feat(pdf): extract standalone PDF ingest package (ESRC-WP-0001)

Bootstrap evidence-source from the citation-evidence src/source PDF slice.

- Headless core (src/pdf): ingest/extract/fingerprint, importing domain
  contracts from @citation-evidence/engine/shared (no local copies)
- Browser upload helpers isolated under src/browser behind a ./browser
  entry point, with an eslint boundary keeping the core browser-free
- pnpm/TS/vitest/eslint scaffold; 52 tests (contract + determinism)
- Fixtures resolved from the sibling citation-evidence checkout, not
  duplicated (real PII) — see docs/ADR-0002; suites skip when absent
- Boundary + fixture decisions recorded as docs/ADR-0001 / ADR-0002
- README/SCOPE rewritten; capability.infotech.pdf-evidence-ingest
  registered, NO_CAPABILITIES removed
- Follow-on workplans ESRC-WP-0002..0004 queued; ESRC-WP-0001 finished

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/browser/byte-store.test.ts

@@ -0,0 +1,97 @@
+import { describe, expect, it, vi } from "vitest";
+
+import type { DocumentId } from "@citation-evidence/engine/shared";
+
+import { createPdfByteStore } from "./byte-store";
+
+function stubUrlHelpers() {
+  let counter = 0;
+  const created: string[] = [];
+  const revoked: string[] = [];
+  const createObjectURL = vi.fn(() => {
+    const url = `blob:stub-${++counter}`;
+    created.push(url);
+    return url;
+  });
+  const revokeObjectURL = vi.fn((url: string) => {
+    revoked.push(url);
+  });
+  return { createObjectURL, revokeObjectURL, created, revoked };
+}
+
+describe("PdfByteStore", () => {
+  it("put / get round-trips bytes and exposes a blob URL", () => {
+    const helpers = stubUrlHelpers();
+    const store = createPdfByteStore(helpers);
+    const bytes = new Uint8Array([0x25, 0x50, 0x44, 0x46]); // %PDF
+    const record = store.put("doc_a" as DocumentId, bytes);
+    expect(record.blobUrl).toBe("blob:stub-1");
+    expect(store.get("doc_a" as DocumentId)?.bytes).toBe(bytes);
+    expect(store.has("doc_a" as DocumentId)).toBe(true);
+    expect(store.list()).toEqual(["doc_a"]);
+    expect(store.size()).toBe(4);
+  });
+
+  it("put replaces an existing entry and revokes the old URL", () => {
+    const helpers = stubUrlHelpers();
+    const store = createPdfByteStore(helpers);
+    const id = "doc_a" as DocumentId;
+    const first = store.put(id, new Uint8Array([1, 2]));
+    const second = store.put(id, new Uint8Array([3, 4, 5]));
+    expect(helpers.revoked).toEqual([first.blobUrl]);
+    expect(store.get(id)?.bytes).toHaveLength(3);
+    expect(second.blobUrl).not.toBe(first.blobUrl);
+  });
+
+  it("delete revokes the blob URL exactly once and is idempotent", () => {
+    const helpers = stubUrlHelpers();
+    const store = createPdfByteStore(helpers);
+    const id = "doc_a" as DocumentId;
+    const record = store.put(id, new Uint8Array([1, 2, 3]));
+    expect(store.delete(id)).toBe(true);
+    expect(helpers.revoked).toEqual([record.blobUrl]);
+    expect(store.delete(id)).toBe(false);
+    expect(helpers.revoked).toHaveLength(1);
+    expect(store.get(id)).toBeNull();
+    expect(store.has(id)).toBe(false);
+  });
+
+  it("clear revokes every URL and empties the store", () => {
+    const helpers = stubUrlHelpers();
+    const store = createPdfByteStore(helpers);
+    const a = store.put("doc_a" as DocumentId, new Uint8Array([1]));
+    const b = store.put("doc_b" as DocumentId, new Uint8Array([2]));
+    store.clear();
+    expect(helpers.revoked.sort()).toEqual([a.blobUrl, b.blobUrl].sort());
+    expect(store.list()).toEqual([]);
+    expect(store.size()).toBe(0);
+  });
+
+  it("uses URL.createObjectURL by default when no override is supplied", () => {
+    const createObjectURL = vi.fn(() => "blob:built-in");
+    const revokeObjectURL = vi.fn();
+    const originalURL = globalThis.URL;
+    Object.defineProperty(globalThis, "URL", {
+      configurable: true,
+      writable: true,
+      value: Object.assign(Object.create(originalURL.prototype as object), {
+        createObjectURL,
+        revokeObjectURL,
+      }),
+    });
+    try {
+      const store = createPdfByteStore();
+      const rec = store.put("doc_z" as DocumentId, new Uint8Array([9]));
+      expect(rec.blobUrl).toBe("blob:built-in");
+      expect(createObjectURL).toHaveBeenCalledTimes(1);
+      store.delete("doc_z" as DocumentId);
+      expect(revokeObjectURL).toHaveBeenCalledWith("blob:built-in");
+    } finally {
+      Object.defineProperty(globalThis, "URL", {
+        configurable: true,
+        writable: true,
+        value: originalURL,
+      });
+    }
+  });
+});

