P7-1 (`PdfTextExtractor`) + P7-2 (`PdfPageV1Chunker`) 의 라이브러리는
완성됐지만 `kebab-app::ingest` 가 `MediaType::Pdf` 를 dispatch 하지 않아
CLI 에서 PDF 가 보이지 않는 상태. P7-3 이 그 와이어링을 다룬다 — P6-4
의 image wiring 패턴과 평행.

핵심 결정 (spec 본문):
- 새 private fn `ingest_one_pdf_asset` (P6-4 의 `ingest_one_image_asset`
  와 평행). `ingest_one_asset` match 에 `MediaType::Pdf` arm 추가.
- per-medium chunker 선택: PDF 는 `PdfPageV1Chunker` 하드코딩 (md 는
  `MdHeadingV1Chunker` 그대로). `config.chunking.chunker_version` 은 PDF
  ingest 에서 무시 (deviation, HOTFIXES 추가 예정).
- encrypted PDF / corrupt PDF → `errors+=1` + `IngestItem.error` 에 P7-1
  의 `qpdf --decrypt` 안내 그대로 보존.
- 빈/scanned candidate 페이지 → asset 인덱싱, 빈 페이지 0 chunk, P7-1
  emit 한 `Provenance::Warning` 그대로 통과. 향후 OCR fallback 까지는
  검색 불가 (out of scope).
- determinism stress: extract → chunk 사이에 `now()` 추가 호출 금지
  (P6-4 와 동일 invariant).
- 11 통합 테스트 + smoke 업데이트 (별도 implementation PR).

`tasks/INDEX.md` P7 components 2 → 3 반영.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@@ -0,0 +137,4 @@
+| Per-page text extraction Err (panic absorbed by `catch_unwind`) | unchanged | yes | `Provenance::Warning` (page N "scanned candidate") — emitted by P7-1, propagated through |
+| Empty page (no `/Contents` stream) | unchanged | yes | `Provenance::Warning` (page N "empty (scanned candidate)") — emitted by P7-1 |
+| Embedding call fails for a chunk | bubbles up via existing markdown path's error handling | partial — depends on existing behaviour | n/a |
+

@@ -0,0 +153,4 @@
+
+## Test plan
+
+| kind | description | fixture / data |

@@ -0,0 +158,4 @@
+| integration | TempDir KB + 1 small text PDF (3 pages) → `kebab ingest` produces 1 doc + 3 chunks; each chunk's `source_spans[0]` is `Page { page: i, .. }`; chunks stored + embedded | `kebab-app/tests/pdf_pipeline.rs` (new), in-memory PDF via `lopdf` builder (mirrors `kebab-parse-pdf::tests::common`) |
+| integration | Re-ingest same PDF → identical `doc_id` and identical `chunk_id` set (P1 idempotency contract) | inline |
+| integration | Encrypted PDF → asset NOT stored; `errors+=1`; `IngestItem.error` mentions `qpdf` / `decrypt` | inline (lopdf builder + fake `/Encrypt` trailer) |
+| integration | Corrupt header PDF → asset NOT stored; `errors+=1`; error message mentions PDF parse failure | inline |

@@ -0,0 +166,4 @@
+| integration | `IngestReport` invariant `scanned == new + updated + skipped + errors` holds when ingesting a mixed corpus including a corrupt PDF | inline |
+| integration | Long PDF (50 pages × ~1.5 KB body each = ~75 KB) produces ≥50 chunks (≥1 per page); embedding loop completes; storage round-trips | inline |
+| smoke | Update `docs/SMOKE.md` so the runbook ingests at least one PDF fixture and verifies search-by-page-text works + `inspect doc` shows `SourceSpan::Page` | docs change |
+

@@ -0,0 +169,4 @@
+
+The opt-in real-Ollama integration tests stay in `kebab-llm-local` / `kebab-parse-image`. P7-3 adds zero LM dependency — PDF text extraction is local-only, so there is no equivalent hermetic-vs-real split to manage.
+
+## Definition of Done

@@ -0,0 +175,4 @@
+
+- [ ] `tasks/p7/p7-3-pdf-ingest-wiring.md` 작성 + self-review (placeholder / 모순 / 모호성 / scope)
+- [ ] `tasks/INDEX.md` "P7 — 3 components" 반영
+- [ ] PR 본문에 design §3.4, §3.5, §6.1, §9.2 링크

@@ -0,0 +189,4 @@
+- [ ] `kebab ask "<query about PDF body>"` produces an `Answer` whose `Citation` includes a `Page` span
+- [ ] HOTFIXES entry written if any spec deviation surfaces during implementation
+- [ ] `docs/SMOKE.md` includes a PDF-fixture step
+

