'] = `(func $math.tan (param $x f64) (result f64)
-    (f64.div (call $math.sin (local.get $x)) (call $math.cos (local.get $x))))`
-
-  ctx.core.stdlib['math.exp'] = `(func $math.exp (param $x f64) (result f64)
-    (local $k i32) (local $t f64) (local $t2 f64) (local $result f64) (local $pow2 f64)
-    (if (f64.ne (local.get $x) (local.get $x)) (then (return (local.get $x))))
-    ;; +Infinity → +Infinity; finite overflow (x > 709) also rounds to +Infinity.
-    (if (result f64) (f64.gt (local.get $x) (f64.const 709.0)) (then (f64.const inf)) (else
-      (if (result f64) (f64.lt (local.get $x) (f64.const -745.0)) (then (f64.const 0.0)) (else
-        (local.set $k (i32.trunc_f64_s (f64.div (local.get $x) (f64.const ${Math.LN2}))))
-        (local.set $t (f64.sub (local.get $x) (f64.mul (f64.convert_i32_s (local.get $k)) (f64.const ${Math.LN2}))))
-        (local.set $t2 (f64.mul (local.get $t) (local.get $t)))
-        (local.set $result (f64.add (f64.const 1.0) (f64.add (local.get $t)
-          (f64.mul (local.get $t2) (f64.add (f64.const 0.5)
-            (f64.mul (local.get $t) (f64.add (f64.const 0.16666666666666666)
-              (f64.mul (local.get $t) (f64.add (f64.const 0.041666666666666664)
-                (f64.mul (local.get $t) (f64.add (f64.const 0.008333333333333333)
-                  (f64.mul (local.get $t) (f64.const 0.001388888888888889)))))))))))))
-        (local.set $pow2 (f64.const 1.0))
-        (if (i32.gt_s (local.get $k) (i32.const 0))
-          (then (block $done (loop $loop
-            (br_if $done (i32.le_s (local.get $k) (i32.const 0)))
-            (local.set $pow2 (f64.mul (local.get $pow2) (f64.const 2.0)))
-            (local.set $k (i32.sub (local.get $k) (i32.const 1)))
-            (br $loop)))
-            (local.set $result (f64.mul (local.get $result) (local.get $pow2))))
-          (else (if (i32.lt_s (local.get $k) (i32.const 0))
-            (then (block $done2 (loop $loop2
-              (br_if $done2 (i32.ge_s (local.get $k) (i32.const 0)))
-              (local.set $pow2 (f64.mul (local.get $pow2) (f64.const 2.0)))
-              (local.set $k (i32.add (local.get $k) (i32.const 1)))
-              (br $loop2)))
-              (local.set $result (f64.div (local.get $result) (local.get $pow2)))))))
-        (local.get $result))))))`
-
-  ctx.core.stdlib['math.expm1'] = `(func $math.expm1 (param $x f64) (result f64)
-    ;; Preserve sign of zero: expm1(±0) = ±0.
-    (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
-    (f64.sub (call $math.exp (local.get $x)) (f64.const 1.0)))`
+      (else (f64.const -1.0))))`)
+
+  // sin/cos over the folded range [0, π/2] use a 5-term MINIMAX polynomial in x² (Horner
+  // form, generated below). It beats the prior 6-term Taylor on both counts: one fewer
+  // multiply (faster — the floatbeat synth is sin-bound), and lower error (sin ≤ 1.9e-8,
+  // cos ≤ 1.3e-7 vs Taylor's ~6e-8 / ~5e-7) — minimax spreads error evenly across the range
+  // instead of piling unused precision near 0. Coeffs fit by scripts/minimax-trig.mjs.
+  const horner = (cs, v) => cs.reduceRight((acc, c, i) =>
+    i === cs.length - 1 ? `(f64.const ${c})`
+      : `(f64.add (f64.const ${c}) (f64.mul (local.get ${v}) ${acc}))`, '')
+  const SIN_C = [1, -0.16666660296130772, 0.008333091744946387, -0.00019811771757028443, 0.000002611054662215034]
+  const COS_C = [1, -0.4999993043717576, 0.04166402742354027, -0.0013856638518363177, 0.00002321737177898552]
+  // 2^f over the reduced range f ∈ [-0.5, 0.5] for $math.exp2 (rel. err ≤ 6e-9). Lets the
+  // base-2 power `2**y` skip the ×ln2 / ÷ln2 round-trip exp(y·ln2) pays — see $math.exp2.
+  const EXP2_C = [1, 0.6931472000619209, 0.24022650999918949, 0.05550340682450019, 0.009618048870444599, 0.0013395279077191057, 0.00015463102004723134]
+  // Range-reduction constants embedded as exact round-trip decimal STRINGS, not
+  // `${Math.PI}`/`${1/Math.PI}` number interpolation. These multiply the (possibly
+  // astronomically large) argument, so any lost digit wrecks large-arg reduction. Under
+  // self-host the kernel formats `${number}` through __ftoa — a 9-significant-digit dtoa
+  // (module/number.js) — which would bake "3.14159265"/"0.318309886" into the WAT and
+  // throw the reduced quadrant off (Δ≈0.2 at x≈2267). A string interpolates verbatim, so
+  // watr parses the full-precision f64 in both legs. Values are native toString's shortest
+  // round-trip reprs of Math.PI, 1/Math.PI, Math.PI/2 — byte-identical to the old output.
+  const PI = '3.141592653589793', INV_PI = '0.3183098861837907', HALF_PI = '1.5707963267948966'
+
+  // Round-to-nearest reduction r = x − q·π ∈ [−π/2, π/2], in pure f64 — no int conversion,
+  // so it never traps and never saturates. A SECOND pass folds the q·π rounding error back
+  // in, keeping r bounded even for astronomically large x where the first pass loses all
+  // precision: Math.sin must return a value in [−1,1] for every finite input, not Inf/garbage.
+  // For |x| ≲ 1e15 the second pass is a no-op (q2 = 0) and the result is bit-identical to a
+  // single reduction. The odd poly r·P(r²) handles r<0 on its own (sin is odd); the sign is the
+  // parity of the total quotient, taken in f64 as q − 2·round(q/2). ×(1/π) avoids a divide.
+  // ±Infinity and NaN must return NaN. Guard before reduction instead of relying on
+  // Inf−Inf·π to mint one: that arithmetic NaN has platform-dependent bits and can
+  // escape as a non-canonical NaN-box on x86/Linux.
+  // The second reduction pass only corrects an r that the first pass left outside
+  // [−π/2, π/2]; for in-range r, q2 is 0 and the pass is a no-op, so gating it on
+  // |r| > π/2 is bit-identical for all finite inputs while sparing the common case
+  // ~6 ops. Both are generic wins for every sin/cos/tan/exp-via-trig call site.
+  wat('math.sin_core', `(func $math.sin_core (param $x f64) (result f64)
+    (local $q f64) (local $q2 f64) (local $r f64) (local $r2 f64)
+    (if (f64.ne (local.get $x) (local.get $x)) (then (return (f64.const nan))))
+    (if (f64.eq (f64.abs (local.get $x)) (f64.const inf)) (then (return (f64.const nan))))
+    ;; |x| ≤ 2⁻²⁷: sin(x) = x to within a fraction of an ulp, and returning x preserves the
+    ;; sign of ±0 (the range reduction below would turn -0 into +0: -0 − (-0·π) = +0).
+    (if (f64.lt (f64.abs (local.get $x)) (f64.const ${2 ** -27})) (then (return (local.get $x))))
+    (local.set $q (f64.nearest (f64.mul (local.get $x) (f64.const ${INV_PI}))))
+    (local.set $r (f64.sub (local.get $x) (f64.mul (local.get $q) (f64.const ${PI}))))
