0.2.7

[Timwood0x10]

CodeScope v0.2.7

First release since v0.2.6 (main, 2026-08-14): 41 commits, 394 files. Two themes: the architectural work for 0.3 (handle-based ABI, Windows parallel indexing), and silent failure — the cases where a tool answered confidently and wrongly, on the default path, for real projects.

New Features

• Handle-based engine ABI (TD-1, three knives) (engine/include/engine.h, engine/src/engine_context.*): engine_create() returns an opaque engine_t, all 73 stateful entry points take it, and the three process-global singletons are gone — no global object is constructed before main(), and a stale call site is a compile error. The still process-wide subsystems (verifier registry, async knowledge builder, index-progress and parse-failure buffers) are recorded in engine_context.h rather than implied. Evidence: a 43-tool differential matrix is byte-identical old vs new, and both binaries index one tree to the same call-edge set.
• index-parallel on Windows (TD-5) (server/src/scheduler/mapped_file.rs): the POSIX open/ftruncate/mmap sequence both schedulers open-coded now lives in one mapping layer with a Windows backend (CreateFileMappingW/MapViewOfFile), and per-run paths come from std::env::temp_dir() instead of /tmp. A windows-smoke job indexes a real project through both schedulers — it has not run yet, so Windows stays Beta (README).
• Graph-shape regression baselines (benchmarks/graph_baseline.sh, make bench-graph): six real checkouts, one per language, compared on files, entities, call edges, resolution_kind, resolve_strategy, deciding_factor and unresolved sites; exact rather than tolerance-based because the shape is reproducible. self 2274 · c-redis 27474 · go-tinygo 4313 · java-spring-petclinic 47 · rust-pyo3 2859 · ts-codebase-memory-mcp 8440 call edges.
• Counterfactual resolver-factor switch (CODESCOPE_RESOLVER_DISABLE_FACTOR=<Name>): drops one factor's weight and score at the single accumulation point, so a factor's real effect is measured by A/B instead of argued from the weight table — relation.reason's decided_by= names ImportMatch for every scored edge on this repository, while disabling factors shows which ones actually move the graph. No weight is changed until that table says otherwise.
• codescope parse-failures / reset-failures (the parse_failures table was write-only), get_communities restored (synchronous label propagation; deterministic; summary-first — the earlier revision returned ~100K tokens for one query), graph_query honours LIMIT and rejects trailing text instead of dropping it silently.

Bug Fixes

Indexing and symbol names

• A Maven tree indexed nothing: the engine's test/docs/samples skip collided with Java package components, so src/main/java/…/samples/petclinic/ was skipped at any depth and discover_modules reported no modules. Both entry points now share applyProjectLanguageContext(); spring-petclinic goes 0 → 30 files.
• Java methods were named after their return type (JavaVisitor::extractName accepted type_identifier and returned the first match): public String processUpdateForm(…) was recorded as an entity called String — 27 such entities in petclinic, 4 Pet, 3 LocalDate — while void methods were correct, and those type-named entities then collected edges. Names are read from the grammar's name field; petclinic 33 → 47 edges and find_definition("processUpdateForm") resolves.
• The same call shape resolved for Go, Java and Python and was dropped for C++, Rust and TypeScript: a parameter is a declaration, but only variable declarations fed the visitors' type table, so a call through a parameter carried a receiver with no type and the ambiguity gate abstained. All three now record parameter types; 6 of 6 languages resolve to the owner's method in test_resolver_language_consistency.
• arity and param_count had no producer: param_count was structurally 0 (no visitor emits a Parameter record) and arity — the input the resolver's overload step reads — was 0 for all 24604 goagent rows, so every candidate scored as "unknown arity". Both come from the CST declaration now, with a variadic tail or a default reported as unknown rather than a count that would penalise a valid call.
• TypeScript annotations lost array element types and leaked object-type fragments: extractTsTypeAnnotation unwrapped element and member types as if they were annotations, so Foo[] yielded nothing, and it returned raw text for shapes it does not model, so const tabs: { id: TabId; … }[] recorded the receiver type of tabs.map(…) as id: TabId. Element types resolve and anything that is not a plain type name is reported as unknown.
• Builtin-name filters deleted calls to the file's own functions (a user format(), valueOf(), write() or free() collided with the filter's list and produced no call record at all): each visitor now keeps the names the file declares, and the filter applies only to unqualified calls.
• C/C++ functions with a qualified return type were invisible (std::string, char *, namespace-nested): the name scan matched the return type and dropped the whole definition — on this repository engine/src/query/** contributed 3 definitions out of 1972 statements. Names are read from the declarator field.
• Aliased imports resolved to nothing (import { pick as pickLib } bound a local name with no record of the imported symbol) and a call omitting a default argument lost its edge (C++ declares the default in the header while the entity's arity comes from the definition).

