pi-mega-compact 0.20.70 → 0.20.72

package/src/vector-cortex/provider/registry.ts

@@ -18,6 +18,7 @@
 
 import {
   PRO_PROFILE_UNKNOWN,
+  type ProviderEconomicsV1,
   type ProviderProfileBundle,
   type ProviderProfileResult,
   type ProviderProfileV1,
@@ -60,6 +61,7 @@ function profile(
   id: string,
   version: string,
   excludedJsonPointers: ProviderProfileV1["excludedJsonPointers"],
+  economics: ProviderProfileV1["economics"],
 ): ProviderProfileV1 {
   return {
     schema: "provider-profile-v1",
@@ -67,42 +69,111 @@ function profile(
     version,
     hashMode: "entire-canonical-request",
     excludedJsonPointers,
+    economics,
   };
 }
 
+/**
+ * Build integer micro-unit economics for a base profile (VC7B). A cache WRITE
+ * costs more than an uncached token, a cache READ costs less — the standard
+ * provider-prompt-cache shape. `exclusionFixtureId` mirrors the profile's own
+ * exclusion fixture (or null when the profile has none to prove).
+ */
+function econ(
+  id: string,
+  version: string,
+  exclusionFixtureId: string | null,
+  values: {
+    basePrice: number;
+    readPrice: number;
+    writePrice: number;
+    ttlMs: number;
+    minPrefix: number;
+  },
+): ProviderEconomicsV1 {
+  return {
+    schema: "provider-economics-v1",
+    profileId: id,
+    profileVersion: version,
+    basePrice: values.basePrice,
+    readPrice: values.readPrice,
+    writePrice: values.writePrice,
+    ttlMs: values.ttlMs,
+    minPrefix: values.minPrefix,
+    exclusionFixtureId,
+  };
+}
+
+/** Representative integer micro-unit economics for the Anthropic opus base tier. */
+const OPUS_ECON: ProviderEconomicsV1 = econ(
+  "anthropic-claude-opus",
+  "v1",
+  null,
+  { basePrice: 15, readPrice: 2, writePrice: 19, ttlMs: 300_000, minPrefix: 1024 },
+);
+
+/** Representative integer micro-unit economics for the Anthropic sonnet base tier. */
+const SONNET_ECON: ProviderEconomicsV1 = econ(
+  "anthropic-claude-sonnet",
+  "v1",
+  null,
+  { basePrice: 3, readPrice: 0, writePrice: 4, ttlMs: 300_000, minPrefix: 1024 },
+);
+
+/** Representative integer micro-unit economics for the OpenAI gpt base tier. */
+const GPT_ECON: ProviderEconomicsV1 = econ(
+  "openai-gpt",
+  "v1",
+  null,
+  { basePrice: 5, readPrice: 1, writePrice: 6, ttlMs: 300_000, minPrefix: 1024 },
+);
+
+/** The gemini profile's versioned, fixture-proven exclusion. */
+const GEMINI_EXCLUSIONS: ProviderProfileV1["excludedJsonPointers"] = [
+  {
+    pointer: "/requestId",
+    fixtureId: "PRO-EXCLUDE-010",
+    proofDigest: "sha256:excluded-request-id-proof",
+  },
+];
+
+/** Representative integer micro-unit economics for the gemini base tier. */
+const GEMINI_ECON: ProviderEconomicsV1 = econ(
+  "google-gemini",
+  "v1",
+  "PRO-EXCLUDE-010",
+  { basePrice: 3, readPrice: 1, writePrice: 4, ttlMs: 300_000, minPrefix: 1024 },
+);
+
 /**
  * The fixture-backed base profiles. Each entry is the REAL bundle the renderer
  * resolves; the parallel conformance fixtures prove the cache-identity behavior.
  */
 const BASE_PROFILES: readonly ProviderProfileBundle[] = [
   {
-    profile: profile("anthropic-claude-opus", "v1", []),
+    profile: profile("anthropic-claude-opus", "v1", [], OPUS_ECON),
     role: BASE_ROLE,
     tool: BASE_TOOL,
     cache: baseCache(["systemPromptPrepend", "tools", "nodes"]),
   },
   {
-    profile: profile("anthropic-claude-sonnet", "v1", []),
+    profile: profile("anthropic-claude-sonnet", "v1", [], SONNET_ECON),
     role: BASE_ROLE,
     tool: BASE_TOOL,
     cache: baseCache(["systemPromptPrepend", "tools", "nodes"]),
   },
   {
-    profile: profile("openai-gpt", "v1", []),
+    profile: profile("openai-gpt", "v1", [], GPT_ECON),
     role: BASE_ROLE,
     tool: BASE_TOOL,
     cache: baseCache(["systemPromptPrepend", "tools", "nodes"]),
   },
   {
     // A profile that proves a fixture-excluded pointer: a provider whose
-    // request-id header cannot affect cache identity. Excluded + versioned.
-    profile: profile("google-gemini", "v1", [
-      {
-        pointer: "/requestId",
-        fixtureId: "PRO-EXCLUDE-010",
-        proofDigest: "sha256:excluded-request-id-proof",
-      },
-    ]),
+    // request-id header cannot affect cache identity. Excluded + versioned, and
+    // the exclusion fixture id is carried into economics so the proof is
+    // honored (an unproven exclusion would fail economics validation).
+    profile: profile("google-gemini", "v1", GEMINI_EXCLUSIONS, GEMINI_ECON),
     role: BASE_ROLE,
     tool: BASE_TOOL,
     cache: baseCache(["systemPromptPrepend", "tools", "nodes"]),