src/browser/byte-store.ts

@@ -0,0 +1,115 @@
+/**
+ * `PdfByteStore` — in-memory store for uploaded PDF bytes, keyed by
+ * `DocumentId`.
+ *
+ * BROWSER SURFACE — not part of the headless ingest core. This mints and
+ * revokes `blob:` URLs and is meant to back an interactive upload/viewer
+ * session. Headless callers should use `src/pdf/` directly and never import
+ * this module.
+ *
+ * Uploaded PDFs are stored in memory only. Bytes survive within a tab
+ * session; reloading the page loses them unless the session was exported.
+ *
+ * One store instance per active session. The session-management layer is
+ * responsible for swapping the active store when the user switches sessions.
+ * The store also owns a small registry of issued `blob:` URLs so it can
+ * revoke them on delete/clear — no cross-cutting cleanup at the app layer.
+ */
+
+import type { DocumentId } from "@citation-evidence/engine/shared";
+
+export interface PdfByteRecord {
+  readonly bytes: Uint8Array;
+  /** A `blob:` URL the viewer can consume directly. */
+  readonly blobUrl: string;
+}
+
+export interface PdfByteStore {
+  put(documentId: DocumentId, bytes: Uint8Array): PdfByteRecord;
+  get(documentId: DocumentId): PdfByteRecord | null;
+  has(documentId: DocumentId): boolean;
+  delete(documentId: DocumentId): boolean;
+  list(): readonly DocumentId[];
+  /** Revoke every blob URL and clear the store. */
+  clear(): void;
+  /** Total bytes currently held — useful for UI dashboards. */
+  size(): number;
+}
+
+export interface CreatePdfByteStoreOptions {
+  /**
+   * Mint a URL for the given bytes. Defaults to `URL.createObjectURL` in
+   * environments that have it; tests can inject a deterministic stub.
+   */
+  readonly createObjectURL?: (blob: Blob) => string;
+  /** Revoke a URL previously minted by `createObjectURL`. */
+  readonly revokeObjectURL?: (url: string) => void;
+}
+
+export function createPdfByteStore(
+  options: CreatePdfByteStoreOptions = {},
+): PdfByteStore {
+  const createUrl =
+    options.createObjectURL ??
+    ((blob: Blob) => {
+      if (typeof URL === "undefined" || typeof URL.createObjectURL !== "function") {
+        throw new Error(
+          "createPdfByteStore: URL.createObjectURL not available — inject a stub via options",
+        );
+      }
+      return URL.createObjectURL(blob);
+    });
+  const revokeUrl =
+    options.revokeObjectURL ??
+    ((url: string) => {
+      if (typeof URL !== "undefined" && typeof URL.revokeObjectURL === "function") {
+        URL.revokeObjectURL(url);
+      }
+    });
+
+  const records = new Map<DocumentId, PdfByteRecord>();
+
+  return {
+    put(documentId, bytes) {
+      // Replace previous record (revoking the prior URL) if any.
+      const prior = records.get(documentId);
+      if (prior) revokeUrl(prior.blobUrl);
+      // Cast: Blob() does accept Uint8Array at runtime, but TS narrows the
+      // buffer type to ArrayBufferLike (could be SharedArrayBuffer) and
+      // refuses without help. The bytes here always come from a fresh
+      // arrayBuffer() call, so a regular ArrayBuffer is guaranteed.
+      const blob = new Blob([bytes as unknown as ArrayBuffer], {
+        type: "application/pdf",
+      });
+      const blobUrl = createUrl(blob);
+      const record: PdfByteRecord = { bytes, blobUrl };
+      records.set(documentId, record);
+      return record;
+    },
+    get(documentId) {
+      return records.get(documentId) ?? null;
+    },
+    has(documentId) {
+      return records.has(documentId);
+    },
+    delete(documentId) {
+      const record = records.get(documentId);
+      if (!record) return false;
+      revokeUrl(record.blobUrl);
+      records.delete(documentId);
+      return true;
+    },
+    list() {
+      return [...records.keys()];
+    },
+    clear() {
+      for (const record of records.values()) revokeUrl(record.blobUrl);
+      records.clear();
+    },
+    size() {
+      let total = 0;
+      for (const r of records.values()) total += r.bytes.byteLength;
+      return total;
+    },
+  };
+}

