[jeremiepas]

Qualification on Apache Fineract (Java, ~4k files) showed the tree-sitter path produces a structurally correct but semantically hollow graph. This closes the four diagnosed gaps.

The gaps (reproduced on a 20-line annotated Java fixture before any fix)

1. Declaration names are wrong. A method returning ClientData was labeled ClientData, not its own name — extractIdentifier picked the first identifier-class child, and the return-type type_identifier precedes the name positionally. Every method sharing a return type collided on one base id, and the contains remap (Map.fromList, last-wins) silently dropped containment edges for all but one.
2. Zero calls edges outside TS/JS. callTypes only matched call_expression/new_expression. Java/C#/PHP/Python/Ruby call nodes never matched, and the callee heuristic ("first identifier in the subtree") returned the receiver of receiver.method(), not the method.
3. Annotations unstructured. @Path(...), @GET, @Scheduled(...) survived only as the head of a flattened signature that also swallowed the method body — no queryable route, verb, cron, or config key.
4. Notable strings invisible. System.getenv("X") never reached the graph since at Function granularity (the default) method bodies are opaque.

A fifth, found during diagnosis: bare type_identifier references leaked as nodes with empty ids.

The fix

• Field-aware AST: TSNodeInfo now carries each child's tree-sitter grammar field name — the Hackage binding already populates it, Core.hs just discarded it. childByField/childrenByField let every other fix below key off grammar semantics instead of positional guessing.
• Names: extractIdentifier prefers the grammar's name field, with a C/C++ declarator-field-chain fallback and a type-list fallback that prefers identifier over type_identifier. type_identifier removed from apiSurfaceTypes (AVI-1440: references emit no node).
• Collision fix: the per-file remap now keys on (base id, line), not base id alone, so overloaded methods each keep their own contains edge.
• calls edges fire for every supported grammar's own call vocabulary (java, c-sharp, php, python, ruby, go, rust, c/cpp, ts/js). Callee resolution is field-based with a rightmost-name rule, so svc.save() targets save, not svc. Call-target indexing now covers class members, not just top-level definitions.
• Annotations (Java fields, Python/TS decorators) land in a structured nodeExtra.annotations array, plus a references edge to the resolved import and a distillation table (route, httpVerb, cron, fixedDelay, topics, configKeys). Signatures are now declaration headers, not full text.
• Notable strings: env-reads, @Value placeholders, and @KafkaListener/@RabbitListener topics resolve to canonical env:/config:/topic: nodes with references edges, working at every granularity (including the opaque-body default).

How it was verified

• Every grammar claim was checked by extracting the exact pinned tarball (per graphos.cabal) and reading src/node-types.json directly — not recollection. Two corrections this surfaced: Ruby's real invocation node is method_call (not bare call); PHP's object_creation_expression has no fields at all.
• cabal test targeted runs: 153 examples, 0 failures, 1 pre-existing pending, across every touched describe block. cabal build all is clean across every component.
• Two real-grammar hspec checks (not hand-built fixtures) run the actual pinned Java and Python parsers end-to-end.
• Full before/after rerun against apache/fineract itself (7250 .java files, shallow clone) — a main-built binary vs this branch's binary, identical flags, identical checkout. See openspec/changes/extraction-semantic-enrichment/FINERACT_VERIFICATION.md for the complete table. Headline, at real-world scale: calls edges 0 → 20,709; methods mislabeled with a bare return-type name 5,803/40,059 (14.5%) → 0; type_identifier leak nodes 2,824 → 0; 33,920 nodes gained a structured annotations array (829 httpVerb, 782 route, 37 configKeys, 1 topics); 39 canonical Env/Config/Topic nodes appeared where there were none.

Bugs found and fixed along the way (not in the original plan)

• Java field_declaration.declarator → variable_declarator.name: the field-name fix for methods had the same positional bug for fields (private String tenantHost; labeled String).
• Ruby's method/singleton_method were missing from functionBoundaryTypes, so every Ruby call's enclosing scope resolved to the module instead of the method.
• The env-read string literal lives inside the call's arguments-field value (an argument_list node), not as a direct child of the call — the first version silently matched zero env-reads until this was corrected.

Known limitations (see openspec/changes/extraction-semantic-enrichment for full detail)

• Cross-file resolution for Java/Kotlin/C#/PHP/Perl is same-file only — the import-binding map calls/references share only understands TS/JS/Python's quoted import shape; these five grammars' unquoted form belongs to the separate Imports.hs pipeline, not wired in here.
• PHP has no annotation support: the pinned grammar has zero attribute-syntax nodes.
• Rust attribute_item/C# attribute_list are not wired into nodeExtra.annotations this pass.
• C# field names are unconfirmed (no node-types.json vendored for it) — falls back to positional scan.
• Kotlin vendoring was investigated and found genuinely blocked, not merely deferred. The only fwcd/tree-sitter-kotlin revision that generates at the required ABI (tag 0.2.11, 2021) has a real parsing bug (return foo(x) never produces a call_expression), unchanged through every release to date; the upstream fix requires ABI 14, which the Haskell tree-sitter package (latest: 0.9.0.3) can't load — the same ceiling already documented for cpp/nix. Closing this needs a cross-cutting ABI upgrade across every vendored grammar, not a Kotlin-only fix. Full investigation in tasks.md task 6.1.
• Also found while cross-checking the Fineract run (documented, not fixed — pre-existing, out of scope): identifier-less Tier-3 nodes (if_statement etc.) share a bare, kind-agnostic, cross-file-collidable id space; and Fineract's @Scheduled(fixedRate=...)/initialDelay usages aren't covered by the cron/fixedDelay-only distillation.

🤖 Generated with Claude Code