+    (if (f64.gt (f64.abs (local.get $r)) (f64.const ${HALF_PI}))
+      (then
+        (local.set $q2 (f64.nearest (f64.mul (local.get $r) (f64.const ${INV_PI}))))
+        (local.set $r (f64.sub (local.get $r) (f64.mul (local.get $q2) (f64.const ${PI}))))
+        (local.set $q (f64.add (local.get $q) (local.get $q2)))))
+    (local.set $q (f64.sub (local.get $q) (f64.mul (f64.const 2) (f64.nearest (f64.mul (local.get $q) (f64.const 0.5))))))
+    (local.set $r2 (f64.mul (local.get $r) (local.get $r)))
+    (local.set $r (f64.mul (local.get $r) ${horner(SIN_C, '$r2')}))
+    ;; Negate for odd quasiperiods
+    (if (f64.gt (f64.abs (local.get $q)) (f64.const 0.5)) (then (local.set $r (f64.neg (local.get $r)))))
+    ;; Clamp to [-1, 1]: polynomial approximation can overshoot by ~1e-8 near peaks.
+    ;; Branchless (f64.min/f64.max) avoids branch misprediction near peaks.
+    (f64.min (f64.max (local.get $r) (f64.const -1.0)) (f64.const 1.0)))`)
+
+  wat('math.sin', `(func $math.sin (param $x f64) (result f64)
+    (call $math.sin_core (local.get $x)))`)
+
+  wat('math.cos_core', `(func $math.cos_core (param $x f64) (result f64)
+    (local $q f64) (local $q2 f64) (local $r f64) (local $r2 f64)
+    (if (f64.ne (local.get $x) (local.get $x)) (then (return (f64.const nan))))
+    (if (f64.eq (f64.abs (local.get $x)) (f64.const inf)) (then (return (f64.const nan))))
+    (local.set $q (f64.nearest (f64.mul (local.get $x) (f64.const ${INV_PI}))))
+    (local.set $r (f64.sub (local.get $x) (f64.mul (local.get $q) (f64.const ${PI}))))
+    (if (f64.gt (f64.abs (local.get $r)) (f64.const ${HALF_PI}))
+      (then
+        (local.set $q2 (f64.nearest (f64.mul (local.get $r) (f64.const ${INV_PI}))))
+        (local.set $r (f64.sub (local.get $r) (f64.mul (local.get $q2) (f64.const ${PI}))))
+        (local.set $q (f64.add (local.get $q) (local.get $q2)))))
+    (local.set $q (f64.sub (local.get $q) (f64.mul (f64.const 2) (f64.nearest (f64.mul (local.get $q) (f64.const 0.5))))))
+    (local.set $r2 (f64.mul (local.get $r) (local.get $r)))
+    (local.set $r ${horner(COS_C, '$r2')})
+    ;; Negate for odd quasiperiods
+    (if (f64.gt (f64.abs (local.get $q)) (f64.const 0.5)) (then (local.set $r (f64.neg (local.get $r)))))
+    ;; Clamp to [-1, 1]: polynomial approximation can overshoot by ~1e-8 near peaks.
+    ;; Branchless (f64.min/f64.max) avoids branch misprediction near peaks.
+    (f64.min (f64.max (local.get $r) (f64.const -1.0)) (f64.const 1.0)))`)
+
+  wat('math.cos', `(func $math.cos (param $x f64) (result f64)
+    (call $math.cos_core (local.get $x)))`)
+
+  wat('math.tan', `(func $math.tan (param $x f64) (result f64)
+    (if (f64.ne (local.get $x) (local.get $x)) (then (return (f64.const nan))))
+    (if (f64.eq (f64.abs (local.get $x)) (f64.const inf)) (then (return (f64.const nan))))
+    (f64.div (call $math.sin (local.get $x)) (call $math.cos (local.get $x))))`)
+
+  // e^x = 2^(x·log2 e) — defer to the faster $math.exp2 (one multiply, no division, and
+  // exp2's NaN/overflow/underflow guards cover exp's). Accurate to exp2's ~6e-9, better
+  // than the old 7-term Taylor, and it shares one code path with `2**`.
+  wat('math.exp', `(func $math.exp (param $x f64) (result f64)
+    (call $math.exp2 (f64.mul (local.get $x) (f64.const ${Math.LOG2E}))))`)
+
+  // 2^y, the dedicated base-2 power. `2**y` lowers here instead of exp(y·ln2): no ×ln2
+  // (so no reciprocal cancellation against exp's ÷ln2), a poly over the tighter [-0.5,0.5],
+  // and the same O(1) IEEE-exponent build of 2^k. ~6e-9 rel. error — well inside tolerance.
+  wat('math.exp2', `(func $math.exp2 (param $y f64) (result f64)
+    (local $k i32) (local $f f64) (local $k2 i32) (local $p f64)
+    (if (f64.ne (local.get $y) (local.get $y)) (then (return (local.get $y))))
+    (if (result f64) (f64.gt (local.get $y) (f64.const 1024.0)) (then (f64.const inf)) (else
+      (if (result f64) (f64.lt (local.get $y) (f64.const -1075.0)) (then (f64.const 0.0)) (else
+        (local.set $k (i32.trunc_f64_s (f64.nearest (local.get $y))))
+        (local.set $f (f64.sub (local.get $y) (f64.convert_i32_s (local.get $k))))
+        (local.set $p ${horner(EXP2_C, '$f')})
+        ;; 2^k via a single IEEE-exponent build for the normal range (the hot path); the
+        ;; two-factor split (2^k2 · 2^(k−k2)) is only needed at the denormal/overflow edges.
+        ;; For normal k both are bit-identical (powers of two multiply exactly) — free speedup.
+        (if (result f64)
+          (i32.and (i32.gt_s (local.get $k) (i32.const -1023)) (i32.lt_s (local.get $k) (i32.const 1024)))
+          (then (f64.mul (local.get $p)
+            (f64.reinterpret_i64 (i64.shl (i64.extend_i32_s (i32.add (local.get $k) (i32.const 1023))) (i64.const 52)))))
+          (else
+            (local.set $k2 (i32.shr_s (local.get $k) (i32.const 1)))
+            (f64.mul (f64.mul (local.get $p)
+              (f64.reinterpret_i64 (i64.shl (i64.extend_i32_s (i32.add (local.get $k2) (i32.const 1023))) (i64.const 52))))
+              (f64.reinterpret_i64 (i64.shl (i64.extend_i32_s (i32.add (i32.sub (local.get $k) (local.get $k2)) (i32.const 1023))) (i64.const 52)))))))))))`)
+
+  // Maclaurin coefficients 1/1!…1/8! for e^x−1 = x·(1 + x/2! + x²/3! + …), Horner-nested
+  // (built with an explicit right-to-left fold so the parens stay balanced — and so the
+  // builder uses only constructs the self-host kernel can compile: Array.reduceRight is
+  // not in jz's runtime, so under the kernel it returns undefined and that token lands
+  // verbatim in the emitted WAT).
+  const expm1Coef = [1, 1 / 2, 1 / 6, 1 / 24, 1 / 120, 1 / 720, 1 / 5040, 1 / 40320]
+  let expm1Series = ''
+  for (let i = expm1Coef.length - 1; i >= 0; i--)
+    expm1Series = expm1Series
+      ? `(f64.add (f64.const ${expm1Coef[i]}) (f64.mul (local.get $x) ${expm1Series}))`
+      : `(f64.const ${expm1Coef[i]})`
+  wat('math.expm1', `(func $math.expm1 (param $x f64) (result f64)
+    ;; expm1(x) = e^x − 1. For |x| < 0.5 sum the series directly: there e^x is within ~1.6
+    ;; of 1, so exp(x)−1 cancels the leading digits (the prior naive form lost up to ~11%
+    ;; near 0); the series doesn't, and the leading x·(…) preserves the sign of ±0. Larger
+    ;; |x| has no cancellation, so exp(x)−1 is accurate.
+    (if (result f64) (f64.lt (f64.abs (local.get $x)) (f64.const 0.5))
+      (then (f64.mul (local.get $x) ${expm1Series}))
+      (else (f64.sub (call $math.exp (local.get $x)) (f64.const 1.0)))))`)
 