src/browser/index.ts

@@ -0,0 +1,20 @@
+/**
+ * `evidence-source/browser` — browser-facing upload helpers.
+ *
+ * BROWSER SURFACE. These helpers assume a browser session: they mint and
+ * revoke `blob:` URLs and hold uploaded bytes in memory for a viewer to
+ * mount. They are deliberately separated from the headless core (the package
+ * root, `src/index.ts`) so that headless consumers never pull in
+ * `URL.createObjectURL`/`Blob` upload machinery.
+ */
+
+export {
+  createPdfByteStore,
+  type CreatePdfByteStoreOptions,
+  type PdfByteRecord,
+  type PdfByteStore,
+} from "./byte-store";
+export {
+  ingestPdfFromFile,
+  type IngestPdfFromFileOptions,
+} from "./upload";

src/browser/upload.test.ts

@@ -0,0 +1,82 @@
+/**
+ * `ingestPdfFromFile` end-to-end: pipes a fixture PDF through the upload path,
+ * asserts the byte store keeps the bytes and the document record carries the
+ * minted `blob:` URL. Corpus resolution/skip lives in `tests/fixtures.ts`.
+ */
+
+import { describe, expect, it, vi } from "vitest";
+
+import { createPdfByteStore } from "./byte-store";
+import { ingestPdfFromFile } from "./upload";
+import { fixtureBytes, fixtureCorpusAvailable } from "../../tests/fixtures";
+
+const hasCorpus = fixtureCorpusAvailable();
+const FIXTURE_FILE = "Fristsetzung zur Bezifferung GÜ an Gegenseite 3 Wochen.pdf";
+
+class FakeFile {
+  readonly name: string;
+  private readonly bytes: Uint8Array;
+  constructor(bytes: Uint8Array, name: string) {
+    this.bytes = bytes;
+    this.name = name;
+  }
+  async arrayBuffer(): Promise<ArrayBuffer> {
+    const out = new ArrayBuffer(this.bytes.byteLength);
+    new Uint8Array(out).set(this.bytes);
+    return out;
+  }
+}
+
+describe.skipIf(!hasCorpus)("ingestPdfFromFile", () => {
+  it("round-trips a fixture PDF through ingest + byte store + blob URL", async () => {
+    const bytes = fixtureBytes(FIXTURE_FILE);
+    const file = new FakeFile(bytes, "demo.pdf") as unknown as File;
+    let counter = 0;
+    const store = createPdfByteStore({
+      createObjectURL: () => `blob:upload-stub-${++counter}`,
+      revokeObjectURL: () => {},
+    });
+
+    const { document, representation } = await ingestPdfFromFile(file, store);
+
+    const stored = store.get(document.id);
+    expect(stored).not.toBeNull();
+    expect(stored!.bytes.byteLength).toBe(bytes.byteLength);
+
+    expect(document.uri).toBe(`blob:upload-stub-${counter}`);
+    expect(document.title).toBe("demo.pdf");
+    expect(document.fingerprint).toMatch(/^[0-9a-f]{64}$/);
+
+    expect(representation.representationType).toBe("pdf-text");
+    expect((representation.canonicalText ?? "").length).toBeGreaterThan(0);
+  }, 30_000);
+
+  it("falls through to ingestPdf with no filename when given a plain Blob", async () => {
+    const bytes = fixtureBytes(FIXTURE_FILE);
+    const blob = {
+      async arrayBuffer() {
+        const out = new ArrayBuffer(bytes.byteLength);
+        new Uint8Array(out).set(bytes);
+        return out;
+      },
+    } as Blob;
+    const store = createPdfByteStore({
+      createObjectURL: () => "blob:no-name",
+      revokeObjectURL: () => {},
+    });
+    const { document } = await ingestPdfFromFile(blob, store);
+    expect(document.title).toBeUndefined();
+    expect(document.uri).toBe("blob:no-name");
+  }, 30_000);
+
+  it("explicit title option overrides the filename", async () => {
+    const bytes = fixtureBytes(FIXTURE_FILE);
+    const file = new FakeFile(bytes, "anonymous-name.pdf") as unknown as File;
+    const store = createPdfByteStore({
+      createObjectURL: vi.fn(() => "blob:override"),
+      revokeObjectURL: vi.fn(),
+    });
+    const { document } = await ingestPdfFromFile(file, store, { title: "Custom" });
+    expect(document.title).toBe("Custom");
+  }, 30_000);
+});