@@ -0,0 +195,4 @@
+- **Scanned PDF OCR fallback** — empty/extract-failed pages stay searchable=false in v1. A future task ("P+ scanned-PDF-ocr") routes those pages through P6-2's `OllamaVisionOcr` after rasterising the page via a PDF renderer (e.g. `mupdf-rs`, `pdfium-render`). Excluded here because (a) it requires a new system / Rust dep we don't have yet, and (b) v1 user scenario is text-embed PDFs (papers, exported reports).
+- **Multi-column reading order / table extraction / formula detection / form-field extraction / bookmark or outline ingestion** — all deferred to future PDF-layout task. P7-1 already lists these as out of scope and the wiring inherits.
+- **Body multilingual via CID font support** — handled at the parser layer (P7-1). UTF-16BE Title metadata works today; non-Latin body text depends on the PDF's font CID mapping.
+- **Per-medium `chunker_version` config** — current `Config::chunking.chunker_version` is single-valued and serves markdown only. PDF ingest ignores it (hard-codes `pdf-page-v1`). A future P+ task either splits the field per medium or introduces a chunker registry. Logged as a deviation in `tasks/HOTFIXES.md` once implementation lands.

@@ -0,0 +139,4 @@
+| Empty page (no `/Contents` stream) | unchanged | yes | `Provenance::Warning` (page N "empty (scanned candidate)") — emitted by P7-1 |
+| `Embedder::embed(...)` Err (any chunk) | `errors+=1` | yes (doc + chunk rows already written before embed call — see below) | n/a |
+
+The embedding call sits *after* `put_document` / `put_blocks` / `put_chunks` in `kebab-app::ingest_one_asset` (markdown path, lines 615+), so a failed embed leaves doc + chunk rows on disk while no vector exists. This is consistent with the markdown path and accepted as v1 behaviour — re-running `kebab ingest` re-attempts the embed for any chunk whose `embedding_id` is missing from the vector store. Whole-asset rollback on embed-fail is a P+ task (atomic ingest transaction).

@@ -0,0 +166,4 @@
+| integration | Mixed page PDF (page 1 text, page 2 empty / scanned, page 3 text) → asset stored; 2 chunks (pages 1 + 3); `doc.provenance.events` contains exactly 1 `Warning` for page 2 marked "scanned candidate" | inline |
+| integration | `kebab inspect doc <pdf_doc_id>` returns the PDF `CanonicalDocument` with per-page `Block::Paragraph` and `SourceSpan::Page` intact | inline |
+| integration | Hybrid search across mixed corpus (1 markdown + 1 PDF) returns the PDF chunk for a query whose terms appear only in the PDF body | inline (real `multilingual-e5-small` embedding) |
+| integration | `IngestReport` invariant `scanned == new + updated + skipped + errors` holds when ingesting a mixed corpus including a corrupt PDF | inline |

@@ -0,0 +172,4 @@
+
+The opt-in real-Ollama integration tests stay in `kebab-llm-local` / `kebab-parse-image`. P7-3 adds zero LM dependency — PDF text extraction is local-only, so there is no equivalent hermetic-vs-real split to manage.
+
+## Definition of Done

@@ -0,0 +193,4 @@
+- [ ] `docs/SMOKE.md` includes a PDF-fixture step
+
+## Out of scope
+

@@ -0,0 +212,4 @@
+- **Encrypted-PDF error message is the only operator-visible signal.** The `kebab ingest` summary shows `errors=1` and the `IngestItem.error` message; a user who didn't read the full output may miss the `qpdf --decrypt` hint. Acceptable in v1 — a future `kebab inspect ingest <run_id>` (P9) renders structured per-asset errors. For now, ensure the test asserts the hint is preserved verbatim from P7-1's bail string.
+- **Determinism stress with `now()` calls.** Same constraint as P6-4 — extract → chunk pipeline must not insert wall-clock reads between steps. P7-1 emits its own `now()` once for all per-page Provenance events; the PDF wiring branch must add no further `now()`s inside the per-asset path.
+- **`pdf-page-v1` per-chunk hash variant `#c{char_start}` is opaque.** Downstream tools comparing `chunk_id`s by exact match work fine (it's still a deterministic blake3 input). Tools attempting to derive a stable position from the `chunk_id` alone would fail — they must read `chunk.source_spans[0].char_start`. Documented in P7-2's HOTFIXES entry; cross-referenced here for findability.
+- **HEIC / RAW image and PDF/A subspecies share `MediaType::Other`.** PDF/A is detected by `kebab-source-fs` as `MediaType::Pdf` (header is still `%PDF-`); PDF subtype variants do not branch separately. Acceptable for v1.