@sanity/ailf 4.1.0 → 4.3.0

package/config/package-surface.ts

@@ -0,0 +1,37 @@
+/**
+ * package-surface.ts — Framework-level package-surface manifest for the
+ * W0198 symbol-resolution preflight.
+ *
+ * The manifest pins each in-scope package to a single semver-major range.
+ * The preflight resolver answers "does symbol X export from package Y" by
+ * reading the installed package's `.d.ts` against this pin. Tasks
+ * reference packages by name; they do not carry per-package version
+ * metadata (per-task overrides remain a future extension point).
+ *
+ * Bumping a major is an editorial event — one PR that updates the pin,
+ * regenerates cached surfaces, and re-runs the historical comparison set.
+ * Patch and minor releases within a pinned major flow silently because
+ * semver disallows the export removals that would change a deduction
+ * outcome.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+
+import { definePackageSurface } from "@sanity/ailf-core"
+
+export default definePackageSurface({
+  packages: [
+    // Sanity App SDK — drives App SDK literacy tasks (e.g. DOC-2117).
+    { pkg: "@sanity/sdk-react", semverPin: "^2.0.0" },
+
+    // Sanity Studio runtime — drives Studio-side literacy tasks.
+    // Pinned to the installed major in this repo (^5.x). The earlier
+    // major (v3) is an explicit per-task override candidate when a task
+    // exercises legacy Studio behavior.
+    { pkg: "sanity", semverPin: "^5.0.0" },
+
+    // @sanity/client — drives direct-client literacy tasks. Pinned to
+    // the installed major in this repo (^7.x).
+    { pkg: "@sanity/client", semverPin: "^7.0.0" },
+  ],
+})

package/config/preflight-scoring.ts

@@ -0,0 +1,26 @@
+/**
+ * preflight-scoring.ts — How heavily the W0198 deterministic preflight
+ * contributes to the `code-correctness` dimension.
+ *
+ * The preflight (`SymbolPreflightReport`) and the LLM rubric both feed
+ * into `code-correctness` per D0010's weighted dimension aggregation.
+ * `codeCorrectnessWeight` sets the relative share between them — `0.4`
+ * means preflight is 40% of the dimension, rubric 60%.
+ *
+ * Bumping this is an editorial decision: a higher weight ties more of the
+ * `code-correctness` score to the deterministic existence-check (less
+ * grader noise on the symbol-existence question, but also less elasticity
+ * for the rubric to penalize stylistic/correctness issues the preflight
+ * cannot see). A lower weight cedes more authority back to the rubric.
+ *
+ * The default of `0.4` is a starting balance; revisit once the
+ * `unresolved` rate stabilizes in CI.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+
+import { definePreflightScoring } from "@sanity/ailf-core"
+
+export default definePreflightScoring({
+  codeCorrectnessWeight: 0.4,
+})

package/dist/_vendor/ailf-core/artifact-registry.d.ts

@@ -41,7 +41,7 @@ export type ArtifactMime = "application/json" | "application/x-ndjson" | "text/m
  */
 export type ArtifactTruncationPolicy = "reject" | "trailing-truncate" | "fielded-truncate" | "trial-oversize";
 /** The union of every artifact type known to AILF. */
-export type ArtifactType = "runManifest" | "scoreSummary" | "pipelineResult" | "pipelineContext" | "documentManifest" | "prComment" | "readinessReport" | "reportSnapshot" | "autoComparison" | "gapReport" | "sinkResults" | "callbackRequest" | "callbackResponse" | "configSnapshot" | "evalConfigGenerated" | "comparisonReport" | "discoveryReport" | "failureModes" | "renderedPrompts" | "rawResults" | "testOutputs" | "graderPrompts" | "graderJudgments" | "traces";
+export type ArtifactType = "runManifest" | "scoreSummary" | "pipelineResult" | "pipelineContext" | "documentManifest" | "prComment" | "readinessReport" | "reportSnapshot" | "autoComparison" | "gapReport" | "sinkResults" | "callbackRequest" | "callbackResponse" | "configSnapshot" | "evalConfigGenerated" | "comparisonReport" | "discoveryReport" | "failureModes" | "renderedPrompts" | "rawResults" | "testOutputs" | "graderPrompts" | "graderJudgments" | "symbolPreflight" | "traces";
 /**
  * Result of parsing a per-entry key into a sanitized filename component.
  * Success carries the sanitized value; failure carries a reason for 4xx responses.