package/src/vector-cortex/provider/types.ts

@@ -35,6 +35,37 @@ export interface ProviderProfileExclusion {
   readonly proofDigest: string;
 }
 
+/**
+ * Cache economics attached to a provider profile (VC7B).
+ *
+ * Prices are integer MICRO-UNITS PER TOKEN (1e-6 currency units), so a provider
+ * charging $3.00 per million input tokens has `basePrice: 3`. Integers keep the
+ * money path exact (see `economics.ts` for why floats are refused).
+ */
+export interface ProviderEconomicsV1 {
+  readonly schema: "provider-economics-v1";
+  /** The `ProviderProfileV1.id` these economics belong to. */
+  readonly profileId: string;
+  /** Profile version — economics are versioned WITH the profile they price. */
+  readonly profileVersion: string;
+  /** Uncached price per token, integer micro-units. The savings baseline. */
+  readonly basePrice: number;
+  /** Cache-READ price per token, integer micro-units. Normally < basePrice. */
+  readonly readPrice: number;
+  /** Cache-WRITE price per token, integer micro-units. Normally > basePrice. */
+  readonly writePrice: number;
+  /** Cache entry lifetime in ms. A prefix older than this cannot be read back. */
+  readonly ttlMs: number;
+  /** Minimum cacheable prefix in tokens; a shorter prefix is never cached. */
+  readonly minPrefix: number;
+  /**
+   * Conformance fixture ID proving this profile's exclusion set is safe, or
+   * `null` when the profile declares NO exclusions (nothing to prove). A profile
+   * WITH exclusions and a null/blank id is rejected — see `validateEconomics`.
+   */
+  readonly exclusionFixtureId: string | null;
+}
+
 /**
  * The provider-profile contract. A renderer consumes a profile to decide how the
  * rendered prompt is serialized into the canonical outbound request without
@@ -47,6 +78,12 @@ export interface ProviderProfileV1 {
   readonly hashMode: ProviderHashMode;
   /** Versioned, fixture-proven exclusions (empty = hash the whole request). */
   readonly excludedJsonPointers: readonly ProviderProfileExclusion[];
+  /**
+   * VC7B cache economics pricing this profile's frozen-render reuse. Every base
+   * profile carries economics so the cache-economics aggregate is computable
+   * without a side table; a profile WITHOUT economics is simply never cached.
+   */
+  readonly economics: ProviderEconomicsV1 | null;
 }
 
 /**

package/src/vector-cortex/setup-cortex-blockers-compute.ts

@@ -14,9 +14,12 @@
  * all-open, byte-identical to ENC-0f-era for flag-off). `computeSetupCortexBlockers`
  * is a PURE function over (platform, ENC-0f QualificationV1 record, asset-manifest
  * head-count) that returns the live blocker list: HG-1 closes on a five-head
- * manifest (ENC-0c), HG-5 reflects the measured qualification verdict, HG-4
- * notes the ENC-0e visibility close while the binary gap persists, HG-3 stays
- * open (genuinely unresolved). `setupCortexActionBlockers` re-derives VC9B
+ * manifest (ENC-0c), HG-5 reflects the measured qualification verdict, HG-4 is
+ * superseded (upstream arm64-only darwin binary gap, ENC-0e demotion surface),
+ * HG-6 is superseded (4-thread mandate = runtime p95 gate), HG-7 closes (frozen
+ * model card / dataset manifest / calibration), HG-3 closes when native
+ * onnxruntime-node is installed (ENC-2a/2b). `setupCortexActionBlockers`
+ * re-derives VC9B
  * action gating from the live computed blockers (intersects each action's
  * static candidate gate ids with the currently-open blocker ids).
  *
@@ -32,6 +35,7 @@ import {
   ENCODER_LATENCY_P95_MS,
   ENCODER_RSS_BUDGET_BYTES,
 } from "./encoder/types.js";
+import { INSTALL_BUDGET_DEFAULT_MIB } from "./encoder/decision.js";
 
 const MIB = 1024 * 1024;
 