src/browser/upload.ts

@@ -0,0 +1,43 @@
+/**
+ * Upload-side ingest path.
+ *
+ * BROWSER SURFACE — not part of the headless ingest core. This is a thin
+ * wrapper over the headless `ingestPdf` that additionally pushes the bytes
+ * into a per-session `PdfByteStore`, which mints a `blob:` URL and stamps it
+ * onto `document.uri` so a viewer adapter can mount the PDF directly.
+ *
+ *   1. Read `file.arrayBuffer()` once into a `Uint8Array`.
+ *   2. Run the existing `ingestPdf(bytes, { filename })` pipeline to produce
+ *      `{document, representation}`.
+ *   3. Push the bytes into the per-session `PdfByteStore`.
+ *   4. Hand the engine inputs back to the caller, which wires them via
+ *      `engine.documents.register(...)`.
+ *
+ * Keeping URL-minting inside the byte store (rather than at the call site)
+ * means there is exactly one place that creates `blob:` URLs and exactly one
+ * place that revokes them.
+ */
+
+import { ingestPdf, type IngestPdfResult } from "../pdf/ingest";
+import type { PdfByteStore } from "./byte-store";
+
+export interface IngestPdfFromFileOptions {
+  /** Override the filename used as the document title. */
+  readonly title?: string;
+}
+
+export async function ingestPdfFromFile(
+  file: File | Blob,
+  store: PdfByteStore,
+  options: IngestPdfFromFileOptions = {},
+): Promise<IngestPdfResult> {
+  const bytes = new Uint8Array(await file.arrayBuffer());
+  const filename = "name" in file && typeof file.name === "string" ? file.name : undefined;
+  const ingested = await ingestPdf(bytes, {
+    ...(filename !== undefined ? { filename } : {}),
+    ...(options.title !== undefined ? { title: options.title } : {}),
+  });
+  const record = store.put(ingested.document.id, bytes);
+  const document = { ...ingested.document, uri: record.blobUrl };
+  return { document, representation: ingested.representation };
+}

src/index.ts

@@ -0,0 +1,20 @@
+/**
+ * `evidence-source` — headless document ingest core (PDF slice).
+ *
+ * This entry point exposes only runtime-agnostic ingest: pass PDF bytes,
+ * receive a `{ document, representation }` pair built entirely from
+ * engine-owned contracts (`@citation-evidence/engine/shared`). No browser,
+ * viewer, or persistence concerns leak through here.
+ *
+ * Browser-facing upload helpers live behind the separate
+ * `@citation-evidence/evidence-source/browser` entry point.
+ */
+
+export {
+  ingestPdf,
+  type IngestPdfInput,
+  type IngestPdfOptions,
+  type IngestPdfResult,
+} from "./pdf/ingest";
+export { extractPdf, type PdfExtractionResult } from "./pdf/extract";
+export { fingerprintBytes } from "./pdf/fingerprint";

