Why

Refs #1366. Document observations need to point back to the source PDF page so readers can inspect the evidence. This is the page-level slice agreed for v0.24.0, following the amended SPEC-89: the observation is the claim, and citations use OKF footnotes joined to stable sources[].id values.

What Changed

• Added an optional typed PDF page locator to document-agent observations: one-based physical page, with an optional separate printed page_label.
• Trusted enrichment generates page-specific OKF-compatible sources entries and Markdown footnotes. Agents provide page numbers, not destination URLs.
• Preserved ordinary observation category, tags, context, and existing uncited document serialization.
• Added real PDF extraction/enrichment/Markdown tests and an API → canonical read-back → on-disk file → database observation integration test, without mocks or parameterized cases.
• Documented the portable representation and its limits in docs/DOCUMENT_CITATIONS.md.

Implementation Details

The page number is validated against the actual extraction page count. Citation resources use project/bundle-root-relative, percent-encoded file paths with standard #page=N fragments. IDs such as document-page-2 are scoped to the note's single trusted source PDF, not an array index; repeated citations share an entry and reordering does not retarget them. Printed labels never change the navigation page.

The existing singular source retains the original PDF checksum and storage-version provenance. The new plural sources supplies citation targets. Trusted-envelope validation checks those entries against the source PDF; generated footnote definitions are not agent-supplied. A separating space prevents trailing Markdown escapes in observation text from escaping the generated marker.

Specifications and prior art

• RFC 8118 §3: PDF fragment identifiers: adopted the one-based physical page=N addressing convention.
• OKF §5.1: provenance sources: adopted stable source IDs, resource references, and Markdown footnote-label joins. This is not a claim of full OKF conformance ([FEATURE] opt-in Open Knowledge Format (OKF) conformance for the note format #1246 remains separate).
• Zotero PDF reader: product precedent for navigating from note/annotation evidence back to its PDF page, not a dependency or proprietary format adopted here.
• W3C Web Annotation TextQuoteSelector: prior art for later exact-quote/context precision. W3C selectors are not implemented in this PR.
• SPEC-89 and the controlling “The observation is the claim” amendment: no separate claim store or competing citation syntax.

Testing

• uv run pytest test-int/test_document_page_citations.py --no-cov -q: 19 passed on SQLite.
• BASIC_MEMORY_TEST_POSTGRES=1 uv run pytest test-int/test_document_page_citations.py --no-cov -q: 19 passed against real PostgreSQL via testcontainers.
• uv run pytest tests/schemas/test_document.py tests/schemas/test_document_agent_temporal.py tests/document_ingestion --no-cov -q: 118 passed.
• just fast-check: passed (lint, formatting, typecheck).
• just doctor: passed the isolated file/API/index/search/status loop.
• codex review --uncommitted: run before pushing; its trailing-backslash finding was fixed with a real Markdown-rendering regression test. The second pass reported no actionable defects (76 focused tests passed in the review).
• An additional coverage-instrumented run failed during FastMCP/Pydantic settings initialization before collection; no coverage percentage is claimed. The normal targeted runs above pass.

Risks / Follow-ups

Review follow-up: generated citation references as well as definitions are reserved across agent-controlled Markdown fields when locators are present. Omitted page labels are compatible with an explicit label in either observation order; only distinct explicit labels conflict. Explicit tests also cover duplicate source IDs and non-PDF sources. The follow-up pre-push review found no actionable regressions (19 citation tests passed).

• This is the Core portable contract and assembly layer. Cloud prompt adoption and UI/viewer resource routing need separate integration; a project-root-relative link is not a new Cloud HTTP route.
• The link opens the current resource in supporting viewers; the retained checksum/version records extraction provenance but does not implement historical-version retrieval or automatic stale-citation detection.
• Extraction offset/page maps, quote selectors/highlighting, evidence watermarks, contradiction findings, and generic OKF conformance remain out of scope. Optional source locator on document observations (page → context, resolvable against the extraction page map) #1366 stays open to track its remaining precision/integration work.
• No migrations, changes to canonical file-service behavior, or additional concurrency machinery.