   // log(x) via bit-level frexp + sqrt(2)-centered split + atanh series.
   //   x = m * 2^k   with bits-extracted k (no loop)
@@ -360,8 +557,9 @@ export default (ctx) => {
   // With 9 polynomial terms and |s|≤0.172, truncation error ≈ 2|s|·z⁹/19 ≈ 4e-17,
   // close to f64 ulp. The whole routine is branchless after edge cases.
   // Edge cases: NaN→NaN, ≤0 distinguishes 0→-Inf, <0→NaN; +Inf passes through.
-  ctx.core.stdlib['math.log'] = `(func $math.log (param $x f64) (result f64)
+  wat('math.log', `(func $math.log (param $x f64) (result f64)
     (local $bits i64) (local $k i32) (local $m f64) (local $s f64) (local $z f64)
+    (local $f f64) (local $w f64) (local $t1 f64) (local $t2 f64) (local $hfsq f64)
     (if (f64.ne (local.get $x) (local.get $x))
       (then (return (local.get $x))))
     (if (f64.le (local.get $x) (f64.const 0.0))
@@ -392,24 +590,30 @@ export default (ctx) => {
       (then
         (local.set $m (f64.mul (local.get $m) (f64.const 0.5)))
         (local.set $k (i32.add (local.get $k) (i32.const 1)))))
-    (local.set $s (f64.div (f64.sub (local.get $m) (f64.const 1.0))
-                           (f64.add (local.get $m) (f64.const 1.0))))
+    ;; s = f/(2+f) with f = m−1 (= (m−1)/(m+1)); then the fdlibm even/odd-split
+    ;; polynomial. Two parallel Horner chains (t1 over even powers, t2 over odd)
+    ;; cut the dependency chain ~in half vs one 9-deep Horner — more ILP, fewer
+    ;; terms — and reconstruct log(m) = f − hfsq + s·(hfsq + t1 + t2). ~1 ulp.
+    (local.set $f (f64.sub (local.get $m) (f64.const 1.0)))
+    (local.set $s (f64.div (local.get $f) (f64.add (local.get $f) (f64.const 2.0))))
     (local.set $z (f64.mul (local.get $s) (local.get $s)))
-    ;; Horner: 1 + z/3 + z²/5 + z³/7 + z⁴/9 + z⁵/11 + z⁶/13 + z⁷/15 + z⁸/17
+    (local.set $w (f64.mul (local.get $z) (local.get $z)))
+    (local.set $t1 (f64.mul (local.get $w) (f64.add (f64.const 0.3999999999940941908)
+      (f64.mul (local.get $w) (f64.add (f64.const 0.2222219843214978396)
+        (f64.mul (local.get $w) (f64.const 0.1531383769920937332)))))))
+    (local.set $t2 (f64.mul (local.get $z) (f64.add (f64.const 0.6666666666666735130)
+      (f64.mul (local.get $w) (f64.add (f64.const 0.2857142874366239149)
+        (f64.mul (local.get $w) (f64.add (f64.const 0.1818357216161805012)
+          (f64.mul (local.get $w) (f64.const 0.1479819860511658591)))))))))
+    (local.set $hfsq (f64.mul (f64.const 0.5) (f64.mul (local.get $f) (local.get $f))))
     (f64.add
       (f64.mul (f64.convert_i32_s (local.get $k)) (f64.const ${Math.LN2}))
-      (f64.mul (f64.const 2.0) (f64.mul (local.get $s) (f64.add (f64.const 1.0)
-        (f64.mul (local.get $z) (f64.add (f64.const 0.3333333333333333)
-          (f64.mul (local.get $z) (f64.add (f64.const 0.2)
-            (f64.mul (local.get $z) (f64.add (f64.const 0.14285714285714285)
-              (f64.mul (local.get $z) (f64.add (f64.const 0.1111111111111111)
-                (f64.mul (local.get $z) (f64.add (f64.const 0.09090909090909091)
-                  (f64.mul (local.get $z) (f64.add (f64.const 0.07692307692307693)
-                    (f64.mul (local.get $z) (f64.add (f64.const 0.06666666666666667)
-                      (f64.mul (local.get $z) (f64.const 0.058823529411764705)))))))))))))))))))))`
-
-  ctx.core.stdlib['math.log2'] = `(func $math.log2 (param $x f64) (result f64)
-    (f64.div (call $math.log (local.get $x)) (f64.const ${Math.LN2})))`
+      (f64.add (f64.sub (local.get $f) (local.get $hfsq))
+        (f64.mul (local.get $s) (f64.add (local.get $hfsq)
+          (f64.add (local.get $t1) (local.get $t2))))))))`)
+
+  wat('math.log2', `(func $math.log2 (param $x f64) (result f64)
+    (f64.div (call $math.log (local.get $x)) (f64.const ${Math.LN2})))`)
 