src/pdf/extract.ts

@@ -0,0 +1,122 @@
+/**
+ * PDF text extraction → canonical text + PageMap + OffsetMap.
+ *
+ * Implements `wiki/ArchitectureOverview.md` §3.4 ("extract canonical text /
+ * build format-specific maps") for the `pdf-text` representation
+ * (`wiki/SharedContracts.md` §1, §3) and §6 (canonical normalization).
+ *
+ * Runtime independence: the PDF.js worker must be configured by the host
+ * application (`GlobalWorkerOptions.workerSrc`) before this module is
+ * called. In Vite/browser code the worker is bundled via the viewer; in
+ * Node tests the test setup file points it at
+ * `pdfjs-dist/legacy/build/pdf.worker.mjs`. No worker setup happens here
+ * so the same module loads cleanly in both runtimes.
+ *
+ * Page boundary semantics: canonical text concatenates per-page normalized
+ * text with a single "\n\n" paragraph separator. The separator is treated
+ * as belonging to the *preceding* page in `OffsetMap`, so the map covers
+ * `[0, canonicalText.length)` with no gaps. The last page has no trailing
+ * separator. This means `pageLength = globalEnd - globalStart` for
+ * every page; for non-last pages it equals (normalized page text length +
+ * 2). See `PageOffsetRange` in the engine's shared document contract.
+ */
+
+import { getDocument } from "pdfjs-dist";
+import type { PDFPageProxy } from "pdfjs-dist";
+import {
+  normalize,
+  type OffsetMap,
+  type PageInfo,
+  type PageMap,
+  type PageOffsetRange,
+} from "@citation-evidence/engine/shared";
+
+const PAGE_SEPARATOR = "\n\n";
+
+export interface PdfExtractionResult {
+  readonly canonicalText: string;
+  readonly pageMap: PageMap;
+  readonly offsetMap: OffsetMap;
+  readonly pageCount: number;
+}
+
+export async function extractPdf(bytes: Uint8Array): Promise<PdfExtractionResult> {
+  // PDF.js mutates the bytes buffer (transfers ownership). Pass a fresh copy
+  // so the caller's Uint8Array stays usable for fingerprinting after extract.
+  const data = new Uint8Array(bytes);
+  const loadingTask = getDocument({ data });
+  const doc = await loadingTask.promise;
+
+  try {
+    const pageCount = doc.numPages;
+    const pageInfos: PageInfo[] = [];
+    const pageNormalizedTexts: string[] = [];
+
+    for (let pageNumber = 1; pageNumber <= pageCount; pageNumber++) {
+      const page = await doc.getPage(pageNumber);
+      try {
+        const viewport = page.getViewport({ scale: 1 });
+        pageInfos.push({
+          page: pageNumber,
+          width: viewport.width,
+          height: viewport.height,
+        });
+
+        const rawText = await extractPageText(page);
+        pageNormalizedTexts.push(normalize(rawText).text);
+      } finally {
+        page.cleanup();
+      }
+    }
+
+    const { canonicalText, offsetMap } = buildOffsetMap(pageNormalizedTexts);
+
+    return {
+      canonicalText,
+      pageMap: pageInfos,
+      offsetMap,
+      pageCount,
+    };
+  } finally {
+    await doc.destroy();
+  }
+}
+
+async function extractPageText(page: PDFPageProxy): Promise<string> {
+  const content = await page.getTextContent();
+  // textContent.items are TextItem | TextMarkedContent. We want only the
+  // TextItem strings (those have a `str` field); marked-content entries are
+  // structural anchors and have no visible text.
+  const parts: string[] = [];
+  for (const item of content.items) {
+    if ("str" in item) {
+      parts.push(item.str);
+      if (item.hasEOL) parts.push("\n");
+    }
+  }
+  return parts.join("");
+}
+
+function buildOffsetMap(pageTexts: readonly string[]): {
+  canonicalText: string;
+  offsetMap: OffsetMap;
+} {
+  const ranges: PageOffsetRange[] = [];
+  let offset = 0;
+  for (let i = 0; i < pageTexts.length; i++) {
+    const text = pageTexts[i]!;
+    const isLast = i === pageTexts.length - 1;
+    const segmentLength = text.length + (isLast ? 0 : PAGE_SEPARATOR.length);
+    const globalStart = offset;
+    const globalEnd = offset + segmentLength;
+    ranges.push({
+      page: i + 1,
+      globalStart,
+      globalEnd,
+      pageLength: segmentLength,
+    });
+    offset = globalEnd;
+  }
+  const canonicalText = pageTexts.join(PAGE_SEPARATOR);
+  return { canonicalText, offsetMap: ranges };
+}