package/dist/_vendor/ailf-core/artifact-registry.js

@@ -241,6 +241,21 @@ const failureModePreviewSchema = z.object({
     severity: z.enum(["low", "medium", "high", "critical"]),
     titlePreview: z.string().max(120),
 });
+/**
+ * Preview shape for `symbolPreflight` manifest entries (W0198 Phase 4).
+ *
+ * The deterministic-lane report can be tens of KB on candidates that
+ * import many bindings. The Studio rollup row only needs the three lane
+ * counts and the deduction total to flag whether a deduction was applied
+ * and how much — drill-down hydrates the full per-binding finding list
+ * from the external artifact.
+ */
+const symbolPreflightPreviewSchema = z.object({
+    exists: z.number().int().nonnegative(),
+    missing: z.number().int().nonnegative(),
+    unresolved: z.number().int().nonnegative(),
+    deductionTotal: z.number().nonnegative(),
+});
 /**
  * Preview shape for `graderJudgments` manifest entries (W0051 / D0033 M7).
  * List views render score + a short reason excerpt; drill-down hydrates
@@ -875,6 +890,38 @@ export const ARTIFACT_REGISTRY = {
             capBytes: 512,
         },
     }),
+    symbolPreflight: buildDescriptor({
+        type: "symbolPreflight",
+        slug: "symbol-preflight",
+        layout: "per-entry",
+        axes: ["run", "mode", "task", "model"],
+        entrySchema: unknownEntry,
+        mime: "application/json",
+        capBytes: 256_000,
+        manifestPreview: {
+            schema: symbolPreflightPreviewSchema,
+            extract: (entry) => {
+                const e = entry;
+                const findings = Array.isArray(e.findings) ? e.findings : [];
+                let exists = 0;
+                let missing = 0;
+                let unresolved = 0;
+                for (const f of findings) {
+                    if (f && typeof f === "object" && "result" in f) {
+                        if (f.result === "exists")
+                            exists++;
+                        else if (f.result === "missing")
+                            missing++;
+                        else if (f.result === "unresolved")
+                            unresolved++;
+                    }
+                }
+                const deductionTotal = typeof e.deduction?.total === "number" ? e.deduction.total : 0;
+                return { exists, missing, unresolved, deductionTotal };
+            },
+            capBytes: 96,
+        },
+    }),
     traces: buildDescriptor({
         type: "traces",
         slug: "traces",

package/dist/_vendor/ailf-core/config-helpers.d.ts

@@ -33,6 +33,8 @@ import type { SinksFile } from "./schemas/sinks.js";
 import type { TestBudgetConfig } from "./schemas/test-budgets.js";
 import type { ModelsConfig } from "./types/index.js";
 import type { GeneralizedTaskDefinition } from "./types/generalized-task.js";
+import type { PackageSurfaceConfig } from "./types/package-surface.js";
+import type { PreflightScoringConfig } from "./types/preflight-scoring.js";
 import type { ModeBase, PresetDefinition } from "./types/plugin-registry.js";
 /**
  * Define an AILF evaluation configuration.
@@ -67,6 +69,39 @@ export declare function defineTask(task: GeneralizedTaskDefinition): Generalized
  *                 misconfigured OpenAI Responses-API fields.
  */
 export declare function defineModels(models: ModelsConfig): ModelsConfig;
+/**
+ * Define the framework-level package-surface manifest consumed by the
+ * W0198 symbol-resolution preflight.
+ *
+ * Validates the obviously author-time errors:
+ * - Each entry has a non-empty `pkg` and `semverPin`.
+ * - `semverPin` is a single-major caret range (`^MAJOR.MINOR.PATCH`).
+ *   Other range syntaxes (`>=`, `~`, `*`) are rejected because the
+ *   resolver's stability commitment is "pinned to one major"; non-caret
+ *   ranges either widen past a major (reintroducing churn) or are
+ *   stricter than needed for an editorial pin.
+ * - No duplicate `pkg` entries.
+ *
+ * @throws {Error} On invalid entries or duplicates.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+export declare function definePackageSurface(config: PackageSurfaceConfig): PackageSurfaceConfig;
+/**
+ * Define the preflight-scoring config consumed by the W0198 two-stage
+ * grader's score merge.
+ *
+ * The single knob is `codeCorrectnessWeight` — preflight's share of the
+ * `code-correctness` dimension (`[0, 1]`). The complementary share belongs
+ * to the LLM rubric. Validates the obviously author-time errors:
+ *
+ * - `codeCorrectnessWeight` is a finite number in `[0, 1]`. Weights outside
+ *   that range either disable one lane entirely (use `0` / `1`) or land in
+ *   nonsense territory (negative, > 1).
+ *
+ * @throws {Error} On out-of-range weights.
+ */
+export declare function definePreflightScoring(config: PreflightScoringConfig): PreflightScoringConfig;
 /**
  * Define rubric templates and scoring profiles.
  *

