pi-mega-compact 0.20.0 → 0.20.1

package/src/vector-cortex/residual/dct.ts

@@ -0,0 +1,166 @@
+/**
+ * vector-cortex/residual/dct.ts — orthonormal DCT-II basis V1 (VC4B).
+ *
+ * Basis V1 is the orthonormal DCT-II matrix generated ANALYTICALLY and never
+ * learned or stored (RESIDUAL_CODEC §byte scope):
+ *
+ *   alpha(0) = sqrt(1/n),  alpha(k>0) = sqrt(2/n)
+ *   C[k][i]  = alpha(k) * cos(pi * (2i + 1) * k / (2n))
+ *
+ * Forward coefficients are the dot product with each basis row (the orthonormal
+ * least-squares solve); the inverse is the transpose applied to the coefficient
+ * vector. Byte mapping is `x = (byte - 127.5) / 127.5`; the inverse maps back
+ * with round-to-nearest-even and clamps to 0..255.
+ *
+ * The full 4096x4096 matrix is ~134 MB of float64, so it is NEVER materialized.
+ * Instead the cosine argument `pi*(2i+1)*k/(2n)` is reduced modulo `2n` against
+ * a precomputed half-period table of length `2n` — this is EXACTLY the same set
+ * of cosine values the matrix would hold (cos is sampled only at the `2n`
+ * distinct arguments `pi*j/(2n)`, j = 0..2n-1), so the transform is bit-stable
+ * and table-generation is O(n) rather than O(n^2).
+ *
+ * Pure numeric transform: no storage, no console, no network (PREVENT-PI-004 /
+ * PREVENT-011).
+ */
+
+import { RESIDUAL_BLOCK_SIZE } from "./types.js";
+
+/** Byte-to-signal mapping midpoint / half-range (RESIDUAL_CODEC §transform). */
+const BYTE_MIDPOINT = 127.5;
+
+/**
+ * Cosine table cache keyed by block length. `table[j] = cos(pi * j / (2n))` for
+ * j = 0..2n-1; every DCT argument `pi*(2i+1)*k/(2n)` reduces to one of these
+ * entries (with a sign) because cos has period `2*pi` = index period `4n`.
+ */
+const cosTables = new Map<number, Float64Array>();
+
+/** Build (or fetch) the `4n`-entry cosine table for block length `n`. */
+function cosTable(n: number): Float64Array {
+  const cached = cosTables.get(n);
+  if (cached) return cached;
+  const period = 4 * n;
+  const table = new Float64Array(period);
+  for (let j = 0; j < period; j++) {
+    table[j] = Math.cos((Math.PI * j) / (2 * n));
+  }
+  cosTables.set(n, table);
+  return table;
+}
+
+/** Orthonormal DCT-II row scale: alpha(0)=sqrt(1/n), else sqrt(2/n). */
+export function alpha(k: number, n: number): number {
+  return k === 0 ? Math.sqrt(1 / n) : Math.sqrt(2 / n);
+}
+
+/**
+ * Map a padded byte block to the signal domain: `x = (byte - 127.5) / 127.5`.
+ * The input must be exactly `n` bytes (the caller zero-pads the final block).
+ */
+export function bytesToSignal(block: Uint8Array): Float64Array {
+  const out = new Float64Array(block.length);
+  for (let i = 0; i < block.length; i++) {
+    out[i] = (block[i]! - BYTE_MIDPOINT) / BYTE_MIDPOINT;
+  }
+  return out;
+}
+
+/**
+ * Map a reconstructed signal back to bytes: invert the affine map, round to
+ * nearest with ties-to-even, and clamp to 0..255 (RESIDUAL_CODEC §transform).
+ */
+export function signalToBytes(signal: Float64Array): Uint8Array {
+  const out = new Uint8Array(signal.length);
+  for (let i = 0; i < signal.length; i++) {
+    const v = signal[i]! * BYTE_MIDPOINT + BYTE_MIDPOINT;
+    const r = roundHalfToEven(v);
+    out[i] = r < 0 ? 0 : r > 255 ? 255 : r;
+  }
+  return out;
+}
+
+/**
+ * Round to nearest, ties to even (banker's rounding). `Math.round` rounds ties
+ * toward +Infinity, which is NOT the rule RESIDUAL_CODEC mandates.