src/pdf/fingerprint.test.ts

@@ -0,0 +1,37 @@
+import { describe, expect, it } from "vitest";
+
+import { fingerprintBytes } from "./fingerprint";
+
+describe("fingerprintBytes", () => {
+  it("produces a 64-char lowercase hex SHA-256", async () => {
+    const digest = await fingerprintBytes(new Uint8Array([0x25, 0x50, 0x44, 0x46]));
+    expect(digest).toMatch(/^[0-9a-f]{64}$/);
+  });
+
+  it("matches the known SHA-256 of the empty input", async () => {
+    // Well-known constant: SHA-256 of zero bytes.
+    const digest = await fingerprintBytes(new Uint8Array(0));
+    expect(digest).toBe(
+      "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    );
+  });
+
+  it("is deterministic across repeated calls on the same bytes", async () => {
+    const bytes = new Uint8Array([1, 2, 3, 4, 5, 6, 7, 8]);
+    const a = await fingerprintBytes(bytes);
+    const b = await fingerprintBytes(bytes);
+    expect(a).toBe(b);
+  });
+
+  it("does not mutate or consume the input buffer", async () => {
+    const bytes = new Uint8Array([9, 8, 7]);
+    await fingerprintBytes(bytes);
+    expect([...bytes]).toEqual([9, 8, 7]);
+  });
+
+  it("distinguishes single-bit differences", async () => {
+    const a = await fingerprintBytes(new Uint8Array([0]));
+    const b = await fingerprintBytes(new Uint8Array([1]));
+    expect(a).not.toBe(b);
+  });
+});

src/pdf/fingerprint.ts

@@ -0,0 +1,31 @@
+/**
+ * SHA-256 fingerprint of raw document bytes.
+ *
+ * Implements the fingerprint half of `wiki/ArchitectureOverview.md` §3.4
+ * (the "compute fingerprint" pipeline step) and populates
+ * `Document.fingerprint` (`wiki/SharedContracts.md` §1).
+ *
+ * Uses Web Crypto's `crypto.subtle.digest`, which is available in browsers
+ * and in Node ≥ 20 (where it is exposed on `globalThis.crypto`). No
+ * platform branching — the API is the same in both environments.
+ */
+
+export async function fingerprintBytes(bytes: Uint8Array): Promise<string> {
+  // Copy into a fresh ArrayBuffer (not SharedArrayBuffer) so the digest call
+  // satisfies TS's updated `BufferSource` type, which excludes
+  // `SharedArrayBuffer`. The copy is O(n) — fine even for large PDFs since
+  // SHA-256 itself is already O(n).
+  const ab = new ArrayBuffer(bytes.byteLength);
+  new Uint8Array(ab).set(bytes);
+  const digest = await crypto.subtle.digest("SHA-256", ab);
+  return bytesToHex(new Uint8Array(digest));
+}
+
+function bytesToHex(bytes: Uint8Array): string {
+  let hex = "";
+  for (let i = 0; i < bytes.length; i++) {
+    const b = bytes[i]!;
+    hex += (b < 0x10 ? "0" : "") + b.toString(16);
+  }
+  return hex;
+}