   // log10 via fdlibm's two-term decomposition: log10(x) = k*log10(2) + log10(m).
   // A plain log(x)/ln(10) double-rounds (rounding of log itself, then of the
@@ -417,7 +621,7 @@ export default (ctx) => {
   // Reducing x = m·2^k, splitting log10(2) and 1/ln(10) into hi/lo halves, and
   // keeping the bulk term (k·log10_2hi, hi·ivln10hi) carry-free recovers the
   // last ulps, so log10(10/100/1000/…) round-trips to exact integers.
-  ctx.core.stdlib['math.log10'] = `(func $math.log10 (param $x f64) (result f64)
+  wat('math.log10', `(func $math.log10 (param $x f64) (result f64)
     (local $bits i64) (local $k i32) (local $m f64) (local $f f64)
     (local $hfsq f64) (local $s f64) (local $z f64) (local $w f64)
     (local $t1 f64) (local $t2 f64) (local $R f64)
@@ -481,12 +685,12 @@ export default (ctx) => {
     (local.set $w (f64.add (local.get $y) (local.get $valhi)))
     (local.set $vallo (f64.add (local.get $vallo)
       (f64.add (f64.sub (local.get $y) (local.get $w)) (local.get $valhi))))
-    (f64.add (local.get $vallo) (local.get $w)))`
+    (f64.add (local.get $vallo) (local.get $w)))`)
 