package/dist/_vendor/ailf-core/config-helpers.js

@@ -144,6 +144,73 @@ export function defineModels(models) {
     return models;
 }
 // ---------------------------------------------------------------------------
+// Package-surface helpers
+// ---------------------------------------------------------------------------
+/**
+ * Define the framework-level package-surface manifest consumed by the
+ * W0198 symbol-resolution preflight.
+ *
+ * Validates the obviously author-time errors:
+ * - Each entry has a non-empty `pkg` and `semverPin`.
+ * - `semverPin` is a single-major caret range (`^MAJOR.MINOR.PATCH`).
+ *   Other range syntaxes (`>=`, `~`, `*`) are rejected because the
+ *   resolver's stability commitment is "pinned to one major"; non-caret
+ *   ranges either widen past a major (reintroducing churn) or are
+ *   stricter than needed for an editorial pin.
+ * - No duplicate `pkg` entries.
+ *
+ * @throws {Error} On invalid entries or duplicates.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+export function definePackageSurface(config) {
+    const seen = new Set();
+    for (const entry of config.packages) {
+        if (!entry.pkg) {
+            throw new Error(`definePackageSurface: every entry must declare a non-empty \`pkg\`.`);
+        }
+        if (!entry.semverPin) {
+            throw new Error(`definePackageSurface: entry "${entry.pkg}" is missing \`semverPin\`.`);
+        }
+        if (!/^\^\d+\.\d+\.\d+(?:[-+].*)?$/.test(entry.semverPin)) {
+            throw new Error(`definePackageSurface: entry "${entry.pkg}" has \`semverPin\` ` +
+                `"${entry.semverPin}". Pins must be single-major caret ranges ` +
+                `(e.g. "^2.0.0"); other range syntaxes are not allowed.`);
+        }
+        if (seen.has(entry.pkg)) {
+            throw new Error(`definePackageSurface: duplicate entry for "${entry.pkg}". ` +
+                `Each package may appear at most once in the manifest.`);
+        }
+        seen.add(entry.pkg);
+    }
+    return config;
+}
+// ---------------------------------------------------------------------------
+// Preflight-scoring helpers
+// ---------------------------------------------------------------------------
+/**
+ * Define the preflight-scoring config consumed by the W0198 two-stage
+ * grader's score merge.
+ *
+ * The single knob is `codeCorrectnessWeight` — preflight's share of the
+ * `code-correctness` dimension (`[0, 1]`). The complementary share belongs
+ * to the LLM rubric. Validates the obviously author-time errors:
+ *
+ * - `codeCorrectnessWeight` is a finite number in `[0, 1]`. Weights outside
+ *   that range either disable one lane entirely (use `0` / `1`) or land in
+ *   nonsense territory (negative, > 1).
+ *
+ * @throws {Error} On out-of-range weights.
+ */
+export function definePreflightScoring(config) {
+    const w = config.codeCorrectnessWeight;
+    if (!Number.isFinite(w) || w < 0 || w > 1) {
+        throw new Error(`definePreflightScoring: \`codeCorrectnessWeight\` must be a finite ` +
+            `number in [0, 1] — got ${String(w)}.`);
+    }
+    return config;
+}
+// ---------------------------------------------------------------------------
 // Rubric helpers
 // ---------------------------------------------------------------------------
 /**