src/pdf/ingest.test.ts

@@ -0,0 +1,119 @@
+/**
+ * Fixture-driven contract tests for the PDF ingest pipeline.
+ *
+ * For each fixture in the shared corpus manifest (ADR-0002):
+ *   1. Read the PDF bytes from disk.
+ *   2. Run `ingestPdf` end-to-end.
+ *   3. Assert the resulting Document + DocumentRepresentation honour the
+ *      manifest contract: media type is application/pdf, fingerprint is a
+ *      64-hex SHA-256, pageMap matches `page_count`, canonicalText contains
+ *      `known_good_quote`, and the offsetMap covers `[0, canonicalText.length)`
+ *      with no gaps.
+ *
+ * Corpus resolution and skip-on-absent behaviour live in `tests/fixtures.ts`.
+ */
+
+import { describe, expect, it } from "vitest";
+
+import { ingestPdf } from "./ingest";
+import { fingerprintBytes } from "./fingerprint";
+import {
+  fixtureBytes,
+  fixtureCorpusAvailable,
+  loadFixtures,
+} from "../../tests/fixtures";
+
+const hasCorpus = fixtureCorpusAvailable();
+const FIXTURES = hasCorpus ? loadFixtures() : [];
+
+describe.skipIf(!hasCorpus)("ingestPdf — fixture corpus", { timeout: 30_000 }, () => {
+  for (const fixture of FIXTURES) {
+    describe(fixture.id, () => {
+      const bytes = fixtureBytes(fixture.filename);
+
+      it("produces a Document with PDF media type and SHA-256 fingerprint", async () => {
+        const { document } = await ingestPdf(bytes, { filename: fixture.filename });
+        expect(document.mediaType).toBe("application/pdf");
+        expect(document.fingerprint).toMatch(/^[0-9a-f]{64}$/);
+        expect(document.title).toBe(fixture.filename);
+        const expected = await fingerprintBytes(bytes);
+        expect(document.fingerprint).toBe(expected);
+      });
+
+      it("produces a pdf-text representation with the expected page count", async () => {
+        const { representation } = await ingestPdf(bytes);
+        expect(representation.representationType).toBe("pdf-text");
+        expect(representation.pageMap?.length).toBe(fixture.page_count);
+        expect(representation.offsetMap?.length).toBe(fixture.page_count);
+      });
+
+      it("canonical text contains the manifest's known-good quote", async () => {
+        const { representation } = await ingestPdf(bytes);
+        const text = representation.canonicalText ?? "";
+        expect(text).toContain(fixture.known_good_quote);
+      });
+
+      it("offsetMap is gap-free and covers [0, canonicalText.length)", async () => {
+        const { representation } = await ingestPdf(bytes);
+        const text = representation.canonicalText ?? "";
+        const offsets = representation.offsetMap ?? [];
+        expect(offsets.length).toBeGreaterThan(0);
+        expect(offsets[0]!.globalStart).toBe(0);
+        expect(offsets.at(-1)!.globalEnd).toBe(text.length);
+        for (let i = 0; i < offsets.length; i++) {
+          const r = offsets[i]!;
+          expect(r.page).toBe(i + 1);
+          expect(r.globalEnd - r.globalStart).toBe(r.pageLength);
+          if (i > 0) expect(r.globalStart).toBe(offsets[i - 1]!.globalEnd);
+        }
+      });
+
+      it("pageMap entries have positive width and height in user-space points", async () => {
+        const { representation } = await ingestPdf(bytes);
+        const pages = representation.pageMap ?? [];
+        for (let i = 0; i < pages.length; i++) {
+          const p = pages[i]!;
+          expect(p.page).toBe(i + 1);
+          expect(p.width).toBeGreaterThan(0);
+          expect(p.height).toBeGreaterThan(0);
+        }
+      });
+    });
+  }
+});
+
+describe.skipIf(!hasCorpus)("ingestPdf — option handling", { timeout: 30_000 }, () => {
+  const fixture = FIXTURES[0];
+  const bytes = fixture ? fixtureBytes(fixture.filename) : new Uint8Array();
+
+  it("uses explicit title over filename", async () => {
+    const { document } = await ingestPdf(bytes, {
+      filename: fixture!.filename,
+      title: "Custom Title",
+    });
+    expect(document.title).toBe("Custom Title");
+  });
+
+  it("omits title entirely when neither filename nor title is supplied", async () => {
+    const { document } = await ingestPdf(bytes);
+    expect(document.title).toBeUndefined();
+  });
+
+  it("propagates uri and metadata when supplied", async () => {
+    const { document } = await ingestPdf(bytes, {
+      uri: "file:///example.pdf",
+      metadata: { source: "test" },
+    });
+    expect(document.uri).toBe("file:///example.pdf");
+    expect(document.metadata).toEqual({ source: "test" });
+  });
+
+  it("accepts ArrayBuffer input", async () => {
+    const ab = bytes.buffer.slice(
+      bytes.byteOffset,
+      bytes.byteOffset + bytes.byteLength,
+    ) as ArrayBuffer;
+    const { document } = await ingestPdf(ab);
+    expect(document.fingerprint).toMatch(/^[0-9a-f]{64}$/);
+  });
+});