   // log1p(x) via Kahan's compensated trick: with u = 1+x, log(u) loses bits when x is
   // small (because u rounds to ~1), but the ratio x/(u-1) is exactly the missing factor.
   // For u==1 (x below ulp), result is just x; preserves -0 from x=-0 path.
-  ctx.core.stdlib['math.log1p'] = `(func $math.log1p (param $x f64) (result f64)
+  wat('math.log1p', `(func $math.log1p (param $x f64) (result f64)
     (local $u f64)
     ;; log1p(+Inf) = +Inf — the ratio trick below would compute Inf/Inf = NaN.
     (if (f64.eq (local.get $x) (f64.const inf)) (then (return (f64.const inf))))
@@ -495,9 +699,9 @@ export default (ctx) => {
       (then (return (local.get $x))))
     (f64.div
       (f64.mul (call $math.log (local.get $u)) (local.get $x))
-      (f64.sub (local.get $u) (f64.const 1.0))))`
+      (f64.sub (local.get $u) (f64.const 1.0))))`)
 
-  ctx.core.stdlib['math.pow'] = `(func $math.pow (param $x f64) (param $y f64) (result f64)
+  wat('math.pow', `(func $math.pow (param $x f64) (param $y f64) (result f64)
     (local $result f64) (local $n i32) (local $neg_base i32) (local $abs_x f64)
     ;; y == 0 -> 1 (covers pow(NaN,0), pow(±0,0), pow(±Inf,0))
     (if (f64.eq (local.get $y) (f64.const 0.0)) (then (return (f64.const 1.0))))
@@ -570,12 +774,12 @@ export default (ctx) => {
     ;; x < 0, non-integer finite y -> NaN
     (if (f64.lt (local.get $x) (f64.const 0.0))
       (then (return (f64.const nan))))
-    (call $math.exp (f64.mul (local.get $y) (call $math.log (local.get $x)))))`
+    (call $math.exp (f64.mul (local.get $y) (call $math.log (local.get $x)))))`)
 
   // fdlibm atan: 4-region argument reduction onto |r| ≤ tan(π/16), then an
   // 11-term odd polynomial split into even/odd parts. Accurate to <1 ulp —
   // the old Taylor series was ~2e-6 off near |x|=0.5. Drives asin/acos/atan2.
-  ctx.core.stdlib['math.atan'] = `(func $math.atan (param $x f64) (result f64)
+  wat('math.atan', `(func $math.atan (param $x f64) (result f64)
     (local $abs_x f64) (local $id i32) (local $r f64) (local $z f64) (local $w f64)
     (local $s1 f64) (local $s2 f64) (local $ahi f64) (local $alo f64) (local $res f64)
     ;; NaN passes through unchanged.
@@ -642,19 +846,20 @@ export default (ctx) => {
     (local.set $res (f64.sub (local.get $ahi)
       (f64.sub (f64.sub (f64.mul (local.get $r) (f64.add (local.get $s1) (local.get $s2))) (local.get $alo))
                (local.get $r))))
-    (f64.copysign (local.get $res) (local.get $x)))`
+    (f64.copysign (local.get $res) (local.get $x)))`)
 