+ */
+export function roundHalfToEven(v: number): number {
+  const floor = Math.floor(v);
+  const diff = v - floor;
+  if (diff > 0.5) return floor + 1;
+  if (diff < 0.5) return floor;
+  // Exact tie: pick the even neighbour.
+  return floor % 2 === 0 ? floor : floor + 1;
+}
+
+/**
+ * Forward orthonormal DCT-II over one block of length `n` (default 4096).
+ * `coefficients[k] = alpha(k) * sum_i x[i] * cos(pi*(2i+1)*k/(2n))`, emitted in
+ * ascending frequency order `k = 0..n-1` (the exact RESIDUAL_CODEC coefficient
+ * order).
+ */
+export function forwardDct(signal: Float64Array): Float64Array {
+  const n = signal.length;
+  const table = cosTable(n);
+  const period = 4 * n;
+  const out = new Float64Array(n);
+  for (let k = 0; k < n; k++) {
+    let sum = 0;
+    // Argument index advances by 2k each step, starting at k (i=0 gives
+    // (2*0+1)*k = k). Reduce modulo the 4n-entry period.
+    let idx = k % period;
+    const step = (2 * k) % period;
+    for (let i = 0; i < n; i++) {
+      sum += signal[i]! * table[idx]!;
+      idx += step;
+      if (idx >= period) idx -= period;
+    }
+    out[k] = alpha(k, n) * sum;
+  }
+  return out;
+}
+
+/**
+ * Inverse orthonormal DCT (the transpose of the forward matrix):
+ * `x[i] = sum_k alpha(k) * c[k] * cos(pi*(2i+1)*k/(2n))`.
+ */
+export function inverseDct(coefficients: Float64Array): Float64Array {
+  const n = coefficients.length;
+  const table = cosTable(n);
+  const period = 4 * n;
+  // Pre-scale each coefficient by its row alpha so the inner loop is a plain
+  // dot product (identical arithmetic to scaling inside the loop for k, and it
+  // keeps the per-i accumulation order stable).
+  const scaled = new Float64Array(n);
+  for (let k = 0; k < n; k++) scaled[k] = alpha(k, n) * coefficients[k]!;
+  const out = new Float64Array(n);
+  for (let i = 0; i < n; i++) {
+    let sum = 0;
+    const step = (2 * i + 1) % period;
+    let idx = 0;
+    for (let k = 0; k < n; k++) {
+      sum += scaled[k]! * table[idx]!;
+      idx += step;
+      if (idx >= period) idx -= period;
+    }
+    out[i] = sum;
+  }
+  return out;
+}
+
+/**
+ * Split a payload into fixed 4096-byte blocks, zero-padding ONLY the final
+ * block. The original length is retained by the caller (the header) so the
+ * padding is dropped on decode.
+ */
+export function splitBlocks(
+  payload: Uint8Array,
+  blockSize: number = RESIDUAL_BLOCK_SIZE,
+): Uint8Array[] {
+  const blocks: Uint8Array[] = [];
+  for (let off = 0; off < payload.length; off += blockSize) {
+    const block = new Uint8Array(blockSize);
+    block.set(payload.subarray(off, Math.min(off + blockSize, payload.length)));
+    blocks.push(block);
+  }
+  return blocks;
+}

package/src/vector-cortex/residual/fixture-payload.ts

@@ -0,0 +1,90 @@
+/**
+ * vector-cortex/residual/fixture-payload.ts — deterministic payload
+ * materialization for the VC4B conformance corpus.
+ *
+ * The residual fixtures describe their payloads GENERATIVELY (`kind` + `length`
+ * + `seed`) rather than embedding megabytes of base64, so this module is the
+ * single normative generator both the fixture producer and the acceptance test
+ * agree on. Every generator is pure and deterministic — the same descriptor
+ * always yields byte-identical output, which is what makes the committed
+ * fixtures meaningful.