src/pdf/ingest.ts

@@ -0,0 +1,88 @@
+/**
+ * PDF ingest pipeline → `{ document, representation }`.
+ *
+ * Implements `wiki/ArchitectureOverview.md` §3.4 ("Raw Source → identify
+ * media type → compute fingerprint → extract metadata → extract canonical
+ * text → build format-specific maps → persist Document +
+ * DocumentRepresentation") for the PDF source format.
+ *
+ * Ingest is a pure function over bytes: it does not persist anything. The
+ * caller (engine repositories, or the app layer) writes the returned
+ * Document + DocumentRepresentation into the chosen store.
+ */
+
+import {
+  newId,
+  type Document,
+  type DocumentRepresentation,
+} from "@citation-evidence/engine/shared";
+import { extractPdf } from "./extract";
+import { fingerprintBytes } from "./fingerprint";
+
+const PDF_MEDIA_TYPE = "application/pdf";
+
+export interface IngestPdfOptions {
+  /** Original filename, used as the default title when no title is given. */
+  readonly filename?: string;
+  /** Optional pre-existing title (overrides filename). */
+  readonly title?: string;
+  /** Optional source URI (e.g. file:// or https://). */
+  readonly uri?: string;
+  /** Free-form metadata persisted on the Document record. */
+  readonly metadata?: Readonly<Record<string, unknown>>;
+}
+
+export interface IngestPdfResult {
+  readonly document: Document;
+  readonly representation: DocumentRepresentation;
+}
+
+export type IngestPdfInput = Uint8Array | ArrayBuffer | Blob;
+
+export async function ingestPdf(
+  input: IngestPdfInput,
+  options: IngestPdfOptions = {},
+): Promise<IngestPdfResult> {
+  const bytes = await toBytes(input);
+  const [fingerprint, extraction] = await Promise.all([
+    fingerprintBytes(bytes),
+    extractPdf(bytes),
+  ]);
+
+  const now = new Date().toISOString();
+  const documentId = newId("document");
+  const representationId = newId("representation");
+  const title = options.title ?? options.filename;
+
+  const document: Document = {
+    id: documentId,
+    mediaType: PDF_MEDIA_TYPE,
+    fingerprint,
+    createdAt: now,
+    updatedAt: now,
+    ...(title !== undefined ? { title } : {}),
+    ...(options.uri !== undefined ? { uri: options.uri } : {}),
+    ...(options.metadata !== undefined ? { metadata: options.metadata } : {}),
+  };
+
+  const representation: DocumentRepresentation = {
+    id: representationId,
+    documentId,
+    representationType: "pdf-text",
+    contentHash: fingerprint,
+    canonicalText: extraction.canonicalText,
+    pageMap: extraction.pageMap,
+    offsetMap: extraction.offsetMap,
+    generatedAt: now,
+  };
+
+  return { document, representation };
+}
+
+async function toBytes(input: IngestPdfInput): Promise<Uint8Array> {
+  if (input instanceof Uint8Array) return input;
+  if (input instanceof ArrayBuffer) return new Uint8Array(input);
+  // Blob (covers `File` in browsers — File extends Blob).
+  const buf = await input.arrayBuffer();
+  return new Uint8Array(buf);
+}