-  ctx.core.stdlib['math.asin'] = `(func $math.asin (param $x f64) (result f64)
+  wat('math.asin', `(func $math.asin (param $x f64) (result f64)
     ;; Domain is [-1, 1]; outside it (including ±Infinity), Math.asin returns NaN.
+    ;; sin/cos output is clamped to [-1, 1] by sin_core/cos_core, so no tolerance needed here.
     (if (result f64) (f64.gt (f64.abs (local.get $x)) (f64.const 1.0))
       (then (f64.const nan))
       (else (call $math.atan (f64.div (local.get $x)
-        (f64.sqrt (f64.sub (f64.const 1.0) (f64.mul (local.get $x) (local.get $x)))))))))`
+        (f64.sqrt (f64.sub (f64.const 1.0) (f64.mul (local.get $x) (local.get $x)))))))))`)
 
-  ctx.core.stdlib['math.acos'] = `(func $math.acos (param $x f64) (result f64)
-    (f64.sub (f64.const ${Math.PI / 2}) (call $math.asin (local.get $x))))`
+  wat('math.acos', `(func $math.acos (param $x f64) (result f64)
+    (f64.sub (f64.const ${HALF_PI}) (call $math.asin (local.get $x))))`)
 
-  ctx.core.stdlib['math.atan2'] = `(func $math.atan2 (param $y f64) (param $x f64) (result f64)
+  wat('math.atan2', `(func $math.atan2 (param $y f64) (param $x f64) (result f64)
     ;; If either argument is NaN, the result is NaN (ECMA-262 21.3.2.5).
     (if (f64.ne (local.get $x) (local.get $x)) (then (return (local.get $x))))
     (if (f64.ne (local.get $y) (local.get $y)) (then (return (local.get $y))))
@@ -662,30 +867,30 @@ export default (ctx) => {
       ;; y is ±0 too: result is ±0 when x is +0, ±π when x is -0; sign taken from y.
       (if (result f64) (f64.eq (local.get $y) (f64.const 0.0))
         (then (f64.copysign
-          (select (f64.const ${Math.PI}) (f64.const 0.0)
+          (select (f64.const ${PI}) (f64.const 0.0)
                   (f64.lt (f64.copysign (f64.const 1.0) (local.get $x)) (f64.const 0.0)))
           (local.get $y)))
         (else
-          (if (result f64) (f64.gt (local.get $y) (f64.const 0.0)) (then (f64.const ${Math.PI / 2})) (else (f64.neg (f64.const ${Math.PI / 2})))))))
+          (if (result f64) (f64.gt (local.get $y) (f64.const 0.0)) (then (f64.const ${HALF_PI})) (else (f64.neg (f64.const ${HALF_PI})))))))
       (else (if (result f64) (f64.ge (local.get $x) (f64.const 0.0))
         (then (call $math.atan (f64.div (local.get $y) (local.get $x))))
         (else (if (result f64) (f64.ge (local.get $y) (f64.const 0.0))
-          (then (f64.add (call $math.atan (f64.div (local.get $y) (local.get $x))) (f64.const ${Math.PI})))
-          (else (f64.sub (call $math.atan (f64.div (local.get $y) (local.get $x))) (f64.const ${Math.PI})))))))))`
+          (then (f64.add (call $math.atan (f64.div (local.get $y) (local.get $x))) (f64.const ${PI})))
+          (else (f64.sub (call $math.atan (f64.div (local.get $y) (local.get $x))) (f64.const ${PI})))))))))`)
 
-  ctx.core.stdlib['math.sinh'] = `(func $math.sinh (param $x f64) (result f64)
+  wat('math.sinh', `(func $math.sinh (param $x f64) (result f64)
     (local $ex f64)
     ;; Preserve sign of zero: sinh(±0) = ±0 (the f64.lt sign test below is false for -0).
     (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
     (local.set $ex (call $math.exp (f64.abs (local.get $x))))
     (local.set $ex (f64.mul (f64.const 0.5) (f64.sub (local.get $ex) (f64.div (f64.const 1.0) (local.get $ex)))))
-    (if (result f64) (f64.lt (local.get $x) (f64.const 0.0)) (then (f64.neg (local.get $ex))) (else (local.get $ex))))`
+    (if (result f64) (f64.lt (local.get $x) (f64.const 0.0)) (then (f64.neg (local.get $ex))) (else (local.get $ex))))`)
 
-  ctx.core.stdlib['math.cosh'] = `(func $math.cosh (param $x f64) (result f64)
+  wat('math.cosh', `(func $math.cosh (param $x f64) (result f64)
     (local $ex f64) (local.set $ex (call $math.exp (f64.abs (local.get $x))))
-    (f64.mul (f64.const 0.5) (f64.add (local.get $ex) (f64.div (f64.const 1.0) (local.get $ex)))))`
+    (f64.mul (f64.const 0.5) (f64.add (local.get $ex) (f64.div (f64.const 1.0) (local.get $ex)))))`)
 
