Implement ESRC-WP-0002/0003/0004: HTML/MD ingest, metadata, recovery

Add ingestHtml and ingestMarkdown with ADR-0003 pageless offset semantics,
PDF intrinsic metadata extraction with caller-wins merge (WP-0003), and
citation recovery primitives including re-ingest reconcile, local quote
search, and pluggable discovery hooks (ADR-0004). Mark all three workplans
finished with contract tests (80 passing).

docs/ADR-0003-html-markdown-representation.md

@@ -0,0 +1,58 @@
+# ADR-0003 — HTML and Markdown representation contract
+
+- **Status:** accepted
+- **Date:** 2026-07-09
+- **Workplan:** ESRC-WP-0002 (T01)
+
+## Context
+
+ESRC-WP-0001 established the PDF ingest boundary. ESRC-WP-0002 extends ingest to
+HTML and Markdown while reusing engine-owned contracts from
+`@citation-evidence/engine/shared`.
+
+`RepresentationType` already reserves `html-dom` and `markdown-rendered` in
+`citation-engine`; no engine schema change is required for this slice.
+
+## Decision
+
+### Representation types
+
+| Source format | `mediaType`        | `representationType`   |
+| ------------- | ------------------ | ---------------------- |
+| HTML          | `text/html`        | `html-dom`             |
+| Markdown      | `text/markdown`    | `markdown-rendered`    |
+
+### Pageless offset semantics
+
+Reflowable formats have no fixed pages:
+
+- `pageMap` is **omitted** (undefined).
+- `offsetMap` is a **single synthetic range** on page 1 covering
+  `[0, canonicalText.length)` with no gaps. This keeps
+  `TextPositionSelector` usable without inventing physical page geometry.
+- `structureMap` remains `never` in the engine contract and is not populated.
+
+### Canonical text
+
+Both pipelines decode source bytes as UTF-8, derive plain text from the format,
+then apply `normalize()` from `@citation-evidence/engine/shared`. HTML strips
+active content (`script`, `style`, `iframe`, `object`, `embed`, `noscript`)
+before text extraction.
+
+### Content hash
+
+`contentHash` is the SHA-256 fingerprint of the **source bytes** (same rule as
+PDF ingest), not a hash of canonical text.
+
+### Fixtures
+
+HTML/Markdown contract tests use small inline fixtures owned by this repo. They
+contain no PII and do not duplicate the upstream PDF corpus (ADR-0002).
+
+## Consequences
+
+- `ingestHtml` and `ingestMarkdown` mirror the PDF `{ document, representation }`
+  shape.
+- Anchoring layers may add DOM/structural selectors later; this slice delivers
+  canonical text and position offsets only.
+- Sanitization is extraction-time stripping, not a sandboxed renderer.