package/dist/_vendor/ailf-core/index.d.ts

@@ -18,7 +18,7 @@ export * from "./examples/index.js";
 export * from "./artifact-registry.js";
 export * from "./batch-signing.js";
 export * from "./constants.js";
-export { defineCanaryTasks, defineConfig, defineFeatures, defineModeBase, defineModels, definePricingTable, definePreset, definePrompts, defineRubrics, defineSchedules, defineSinks, defineSources, defineTask, defineTestBudgets, defineThresholds, } from "./config-helpers.js";
+export { defineCanaryTasks, defineConfig, defineFeatures, defineModeBase, defineModels, definePackageSurface, definePreflightScoring, definePricingTable, definePreset, definePrompts, defineRubrics, defineSchedules, defineSinks, defineSources, defineTask, defineTestBudgets, defineThresholds, } from "./config-helpers.js";
 export type { PricingEntry, PromptEntry, SourceEntry, } from "./config-helpers.js";
 export { env } from "./env-helper.js";
 export { NoOpArtifactWriter, NotImplementedError, } from "./ports/artifact-writer.js";

package/dist/_vendor/ailf-core/index.js

@@ -21,7 +21,7 @@ export * from "./constants.js";
 // ---------------------------------------------------------------------------
 // Architecture overhaul — Phase 0 helpers
 // ---------------------------------------------------------------------------
-export { defineCanaryTasks, defineConfig, defineFeatures, defineModeBase, defineModels, definePricingTable, definePreset, definePrompts, defineRubrics, defineSchedules, defineSinks, defineSources, defineTask, defineTestBudgets, defineThresholds, } from "./config-helpers.js";
+export { defineCanaryTasks, defineConfig, defineFeatures, defineModeBase, defineModels, definePackageSurface, definePreflightScoring, definePricingTable, definePreset, definePrompts, defineRubrics, defineSchedules, defineSinks, defineSources, defineTask, defineTestBudgets, defineThresholds, } from "./config-helpers.js";
 export { env } from "./env-helper.js";
 export { NoOpArtifactWriter, NotImplementedError, } from "./ports/artifact-writer.js";
 export { assoc, resolveVariantMode, splitTaskVariant, } from "./artifact-capture/association.js";

package/dist/_vendor/ailf-core/ports/context.d.ts

@@ -19,6 +19,7 @@ import type { CacheStore } from "./cache-store.js";
 import type { DocFetcher } from "./doc-fetcher.js";
 import type { EvalRunner } from "./eval-runner.js";
 import type { Logger } from "./logger.js";
+import type { PackageSurfaceResolver } from "./package-surface-resolver.js";
 import type { ProgressReporter } from "./progress-reporter.js";
 import type { TaskSource } from "./task-source.js";
 /**
@@ -78,6 +79,16 @@ export interface ResolvedConfig {
     noRemoteCache: boolean;
     /** Grader replications for consistency measurement */
     graderReplications?: number;
+    /**
+     * Grader context policy. Controls whether canonical reference docs are
+     * injected into the grader's assertion `rubricPrompt`. `"rubric-only"`
+     * sends only the rubric template + criteria; `"with-docs"` adds the
+     * canonical doc content as authoritative ground truth.
+     *
+     * Sourced from EvalConfig `grader.context` or the equivalent CLI/env
+     * surface. Defaults to `"rubric-only"` at the EvalConfig boundary.
+     */
+    graderContext?: "rubric-only" | "with-docs";
     /** Base directory for user-facing pipeline output artifacts. */
     outputDir: string;
     /** Output path override */
@@ -237,6 +248,13 @@ export interface AppContext {
     readonly evalRunner: EvalRunner;
     /** Structured logger */
     readonly logger: Logger;
+    /**
+     * Package-surface resolver for the W0198 symbol-resolution preflight.
+     * Optional during the staged rollout — Phase 1 wires the adapter, later
+     * phases consume it. When unset, the preflight step (Phase 4+) treats
+     * every binding as `unresolved` so candidates are not penalized.
+     */
+    readonly packageSurfaceResolver?: PackageSurfaceResolver;
     /**
      * Progress reporter — carries `phase-start / phase-progress / phase-complete`
      * events for long-running pipeline spans (W0053). The composition root always

