sigmap 8.29.0 → 8.31.0

package/CHANGELOG.md

@@ -10,6 +10,30 @@ Format: [Semantic Versioning](https://semver.org/)
 
 ---
 
+## [8.31.0] — 2026-09-08
+
+### Added
+- `sigmap lines <file> <start>-<end>` — the CLI twin of the `get_lines` MCP tool, for environments where MCP is unavailable. Accepts a `:94` anchor pasted straight off a signature with `--context N`. Delegates to the same handler as MCP, so it shares the project-root sandbox, EOF clamping and secret redaction (#568)
+- Spring interface calls now link to their implementation, so blast radius on the class that owns the code is no longer empty. Exactly one implementation resolves outright; several resolve only via a single `@Primary`; anything still ambiguous produces no edge rather than a guess (#565)
+- A JVM call-graph gate (`npm run validate:callgraph-jvm`) asserting named caller→callee pairs and edge volume against a committed baseline — offline and deterministic. Verified to fail on a simulated revert (#567)
+
+### Fixed
+- The dependency graph was **empty** on Maven/Gradle repos: `buildFromCwd` hard-coded `srcDirs` to `src`/`app`/`lib`/`R`/`inst` and never read the project config, then capped the walk at 8 directories. Java package-import resolution already worked but was never reached. On a 524-file Spring repo: 0 → 524 nodes, 341 with importers (#561)
+- The Java call graph discarded every `receiver.method(` call — 58% of call sites in a real Spring module — so controller→service edges did not exist. Receiver types are now resolved from field and local declarations; unresolvable receivers still produce no edge. Interface method declarations are indexed so calls to them have a target. On the same repo: 0 → 10,213 edges (#563)
+
+### Changed
+- Retrieval baseline re-recorded. The hard split moved 76.7% → 74.4% purely because the index is regenerated under a token budget and this release added ~1,000 lines; holding the index fixed, pre- and post-merge code produce identical results and the same 23 misses. MRR rose 0.639 → 0.644 (#569)
+
+## [8.30.0] — 2026-09-07
+
+### Added
+- `sigmap-task` — an invokable prompt skill that drives the full grounding loop from the CLI, for environments where MCP is unavailable. Installs to `.github/prompts/sigmap-task.prompt.md` for Copilot (`/sigmap-task`) and through the normal skill path for other clients (#554)
+
+### Fixed
+- Java extractor: lifted three hard-coded caps that hid 85% of the API surface on real Java repos — an 8-member-per-class limit applied silently, a 25-signature-per-file cap that shadowed `maxSigsPerFile`, and a 5,000-character class-body scan limit. Omissions are now disclosed with a `… +N more` marker, matching the JS/TS path (#552)
+- Retrieval: generated data holders no longer outrank real logic. A file whose members are overwhelmingly trivial accessors is demoted like other generated code, with an escape hatch when the query asks for an entity, model, DTO or accessor. Entities remain retrievable by their own symbols (#552)
+- `sigmap mcp install vscode` wrote a config VS Code ignores — it emitted a top-level `mcpServers` key where VS Code requires `servers` with an explicit `type`. A config written by an earlier version is now migrated rather than left in place (#557)
+
 ## [8.29.0] — 2026-09-01
 
 Minor release — **"Retrieval Index Split" (v8.29)**: the ranker stops reading the token-budgeted prompt artifact, and the ranking features that were silently inert start executing. Plus the first benchmark corpus this project did not author itself.

package/README.md

@@ -122,8 +122,8 @@ Ask → Rank → Context → Validate → Judge → Learn
 
 <!--SM:benchmarkBlock-->
 ```
-Benchmark : sigmap-v8.29-main (21 repositories, including R language)
-Date      : 2026-09-01
+Benchmark : sigmap-v8.31-main (21 repositories, including R language)
+Date      : 2026-09-08
 
 Hit@5          : 81.1%   (grep-agent baseline 44.0%  — 1.73× lift)
 Token reduction: 96.8%   (across 21 repos)