-  ctx.core.stdlib['math.tanh'] = `(func $math.tanh (param $x f64) (result f64)
+  wat('math.tanh', `(func $math.tanh (param $x f64) (result f64)
     (local $e2x f64)
     ;; Preserve sign of zero: tanh(±0) = ±0 (the f64.lt sign test below is false for -0).
     (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
@@ -693,30 +898,30 @@ export default (ctx) => {
       (then (if (result f64) (f64.lt (local.get $x) (f64.const 0.0)) (then (f64.const -1.0)) (else (f64.const 1.0))))
       (else (local.set $e2x (call $math.exp (f64.mul (f64.const 2.0) (f64.abs (local.get $x)))))
         (local.set $e2x (f64.div (f64.sub (local.get $e2x) (f64.const 1.0)) (f64.add (local.get $e2x) (f64.const 1.0))))
-        (if (result f64) (f64.lt (local.get $x) (f64.const 0.0)) (then (f64.neg (local.get $e2x))) (else (local.get $e2x))))))`
+        (if (result f64) (f64.lt (local.get $x) (f64.const 0.0)) (then (f64.neg (local.get $e2x))) (else (local.get $e2x))))))`)
 
-  ctx.core.stdlib['math.asinh'] = `(func $math.asinh (param $x f64) (result f64)
+  wat('math.asinh', `(func $math.asinh (param $x f64) (result f64)
     ;; ±Infinity and NaN pass through unchanged. (log(±Inf + sqrt(Inf²+1)) → NaN otherwise.)
     (if (i32.eqz (call $math.isFinite (local.get $x))) (then (return (local.get $x))))
     ;; Preserve sign of zero: asinh(±0) = ±0.
     (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
-    (call $math.log (f64.add (local.get $x) (f64.sqrt (f64.add (f64.mul (local.get $x) (local.get $x)) (f64.const 1.0))))))`
+    (call $math.log (f64.add (local.get $x) (f64.sqrt (f64.add (f64.mul (local.get $x) (local.get $x)) (f64.const 1.0))))))`)
 
-  ctx.core.stdlib['math.acosh'] = `(func $math.acosh (param $x f64) (result f64)
+  wat('math.acosh', `(func $math.acosh (param $x f64) (result f64)
     (if (f64.eq (local.get $x) (f64.const inf)) (then (return (f64.const inf))))
     ;; acosh is defined only for x >= 1; everything below (incl. -Inf) is NaN.
     (if (result f64) (f64.lt (local.get $x) (f64.const 1.0)) (then (f64.const nan)) (else
-      (call $math.log (f64.add (local.get $x) (f64.sqrt (f64.sub (f64.mul (local.get $x) (local.get $x)) (f64.const 1.0))))))))`
+      (call $math.log (f64.add (local.get $x) (f64.sqrt (f64.sub (f64.mul (local.get $x) (local.get $x)) (f64.const 1.0))))))))`)
 
-  ctx.core.stdlib['math.atanh'] = `(func $math.atanh (param $x f64) (result f64)
+  wat('math.atanh', `(func $math.atanh (param $x f64) (result f64)
     ;; Preserve sign of zero: atanh(±0) = ±0.
     (if (f64.eq (local.get $x) (f64.const 0.0)) (then (return (local.get $x))))
     ;; ±Infinity → NaN. Without this the (1+x)/(1-x) ratio is Inf/Inf, whose
     ;; sign-nondeterministic arithmetic NaN would escape non-canonical on x86.
     (if (f64.eq (f64.abs (local.get $x)) (f64.const inf)) (then (return (f64.const nan))))
-    (f64.mul (f64.const 0.5) (call $math.log (f64.div (f64.add (f64.const 1.0) (local.get $x)) (f64.sub (f64.const 1.0) (local.get $x))))))`
+    (f64.mul (f64.const 0.5) (call $math.log (f64.div (f64.add (f64.const 1.0) (local.get $x)) (f64.sub (f64.const 1.0) (local.get $x))))))`)
 
-  ctx.core.stdlib['math.cbrt'] = `(func $math.cbrt (param $x f64) (result f64)
+  wat('math.cbrt', `(func $math.cbrt (param $x f64) (result f64)
     (local $y f64)
     ;; ±Infinity and NaN pass through; preserve sign of zero.
     (if (i32.eqz (call $math.isFinite (local.get $x))) (then (return (local.get $x))))
@@ -728,31 +933,37 @@ export default (ctx) => {
         (local.set $y (call $math.pow (local.get $x) (f64.const 0.3333333333333333)))
         (local.set $y (f64.div (f64.add (f64.mul (f64.const 2.0) (local.get $y)) (f64.div (local.get $x) (f64.mul (local.get $y) (local.get $y)))) (f64.const 3.0)))
         (local.set $y (f64.div (f64.add (f64.mul (f64.const 2.0) (local.get $y)) (f64.div (local.get $x) (f64.mul (local.get $y) (local.get $y)))) (f64.const 3.0)))
-        (local.get $y))))`
+        (local.get $y))))`)
 
   // Small finite-test helper (NaN→0, ±Inf→0, finite→1). Used by transcendental
   // functions that need to short-circuit on infinite inputs.
-  ctx.core.stdlib['math.isFinite'] = `(func $math.isFinite (param $x f64) (result i32)
+  wat('math.isFinite', `(func $math.isFinite (param $x f64) (result i32)
     (i32.and
       (f64.eq (local.get $x) (local.get $x))
-      (f64.lt (f64.abs (local.get $x)) (f64.const inf))))`
+      (f64.lt (f64.abs (local.get $x)) (f64.const inf))))`)
 