+ *
+ * Pure byte generation: no storage, no console, no network (PREVENT-PI-004 /
+ * PREVENT-011).
+ */
+
+/** A generative payload descriptor as carried by a residual fixture. */
+export interface PayloadDescriptor {
+  readonly kind: string;
+  readonly length?: number;
+  readonly seed?: number;
+  readonly value?: number;
+  readonly outlierOffset?: number;
+  readonly bytesBase64?: string;
+}
+
+/** Linear congruential generator (numerical-recipes constants), deterministic. */
+function lcgBytes(length: number, seed: number): Uint8Array {
+  const out = new Uint8Array(length);
+  let state = seed >>> 0;
+  for (let i = 0; i < length; i++) {
+    state = (Math.imul(state, 1664525) + 1013904223) >>> 0;
+    out[i] = (state >>> 24) & 0xff;
+  }
+  return out;
+}
+
+/**
+ * Materialize a fixture payload descriptor into exact bytes.
+ *
+ *   empty        zero bytes
+ *   zeros        `length` 0x00 bytes
+ *   constant     `length` copies of `value`
+ *   sequence     `i % 256`
+ *   lcg          deterministic pseudorandom stream from `seed`
+ *   text         repeating printable ASCII (valid UTF-8)
+ *   invalid-utf8 a stream containing lone continuation/overlong bytes
+ *   dc-outlier   all 0xff except one 0x00 at `outlierOffset` (forces the exact
+ *                correction stream: a DC-dominant block whose coarse scale
+ *                cannot reproduce the outlier's neighbourhood)
+ *   alternating  0x00/0xff alternation (maximum Nyquist energy)
+ */
+export function materializePayload(d: PayloadDescriptor): Uint8Array {
+  const length = d.length ?? 0;
+  switch (d.kind) {
+    case "empty":
+      return new Uint8Array(0);
+    case "zeros":
+      return new Uint8Array(length);
+    case "constant":
+      return new Uint8Array(length).fill((d.value ?? 0) & 0xff);
+    case "sequence":
+      return Uint8Array.from({ length }, (_v, i) => i % 256);
+    case "lcg":
+      return lcgBytes(length, d.seed ?? 1);
+    case "text": {
+      const alphabet = "abcdefghijklmnopqrstuvwxyz0123456789 ";
+      return Uint8Array.from({ length }, (_v, i) =>
+        alphabet.charCodeAt(i % alphabet.length),
+      );
+    }
+    case "invalid-utf8": {
+      // Deliberately invalid: lone 0x80 continuation, 0xff (never valid), and
+      // the overlong 0xc0 0xaf encoding of "/". Never normalized by the codec.
+      const pattern = [0xff, 0x80, 0xc0, 0xaf, 0x00, 0xfe];
+      return Uint8Array.from({ length }, (_v, i) => pattern[i % pattern.length]!);
+    }
+    case "dc-outlier": {
+      const out = new Uint8Array(length).fill(255);
+      const at = d.outlierOffset ?? 0;
+      if (at < length) out[at] = 0;
+      return out;
+    }
+    case "alternating":
+      return Uint8Array.from({ length }, (_v, i) => (i % 2 === 0 ? 0 : 255));
+    case "literal":
+      return new Uint8Array(Buffer.from(d.bytesBase64 ?? "", "base64"));
+    default:
+      throw new Error(`unknown residual payload kind: ${d.kind}`);
+  }
+}

package/src/vector-cortex/residual/gf256.ts

@@ -0,0 +1,196 @@
+/**
+ * vector-cortex/residual/gf256.ts — GF(2^8) field arithmetic and matrix algebra
+ * for the VC4B Reed–Solomon erasure parity.
+ *
+ * Field: GF(2^8) with primitive polynomial `0x11d`, elements represented as
+ * polynomial-basis bytes (RESIDUAL_CODEC §erasure parity). Log/antilog tables
+ * are built once from generator 2 and drive constant-time multiply/divide.