package/dist/_vendor/ailf-core/ports/doc-fetcher.d.ts

@@ -86,6 +86,30 @@ export interface UrlFetchSummary {
     totalFailed: number;
     totalFetched: number;
 }
+/**
+ * Per-task entry in the symbol-reference index manifest (W0197).
+ *
+ * One entry is emitted per task that the fetcher processed, regardless of
+ * whether extraction yielded any symbols. The grader-context consumer
+ * uses `symbolCount > 0` to decide between symbol-index and full-doc
+ * injection — distinguishing "extraction ran and produced nothing"
+ * (entry present, count zero — observable fallback) from "fetcher hadn't
+ * run with W0197" (no entry — silent fallback).
+ */
+export interface SymbolIndexManifestEntry {
+    taskId: string;
+    /** Repo-relative path to the per-task index artifact. */
+    path: string;
+    /** Total deduped symbols in the index (post-merge). */
+    symbolCount: number;
+    /** Per-tier breakdown for observability. */
+    tierBreakdown: {
+        typeDef: number;
+        heading: number;
+        inlineCode: number;
+        codeBlock: number;
+    };
+}
 /** Metadata about the fetch operation, for downstream pipeline consumption */
 export interface FetchMetadata {
     /** Document manifest for traceability (slug, _id, _rev, title) */
@@ -96,6 +120,12 @@ export interface FetchMetadata {
     documentOverlay?: DocumentOverlaySummary;
     /** URL fetch operations summary */
     urlFetch?: UrlFetchSummary;
+    /**
+     * Per-task symbol-index manifest (W0197). Always populated when the
+     * fetcher emits per-task index artifacts; the in-memory copy mirrors
+     * the on-disk `contexts/canonical-symbols/manifest.json`.
+     */
+    symbolIndexes?: SymbolIndexManifestEntry[];
 }
 /** Complete result of a doc fetch operation */
 export interface FetchResult {

package/dist/_vendor/ailf-core/ports/index.d.ts

@@ -9,10 +9,12 @@ export { NoOpArtifactWriter } from "./artifact-writer.js";
 export type { CacheEntryMetadata, CacheKey, CacheLookupResult, CacheRecordInput, CacheStore, } from "./cache-store.js";
 export type { ConfigSource } from "./config-source.js";
 export type { AppContext, ReportSinkPort, ReportStorePort, ResolvedConfig, } from "./context.js";
-export type { DocContext, DocFetcher, DocSourceConfig, DocumentManifestEntry, DocumentOverlaySummary, FetchMetadata, FetchResult, ReleaseImpact, UrlFetchEntry, UrlFetchSummary, } from "./doc-fetcher.js";
+export type { DocContext, DocFetcher, DocSourceConfig, DocumentManifestEntry, DocumentOverlaySummary, FetchMetadata, FetchResult, ReleaseImpact, SymbolIndexManifestEntry, UrlFetchEntry, UrlFetchSummary, } from "./doc-fetcher.js";
 export type { EvalRunConfig, EvalRunner } from "./eval-runner.js";
 export type { CompilationContext, CompileResultAssertion, CompileResultPrompt, CompileResultProvider, CompileResultTestCase, ModeCompileResult, ModeHandler, ModeProviderEntry, ModeRubricConfig, PromptTemplate, } from "./mode-handler.js";
 export type { Logger } from "./logger.js";
+export type { PackageSurface, PackageSurfaceResolver, PackageSurfaceSymbol, PackageSurfaceUnresolvedReason, } from "./package-surface-resolver.js";
+export { PackageSurfaceResolverError } from "./package-surface-resolver.js";
 export type { PipelineStep } from "./pipeline-step.js";
 export type { ArtifactWriterProgressOptions, PhaseCompleteEvent, PhaseProgressEvent, PhaseStartEvent, ProgressReporter, } from "./progress-reporter.js";
 export { ARTIFACT_EXPORT_PHASE_ID, NoOpProgressReporter, } from "./progress-reporter.js";

