fix: classify repeated edge occurrences

[hsneves]

Depends on #2177. The new work in this stacked PR is commit fb77da2.

Summary

• preserve portable source occurrence evidence for Python type references and calls
• normalize edges after endpoint remapping, separating legitimate repeated occurrences from producer duplicates
• classify zero-node extraction outcomes and treat unrecognized JSON data as intentional skips
• expose occurrence evidence in query/path/explain, HTML, GraphML, Cypher, Neo4j, and FalkorDB without collapsing parallel edges
• extend diagnostics and generated skills with the new classifications

Validation

• uv run ruff check .
• uv run python -m tools.skillgen --check
• uv run python -m tools.skillgen --schema-singleton
• uv run python -m tools.skillgen --monolith-roundtrip
• uv run pytest -q (3607 passed, 33 skipped)
• real monorepo AST baseline: 457 duplicate occurrences reconciled as 418 legitimate + 39 producer duplicates; 14 intentional JSON skips; zero unclassified duplicates

[hsneves]

Follow-up validation on the real monorepo exposed and fixed a Python 3.10 packaging gap: tomli was development-only, so installed CLI environments could not parse pyproject.toml. Commit 5430a55 moves it to runtime dependencies. Final validation: wheel build succeeded; full suite is now 3608 passed / 33 skipped; the installed Python 3.10 CLI materialized both manifests and the final corpus has zero unexpected-empty files.