Determinism — one tree, one binary, one graph: the two resolver loader tables are ordered by content (a worker-dependent rowid order decided which interface six call sites on goagent resolved to), and entity ids derive from the record's semantic identity instead of SQLite's scan order.

Silent success and silent drops — a run that indexed nothing no longer reports ok:true (0-file and no-symbol runs report ok:false/complete:false); failed CSR builds, bulk COMMITs, nested transactions, migrations and batch inserts roll back and report instead of committing whatever landed; index-parallel installs its database at the resolved path (it only reported a temp path, so the next command queried an empty DB) and runs the post-index pass, so the DB it hands back is searchable; force_index_files is no longer filtered at the graph stage (forcing engine/tests added 13174 records and zero entities); createSchema no longer has a use-after-free; the CMake generator is resolved from the build tree, because a mismatched generator silently kept new translation units out of the archive.

MCP surface — isError was serialized as is_error, so a spec-compliant client saw no failure flag for any tool result; an MCP-indexed project had no projects row (every entity landed under project_id = 0, next to a full graph); front-end build output was indexed (dist/: 700 minified nodes beside 224 real symbols); array-shaped node_types/edge_types filters were silently ignored; project_overview.entry_points read the legacy table and was always empty; verify_summary and verify_review scored the same empty input 1.0 and 0.0; ambiguous names returned an empty list with no next step.

Paths and tool contracts — detect_changes reported "nothing changed" for the input shape its own schema advertises and for any path spelling other than the stored one; find_definition / find_references / find_callers / find_callees missed a file named through a symlink; shortest_path and codescope_trace advertised no required arguments; a failing engine test could pass the gate (the CHECK counter was consulted only by return checkFailures() ? 1 : 0, which 8 test files never wrote).

🧹 Cleanup

• One JSON implementation (TD-3): every hand-built envelope migrated to util::JsonWriter, escaping converged on util::jsonEscapeString, and the duplicate escapers and declarations deleted.
• ~1.5k lines of dead code removed (scanner_visitor, the superseded resolver stack, the zero-call-site scorers, five duplicate copies of one scoring rule), the superseded Planner/VerdictBuilder chain deleted, and verify_statement retired now that verify_claim is the single entry point.

Testing

• Every engine test uses CHECK() (929 assertion sites migrated, no live assert() remains) and a recorded failure makes the binary exit non-zero — enforced by an exit hook, because 8 files printed a banner and returned 0.
• New suites: test_resolver_language_consistency (six languages, one call shape), test_import_alias_resolution, test_function_arity, test_receiver_type_canonical, test_dispatch_order_determinism, test_metrics_persist, test_force_index_files, test_engine_handles (instance and ABI boundaries), test_resolver_language_filter, test_gitignore_build_dirs, plus property-based tests over clamping, exclude-path escaping and MCP payload round-trips.
• make check green: 106 engine test binaries, 136 server tests; call-graph accuracy gate TP 36 / FP 0 / FN 0; make bench-graph matches all six baselines.