+ *
+ * Matrix inversion and recovery use DETERMINISTIC left-to-right pivot search and
+ * GF Gaussian elimination — no randomness, no iteration-order dependence, so
+ * every implementation reaches byte-identical results.
+ *
+ * Pure arithmetic: no storage, no console, no network (PREVENT-PI-004 /
+ * PREVENT-011).
+ */
+
+import { GF_PRIMITIVE_POLYNOMIAL } from "./types.js";
+
+const FIELD_SIZE = 256;
+
+/** `EXP[i] = 2^i` in GF(2^8) (doubled length so multiply needs no modulo). */
+const EXP = new Uint8Array(FIELD_SIZE * 2);
+/** `LOG[x] = i` such that `2^i = x`; `LOG[0]` is unused (0 has no logarithm). */
+const LOG = new Uint8Array(FIELD_SIZE);
+
+(function buildTables(): void {
+  let x = 1;
+  for (let i = 0; i < FIELD_SIZE - 1; i++) {
+    EXP[i] = x;
+    LOG[x] = i;
+    x <<= 1;
+    if (x & 0x100) x ^= GF_PRIMITIVE_POLYNOMIAL;
+  }
+  // Mirror the cycle so `EXP[a + b]` is valid for a,b <= 254 without a modulo.
+  for (let i = FIELD_SIZE - 1; i < EXP.length; i++) {
+    EXP[i] = EXP[i - (FIELD_SIZE - 1)]!;
+  }
+})();
+
+/** GF(2^8) addition (and subtraction) is XOR. */
+export function gfAdd(a: number, b: number): number {
+  return (a ^ b) & 0xff;
+}
+
+/** GF(2^8) multiplication via log/antilog tables. */
+export function gfMul(a: number, b: number): number {
+  if (a === 0 || b === 0) return 0;
+  return EXP[LOG[a]! + LOG[b]!]!;
+}
+
+/** GF(2^8) division; dividing by zero is a programming error and throws. */
+export function gfDiv(a: number, b: number): number {
+  if (b === 0) throw new RangeError("gf256: division by zero");
+  if (a === 0) return 0;
+  return EXP[LOG[a]! + (FIELD_SIZE - 1) - LOG[b]!]!;
+}
+
+/** Multiplicative inverse of a non-zero field element. */
+export function gfInv(a: number): number {
+  if (a === 0) throw new RangeError("gf256: zero has no inverse");
+  return EXP[FIELD_SIZE - 1 - LOG[a]!]!;
+}
+
+/** `base^exponent` in GF(2^8) (exponent is a non-negative integer). */
+export function gfPow(base: number, exponent: number): number {
+  if (exponent === 0) return 1;
+  if (base === 0) return 0;
+  return EXP[(LOG[base]! * exponent) % (FIELD_SIZE - 1)]!;
+}
+
+/** A dense row-major GF(2^8) matrix. */
+export interface GfMatrix {
+  readonly rows: number;
+  readonly cols: number;
+  readonly data: Uint8Array;
+}
+
+/** Allocate a zero matrix. */
+export function gfMatrix(rows: number, cols: number): GfMatrix {
+  return { rows, cols, data: new Uint8Array(rows * cols) };
+}
+
+/** Read `m[r][c]`. */
+export function gfAt(m: GfMatrix, r: number, c: number): number {
+  return m.data[r * m.cols + c]!;
+}
+
+/** Write `m[r][c] = v`. */
+export function gfSet(m: GfMatrix, r: number, c: number, v: number): void {
+  m.data[r * m.cols + c] = v;
+}
+
+/**
+ * Build the `rows x cols` Vandermonde matrix `V[r][c] = alpha_r^c` with the
+ * distinct evaluation points `alpha_r = r + 1` (RESIDUAL_CODEC: rows r=0..8,
+ * columns c=0..5).