-  ctx.core.stdlib['math.hypot'] = `(func $math.hypot (param $x f64) (param $y f64) (result f64)
+  wat('math.hypot', `(func $math.hypot (param $x f64) (param $y f64) (result f64)
     ;; Any ±Infinity argument ⇒ +Infinity, even when the other is NaN (ECMA-262 21.3.2.18).
     (if (f64.eq (f64.abs (local.get $x)) (f64.const inf)) (then (return (f64.const inf))))
     (if (f64.eq (f64.abs (local.get $y)) (f64.const inf)) (then (return (f64.const inf))))
-    (f64.sqrt (f64.add (f64.mul (local.get $x) (local.get $x)) (f64.mul (local.get $y) (local.get $y)))))`
+    (f64.sqrt (f64.add (f64.mul (local.get $x) (local.get $x)) (f64.mul (local.get $y) (local.get $y)))))`)
 
-  ctx.core.stdlib['math.random'] = `(func $math.random (result f64)
+  // xorshift32 → [0,1). In entropy mode a one-shot prologue replaces the fixed
+  // initial state with host entropy on first call (branch is well-predicted after).
+  const rngSeedPrologue = rngEntropy ? `(if (i32.eqz (global.get $math.rng_seeded))
+      (then (global.set $math.rng_state (call $__rng_seed)) (global.set $math.rng_seeded (i32.const 1))))
+    ` : ``
+  wat('math.random', `(func $math.random (result f64)
     (local $s i32)
-    (local.set $s (global.get $math.rng_state))
+    ${rngSeedPrologue}(local.set $s (global.get $math.rng_state))
     (local.set $s (i32.xor (local.get $s) (i32.shl (local.get $s) (i32.const 13))))
     (local.set $s (i32.xor (local.get $s) (i32.shr_u (local.get $s) (i32.const 17))))
     (local.set $s (i32.xor (local.get $s) (i32.shl (local.get $s) (i32.const 5))))
     (global.set $math.rng_state (local.get $s))
-    (f64.div (f64.convert_i32_u (i32.and (local.get $s) (i32.const 0x7FFFFFFF))) (f64.const 2147483647.0)))`
+    (f64.div (f64.convert_i32_u (i32.and (local.get $s) (i32.const 0x7FFFFFFF))) (f64.const 2147483647.0)))`,
+    rngEntropy ? ['__rng_seed'] : [])
 
-  ctx.core.stdlib['math.sumPrecise'] = `(func $math.sumPrecise (param $arr i64) (result f64)
+  wat('math.sumPrecise', `(func $math.sumPrecise (param $arr i64) (result f64)
     ;; Exact summation via a 2304-bit fixed-point accumulator (36 i64 words,
     ;; little-endian two's complement) holding sum*2^1074. Every finite f64 is an
     ;; integer multiple of 2^-1074, so the running sum carries zero rounding
@@ -911,8 +1122,25 @@ export default (ctx) => {
       (then (f64.reinterpret_i64 (i64.shl (i64.extend_i32_u (i32.add (local.get $k) (i32.const 1023))) (i64.const 52))))
       (else (f64.reinterpret_i64 (i64.shl (i64.const 1) (i64.extend_i32_u (i32.add (local.get $k) (i32.const 1074))))))))
     (local.set $res (f64.mul (f64.convert_i64_u (local.get $top)) (local.get $pow)))
-    (select (f64.neg (local.get $res)) (local.get $res) (local.get $resultNeg)))`
+    (select (f64.neg (local.get $res)) (local.get $res) (local.get $resultNeg)))`)
 
-  // Global for random state
-  ctx.scope.globals.set('math.rng_state', '(global $math.rng_state (mut i32) (i32.const 12345))')
+  // Global for random state — seeded with the fixed constant (deterministic) or,
+  // in entropy mode, overwritten from the host on first Math.random() call.
+  declGlobal('math.rng_state', 'i32', rngSeedConst)
+  if (rngEntropy) {
+    declGlobal('math.rng_seeded', 'i32')
+    // One i32 of host entropy, floored at 1 (xorshift32 is dead at state 0).
+    wat('__rng_seed', ctx.transform.host === 'wasi'
+      ? `(func $__rng_seed (result i32)
+    (local $buf i32) (local $s i32)
+    (local.set $buf (call $__alloc (i32.const 4)))
+    (drop (call $__random_get (local.get $buf) (i32.const 4)))
+    (local.set $s (i32.load (local.get $buf)))
+    (select (local.get $s) (i32.const 1) (local.get $s)))`
+      : `(func $__rng_seed (result i32)
+    (local $s i32)
+    (local.set $s (call $__env_rng_seed))
+    (select (local.get $s) (i32.const 1) (local.get $s)))`,
+      ctx.transform.host === 'wasi' ? ['__alloc'] : [])
+  }
 }