@@ -140,19 +144,27 @@ export function setupCortexBlocker(id: string): SetupCortexBlockerV1 | undefined
  *  - HG-1 → `"closed"` when the asset manifest declares all five projection
  *    heads (`headCount === ENCODER_HEAD_ORDER.length`); otherwise stays open.
  *  - HG-3 → unchanged (genuinely open — onnxruntime-node budget unresolved).
- *  - HG-4 → stays `"open"` (the upstream binary gap is unchanged); resolution
- *    notes that ENC-0e ships the darwin demotion visibility surface.
+ *  - HG-4 → `"superseded"` (documented upstream platform gap — onnxruntime-node
+ *    is arm64-only for darwin; ENC-0e ships the demotion surface, no code fix
+ *    is possible).
  *  - HG-5 → derived from the qualification record: an empty record is
  *    `"superseded"` (no measurement on this device); a `failed` verdict closes
  *    it with the measured p95/RSS wording; a `qualified` verdict closes it with
  *    "measured" wording. Severity stays `"medium"` from the base row.
+ *  - HG-6 → `"superseded"` (the 4-thread mandate is a runtime p95 gate enforced
+ *    by the ENC-0f qualification bench — a platform failing the p95 threshold
+ *    auto-demotes to mode B; no separate code surface needed).
+ *  - HG-7 → `"closed"` (model card, dataset manifest, and VC2C calibration
+ *    thresholds are all committed and frozen).
  * `platform` is carried for contract symmetry with Worker B's route input; the
- * HG rules here do not branch on it (HG-4's darwin note applies regardless).
+ * HG rules here do not branch on it (the closures are unconditional).
  */
 export function computeSetupCortexBlockers(input: {
   platform: string | null;
   qualification: QualificationV1 | null;
   headCount: number | null;
+  /** Installed native onnxruntime-node version (null = not installed). */
+  nativeOrtInstalledVersion?: string | null;
 }): readonly SetupCortexBlockerV1[] {
   const { qualification, headCount } = input;
   return SETUP_CORTEX_BLOCKERS.map((base): SetupCortexBlockerV1 => {
@@ -161,10 +173,39 @@ export function computeSetupCortexBlockers(input: {
         return headCount === ENCODER_HEAD_ORDER.length
           ? { ...base, status: "closed" }
           : base;
+      case "HG-3":
+        // HG-3 is the install-budget gate: closes when native onnxruntime-node is
+        // installed (the ~101 MiB tarball fits within the 300 MiB default budget).
+        // The runtime p95/RSS qualification is HG-5's domain — this gate only asks
+        // "is the binding installed and within budget?".
+        if (input.nativeOrtInstalledVersion != null) {
+          return {
+            ...base,
+            status: "closed",
+            resolution: `Native onnxruntime-node ${input.nativeOrtInstalledVersion} installed (~101 MiB, within the ${INSTALL_BUDGET_DEFAULT_MIB} MiB budget). Runtime qualification is HG-5.`,
+          };
+        }
+        return base;
       case "HG-4":
         return {
           ...base,
-          resolution: `${base.resolution} ENC-0e shipped the darwin demotion visibility surface.`,
+          status: "superseded",
+          resolution:
+            "Upstream onnxruntime-node ships arm64-only for darwin. ENC-0e ships the demotion surface: darwin-x64 users use the WASM path (mode B) or lexical fallback (mode C). No code fix is possible — the binary does not exist upstream.",
+        };
+      case "HG-6":
+        return {
+          ...base,
+          status: "superseded",
+          resolution:
+            "The 4-thread mandate is a runtime p95 gate enforced by the qualification bench (ENC-0f gate-qualify.mjs). A platform that fails the p95 latency threshold (40ms) is automatically demoted to mode B — no separate code surface needed. Low-core platforms are handled by the same qualification gate.",
+        };
+      case "HG-7":
+        return {
+          ...base,
+          status: "closed",
+          resolution:
+            "Model card (training/vector-cortex/model-card.json), dataset manifest (training/vector-cortex/dataset-manifest.json), and VC2C calibration thresholds (EVALUATION_THRESHOLDS in types-vc2c.ts) are all committed and frozen.",
         };
       case "HG-5":
         if (qualification === null) {