Resolve Python receiver calls and audit impact coverage

[forhappy]

Python service callee queries could report zero despite attribute-chain calls, while impact summaries counted only the capped presentation results. This change preserves source invocations, resolves supported nominal receivers, and independently audits direct impact.

• Resolve annotated and injected Python receivers, unit-of-work member chains and source-backed factory returns. Unknown receivers retain source inventories and optional inferred edges; ambiguous or mutated receivers remain unresolved.
• Aggregate class-owned methods and initializers while retaining actual call endpoints. Report resolved and unresolved source-call counts instead of a false zero. Producers without inventories retain their existing relationship summaries.
• Expand reverse impact defaults and add CLI --relation / MCP relations filters. Count and group direct/transitive dependents independently of the presentation cap, and fail a complete audit if selected direct dependents were omitted. Every known direct contact is accounted for as a dependent, outgoing context or exclusion; outgoing references are not claimed affected code.
• Preserve qualified external Rust factory receivers under existing inference rules, with negative coverage for absent source-local methods.

Compatibility: optional graph callSites and query callSummary / impactSummary fields; disposable AST cache version 14. Rebuild Python graphs to obtain inventories. Historical graphs remain readable and immutable. Command, contract and migration documentation is included.

Public-repository qualification used a read-only pinned Dify checkout. A source-first seeded sample captured 2,550/2,550 source calls across 10 production services: 698 sites had all-exact target evidence and 1,537 had inferred edge evidence. These are capture measurements, not a claim of exact runtime dispatch. All five widely used symbol audits accounted for their complete published direct adjacency. The graph retained its partial-source warning for 15 quarantined edges. Reproduction steps and reviewed source links are in docs/implementation/python-call-impact-qualification.md.

The fixture graph adds 17 explicitly inferred receiver nodes. Five density/exact-isolation baselines are adjusted for that addition; the absolute exact-edge, exact cross-file-edge and typed endpoint floors, relationship-specific gates and semantic assertions are retained. Exact edges increase from 956 to 962 and exact cross-file edges remain 56.

Validation:

• cargo fmt --all -- --check
• Workspace Clippy (--lib --bins --locked -- -D warnings) and workspace library/binary tests
• Full CLI integration suite, targeted query/output/model/MCP/core contracts, and 220 resolver integration tests
• Native qualification-helper Clippy; deterministic query-relevance gate
• ./scripts/qualify_code_graph_v1.sh --fixtures-only
• Product boundary gate and public Dify qualification

JavaScript, packaging and other platform matrices are left to CI; no JavaScript sources changed. Rust builds and qualification artifacts were kept in this checkout's external target directory.