package/dist/_vendor/ailf-core/ports/index.js

@@ -5,5 +5,6 @@
  * Adapters (in packages/eval) implement these interfaces.
  */
 export { NoOpArtifactWriter } from "./artifact-writer.js";
+export { PackageSurfaceResolverError } from "./package-surface-resolver.js";
 export { ARTIFACT_EXPORT_PHASE_ID, NoOpProgressReporter, } from "./progress-reporter.js";
 export { canonicalDocRefLabel, isIdRef, isPathRef, isPerspectiveRef, isSlugRef, isTemplatedAssertion, } from "./task-source.js";

package/dist/_vendor/ailf-core/ports/mode-handler.d.ts

@@ -17,6 +17,7 @@
  * @see packages/eval/src/pipeline/compiler/mode-handlers/
  */
 import type { GeneralizedTaskDefinition } from "../types/generalized-task.js";
+import type { PreflightRubricContext } from "../types/preflight-scoring.js";
 /**
  * A prompt template owned by a mode handler.
  *
@@ -44,6 +45,28 @@ export interface CompilationContext {
     models?: ModeProviderEntry[];
     /** Rubric config (templates, weights) — loaded from config/rubrics */
     rubricConfig?: ModeRubricConfig;
+    /**
+     * Grader context policy.
+     *
+     * - `"rubric-only"`: the grader sees only the rubric template + criteria +
+     *   the candidate's response.
+     * - `"with-docs"`: the canonical reference content for the task is injected
+     *   into the assertion's `rubricPrompt` so the grader has authoritative
+     *   ground truth alongside the rubric.
+     *
+     * Mode handlers without canonical doc context (agent-harness, knowledge-probe)
+     * ignore this field — the `canonicalReference` parameter on
+     * `resolveTemplatedAssertion` is optional, so unset means current behavior.
+     */
+    graderContext?: "rubric-only" | "with-docs";
+    /**
+     * W0198 Phase 6 — when supplied, mode handlers prefix the
+     * `code-correctness` rubric with a "DETERMINISTIC PREFLIGHT" system
+     * instruction telling the grader to treat the preflight's existence
+     * findings as ground truth. Sourced upstream from the package-surface
+     * manifest; absence collapses cleanly to the pre-W0198 rubric.
+     */
+    preflightContext?: PreflightRubricContext;
 }
 /** A model provider entry for compilation */
 export interface ModeProviderEntry {