+ */
+export function vandermonde(rows: number, cols: number): GfMatrix {
+  const m = gfMatrix(rows, cols);
+  for (let r = 0; r < rows; r++) {
+    for (let c = 0; c < cols; c++) {
+      gfSet(m, r, c, gfPow((r + 1) & 0xff, c));
+    }
+  }
+  return m;
+}
+
+/** Matrix product `a x b` over GF(2^8). */
+export function gfMatMul(a: GfMatrix, b: GfMatrix): GfMatrix {
+  const out = gfMatrix(a.rows, b.cols);
+  for (let r = 0; r < a.rows; r++) {
+    for (let c = 0; c < b.cols; c++) {
+      let acc = 0;
+      for (let i = 0; i < a.cols; i++) {
+        acc ^= gfMul(gfAt(a, r, i), gfAt(b, i, c));
+      }
+      gfSet(out, r, c, acc);
+    }
+  }
+  return out;
+}
+
+/** Extract the contiguous row range `[start, start+count)` of a matrix. */
+export function gfSubRows(m: GfMatrix, start: number, count: number): GfMatrix {
+  const out = gfMatrix(count, m.cols);
+  out.data.set(m.data.subarray(start * m.cols, (start + count) * m.cols));
+  return out;
+}
+
+/** Gather the given row indices (in order) into a new matrix. */
+export function gfPickRows(m: GfMatrix, indices: readonly number[]): GfMatrix {
+  const out = gfMatrix(indices.length, m.cols);
+  indices.forEach((src, dest) => {
+    out.data.set(m.data.subarray(src * m.cols, (src + 1) * m.cols), dest * m.cols);
+  });
+  return out;
+}
+
+/**
+ * Invert a square GF(2^8) matrix by Gauss–Jordan elimination with a
+ * DETERMINISTIC left-to-right pivot search (the first row at or below the
+ * current column with a non-zero entry). Returns null when the matrix is
+ * singular.
+ */
+export function gfInvert(m: GfMatrix): GfMatrix | null {
+  if (m.rows !== m.cols) return null;
+  const n = m.rows;
+  const work = gfMatrix(n, n);
+  work.data.set(m.data);
+  const inv = gfMatrix(n, n);
+  for (let i = 0; i < n; i++) gfSet(inv, i, i, 1);
+
+  const swapRows = (mat: GfMatrix, a: number, b: number): void => {
+    if (a === b) return;
+    for (let c = 0; c < mat.cols; c++) {
+      const t = gfAt(mat, a, c);
+      gfSet(mat, a, c, gfAt(mat, b, c));
+      gfSet(mat, b, c, t);
+    }
+  };
+
+  for (let col = 0; col < n; col++) {
+    // Deterministic pivot: the lowest-index row >= col with a non-zero entry.
+    let pivot = -1;
+    for (let r = col; r < n; r++) {
+      if (gfAt(work, r, col) !== 0) {
+        pivot = r;
+        break;
+      }
+    }
+    if (pivot === -1) return null; // singular
+    swapRows(work, col, pivot);
+    swapRows(inv, col, pivot);
+
+    // Normalize the pivot row.
+    const pivotValue = gfAt(work, col, col);
+    if (pivotValue !== 1) {
+      const scale = gfInv(pivotValue);
+      for (let c = 0; c < n; c++) {
+        gfSet(work, col, c, gfMul(gfAt(work, col, c), scale));
+        gfSet(inv, col, c, gfMul(gfAt(inv, col, c), scale));
+      }
+    }
+    // Eliminate the column from every other row.
+    for (let r = 0; r < n; r++) {
+      if (r === col) continue;
+      const factor = gfAt(work, r, col);
+      if (factor === 0) continue;
+      for (let c = 0; c < n; c++) {
+        gfSet(work, r, c, gfAdd(gfAt(work, r, c), gfMul(factor, gfAt(work, col, c))));
+        gfSet(inv, r, c, gfAdd(gfAt(inv, r, c), gfMul(factor, gfAt(inv, col, c))));
+      }
+    }
+  }
+  return inv;
+}