Improve Markdown document graph resolution

[forhappy]

Summary

• give every Markdown structural block a deterministic section-qualified identity
• match duplicate automatic heading slugs in source order and decode bounded percent-encoded fragments
• resolve exact, extensionless, repository-root, directory-index, and unique-wikilink document targets against the complete project inventory
• connect Markdown documentation to file inventory for Python, Rust, Go, Java, TypeScript, and every other extracted language
• keep missing fragments and ambiguous targets unresolved instead of selecting a candidate or falling back to a document root
• retain exact document relationships at the default low inference level and advance extraction semantics to v12

Why

The per-file Markdown extractor previously guessed a .md endpoint before the full project inventory was available. Cross-file fragments stopped at the document root, extensionless and root-relative paths could not be selected accurately, and the generic endpoint-rewrite provenance correctly classified the later repair as inferred. Default-low publication therefore retained only two same-file HTTPX references and filtered the project-resolved links.

The new flow preserves source spelling and the exact link site in compass-languages, then performs unique target selection in compass-resolve. Because the resulting relationship is an exact source-backed document fact rather than heuristic endpoint repair, it survives low inference. Ambiguity remains explicit and unpublished.

Impact

• agents can navigate directly from documentation prose to the referenced heading or source file instead of retrieving and searching a whole document
• repeated headings and structural blocks have stable, scoped identities that improve search precision
• .md, .markdown, .mdx, .qmd, and .skill documents share the same bounded resolution rules
• language-specific code extractors and resolvers are unchanged; document-to-code coverage is verified across Python, Rust, Go, Java, and TypeScript file inventories
• graph schema v1 and relationship direction are unchanged

HTTPX default-low qualification

Metric	Baseline	Candidate
Files indexed	102	102
Nodes	4,867	4,867
Edges	7,123	7,152
Published Markdown references	2	31
Exact project-resolved document references	0	29

All 29 added edges have real endpoints and exact link-site evidence; no nodes were added. A fresh release-mode build completed in 1.25s. The cold release-mode CompassQL lookup was 0.13s on the baseline and 0.21s on the candidate; latency is reported for context and was not an optimization criterion.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --lib --bins --locked -- -D warnings
• cargo test --workspace --lib --bins --locked
• full compass-languages and compass-resolve test suites
• focused default-low end-to-end Compass Core document publication test
• ./scripts/qualify_code_graph_v1.sh --fixtures-only
• sh scripts/check_product_boundary.sh
• fresh release-mode HTTPX build plus CompassQL document-link query

An additional cargo clippy -p compass-languages -p compass-resolve -p compass-core --all-targets --all-features --locked -- -D warnings run reached two pre-existing test-only violations (expect_err in framework_qualification.rs and items_after_test_module in evidence/projection/mod.rs). The documented workspace lib/bin Clippy gate passes.