package/dist/_vendor/ailf-core/ports/package-surface-resolver.d.ts

@@ -0,0 +1,71 @@
+/**
+ * Port: Reads the public symbol surface of an installed package.
+ *
+ * The W0198 symbol-resolution preflight resolves candidate imports against
+ * the actual package surface (its `.d.ts`) to answer "does symbol X export
+ * from package Y" deterministically. This port lets the preflight do that
+ * without depending on a specific resolution mechanism — adapters decide
+ * whether to read installed `node_modules`, an in-memory map (tests), or a
+ * future cached snapshot.
+ *
+ * Adapters:
+ * - `DtsPackageSurface` (`packages/eval/src/adapters/package-surface/`) —
+ *   reads installed `.d.ts` from `node_modules/<pkg>`, follows one hop of
+ *   `export * from "./other"`.
+ * - `InMemoryPackageSurface` — test double; maps package names to a fixed
+ *   surface for deterministic unit tests.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+/**
+ * One exported binding from a package's public surface.
+ *
+ * - `name` — the bare exported identifier (e.g. `"useEditDocument"`).
+ * - `source` — provenance of the binding within the package:
+ *   - `"types"` — declared in a `.d.ts` (the only source the
+ *     `DtsPackageSurface` adapter currently emits).
+ *   - `"runtime"` — reserved for future adapters that introspect the
+ *     loaded module's runtime exports.
+ */
+export interface PackageSurfaceSymbol {
+    readonly name: string;
+    readonly source: "types" | "runtime";
+}
+/**
+ * Resolved public surface of a single package at a single resolved version.
+ *
+ * - `pkg` — npm package name as the caller asked.
+ * - `version` — version actually resolved on disk (from the package's
+ *   `package.json`). May differ from a manifest pin's literal range; the
+ *   pin says "any major-N", and this is the concrete version present.
+ * - `symbols` — every top-level exported binding the adapter discovered.
+ *   Order is implementation-defined (callers should not rely on it).
+ */
+export interface PackageSurface {
+    readonly pkg: string;
+    readonly version: string;
+    readonly symbols: readonly PackageSurfaceSymbol[];
+}
+/**
+ * Adapters throw a `PackageSurfaceResolverError` when they cannot answer
+ * a `resolveExports` call. The `reason` field maps directly to the
+ * `unresolved` finding kinds the W0198 preflight surfaces — callers
+ * catch this error and convert it into a per-binding `unresolved`
+ * finding rather than a `missing` deduction.
+ */
+export declare class PackageSurfaceResolverError extends Error {
+    readonly pkg: string;
+    readonly reason: PackageSurfaceUnresolvedReason;
+    constructor(reason: PackageSurfaceUnresolvedReason, pkg: string, message: string);
+}
+export type PackageSurfaceUnresolvedReason = "package-not-installed" | "types-entry-missing" | "parse-failed";
+/**
+ * Reads the public symbol surface of an installed package.
+ *
+ * Adapters MAY cache results within a single resolver instance — same
+ * `(pkg, version)` should yield the same `PackageSurface` for the
+ * lifetime of the resolver.
+ */
+export interface PackageSurfaceResolver {
+    resolveExports(pkg: string): Promise<PackageSurface>;
+}

package/dist/_vendor/ailf-core/ports/package-surface-resolver.js

@@ -0,0 +1,36 @@
+/**
+ * Port: Reads the public symbol surface of an installed package.
+ *
+ * The W0198 symbol-resolution preflight resolves candidate imports against
+ * the actual package surface (its `.d.ts`) to answer "does symbol X export
+ * from package Y" deterministically. This port lets the preflight do that
+ * without depending on a specific resolution mechanism — adapters decide
+ * whether to read installed `node_modules`, an in-memory map (tests), or a
+ * future cached snapshot.
+ *
+ * Adapters:
+ * - `DtsPackageSurface` (`packages/eval/src/adapters/package-surface/`) —
+ *   reads installed `.d.ts` from `node_modules/<pkg>`, follows one hop of
+ *   `export * from "./other"`.
+ * - `InMemoryPackageSurface` — test double; maps package names to a fixed
+ *   surface for deterministic unit tests.
+ *
+ * @see docs/design-docs/two-stage-grader-symbol-preflight.md
+ */
+/**
+ * Adapters throw a `PackageSurfaceResolverError` when they cannot answer
+ * a `resolveExports` call. The `reason` field maps directly to the
+ * `unresolved` finding kinds the W0198 preflight surfaces — callers
+ * catch this error and convert it into a per-binding `unresolved`
+ * finding rather than a `missing` deduction.
+ */
+export class PackageSurfaceResolverError extends Error {
+    pkg;
+    reason;
+    constructor(reason, pkg, message) {
+        super(message);
+        this.name = "PackageSurfaceResolverError";
+        this.pkg = pkg;
+        this.reason = reason;
+    }
+}

package/dist/_vendor/ailf-core/schemas/eval-config.d.ts

@@ -44,6 +44,12 @@ export declare const EvalConfigSchema: z.ZodObject<{
         gapAnalysis: z.ZodOptional<z.ZodBoolean>;
         apiUrl: z.ZodOptional<z.ZodString>;
     }, z.core.$strip>>;
+    grader: z.ZodOptional<z.ZodObject<{
+        context: z.ZodOptional<z.ZodEnum<{
+            "rubric-only": "rubric-only";
+            "with-docs": "with-docs";
+        }>>;
+    }, z.core.$strip>>;
     output: z.ZodOptional<z.ZodObject<{
         dir: z.ZodOptional<z.ZodString>;
     }, z.core.$strip>>;

package/dist/_vendor/ailf-core/schemas/eval-config.js

@@ -98,6 +98,20 @@ export const EvalConfigSchema = z
         apiUrl: z.string().url().optional(),
     })
         .optional(),
+    /**
+     * Grader configuration.
+     *
+     * - `context: "rubric-only"` — the grader sees only the rubric template
+     *   + criteria + candidate response.
+     * - `context: "with-docs"` — canonical reference content for each task
+     *   is injected into the assertion's `rubricPrompt` so the grader has
+     *   authoritative ground truth.
+     */
+    grader: z
+        .object({
+        context: z.enum(["rubric-only", "with-docs"]).optional(),
+    })
+        .optional(),
     /**
      * Output configuration (W0077 Phase 6c). Replaces the retired
      * `--output-dir` CLI flag. Path is resolved relative to the caller's

package/dist/_vendor/ailf-core/schemas/index.d.ts

@@ -17,4 +17,5 @@ export * from "./pipeline.js";
 export * from "./report.js";
 export * from "./schedules.js";
 export * from "./sinks.js";
+export * from "./symbol-preflight-report.js";
 export * from "./test-budgets.js";

package/dist/_vendor/ailf-core/schemas/index.js

@@ -17,4 +17,5 @@ export * from "./pipeline.js";
 export * from "./report.js";
 export * from "./schedules.js";
 export * from "./sinks.js";
+export * from "./symbol-preflight-report.js";
 export * from "./test-budgets.js";

package/dist/_vendor/ailf-core/schemas/symbol-preflight-report.d.ts

@@ -0,0 +1,51 @@
+/**
+ * @sanity/ailf-core — SymbolPreflightReport schema
+ *
+ * The trust-boundary parser for the W0198 deterministic-lane artifact.
+ * Used when reading a previously-emitted preflight report back from
+ * disk / GCS so a downstream step (re-grading, comparison, dashboard
+ * rendering) can trust its shape.
+ *
+ * The schema asserts `satisfies z.ZodType<SymbolPreflightReport>`
+ * against the domain type in
+ * `packages/core/src/types/symbol-preflight-report.ts` (D0045 / W0187),
+ * so any drift between the two is a build error.
+ */
+import { z } from "zod";
+export declare const SymbolPreflightReportSchema: z.ZodObject<{
+    candidate: z.ZodObject<{
+        taskId: z.ZodString;
+        testIndex: z.ZodNumber;
+    }, z.core.$strip>;
+    findings: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
+        result: z.ZodLiteral<"exists">;
+        pkg: z.ZodString;
+        version: z.ZodString;
+        binding: z.ZodString;
+        source: z.ZodEnum<{
+            types: "types";
+            runtime: "runtime";
+        }>;
+    }, z.core.$strip>, z.ZodObject<{
+        result: z.ZodLiteral<"missing">;
+        pkg: z.ZodString;
+        version: z.ZodString;
+        binding: z.ZodString;
+    }, z.core.$strip>, z.ZodObject<{
+        result: z.ZodLiteral<"unresolved">;
+        pkg: z.ZodString;
+        binding: z.ZodString;
+        reason: z.ZodEnum<{
+            "package-not-installed": "package-not-installed";
+            "types-entry-missing": "types-entry-missing";
+            "parse-failed": "parse-failed";
+            "reexport-hop-unfollowed": "reexport-hop-unfollowed";
+        }>;
+    }, z.core.$strip>], "result">>;
+    deduction: z.ZodObject<{
+        perMissing: z.ZodNumber;
+        cap: z.ZodNumber;
+        total: z.ZodNumber;
+    }, z.core.$strip>;
+}, z.core.$strip>;
+export type { SymbolPreflightReport } from "../types/symbol-preflight-report.js";