sigmap 8.31.0 → 8.32.1

package/CHANGELOG.md

@@ -10,6 +10,34 @@ Format: [Semantic Versioning](https://semver.org/)
 
 ---
 
+## [8.32.1] — 2026-09-13
+
+
+### Fixed
+- **Retraction of a v8.32.0 claim.** That release stated "every extractor now discloses what a ceiling dropped". It did not. `vue.js` was registered but unreachable (`.vue` dispatches to `vue_sfc`), so disclosure was added to dead code while the live `.vue` handler kept truncating silently; `.tsx` (every React component), `.properties`, `.toml` and `.md` still cut output with a bare `slice()`; and `r.js` called `capWithNotice` but eight inner caps stopped collection at the ceiling so it never fired — and when forced to, reported `+1 more` where 50 signatures were hidden. All fixed, ceilings unchanged, with a test that fails against the pre-fix extractors. The root cause was that three of those languages have no test fixture (#582, #583, #584, #588)
+- The walk depth hid most code in JVM package layouts: `maxDepth: 6` indexed **6 of 47** Java files on spring-petclinic, because Java puts one directory per package segment. #561 had already raised the *graph* walk to 12, so extraction was the shallower half of an inconsistent pair. Depth now rises to 12 only for JVM layouts; an explicit `maxDepth` always wins. Gated JVM corpus 16.4% → 23.0% (#590)
+
+### Changed
+- Published retrieval hit@5 moves 81.1% → 78.9%. One of 18 repos accounts for it: spring-petclinic 100% → 60%, previously measured against an index holding 6 of 47 Java files. The leak-free `mined` corpus is flat and the leak-free `jvm` corpus improves, so the prior figure was inflated by under-indexing rather than this being a ranking regression. The two affected tasks are tracked as a ranking weakness the missing files were concealing (#592)
+
+---
+
+## [8.32.0] — 2026-09-12
+
+### Added
+- A gated **JVM retrieval corpus** — 61 tasks mined from `spring-petclinic` (32) and `akka` (29), leak-checked so no query shares a stemmed token with its expected file's basename. It scores against *other* repos, which puts it outside the feedback loop that makes the `hard` split move whenever sigmap's own source changes. `scripts/mine-corpus.mjs` gained `--repo` and gates its sigmap-specific rules on the target being sigmap itself, with byte-identical default output (#577, closes #575)
+- `test/integration/docs-markdown.test.js` — guards the VitePress sources against the two failure modes that can break the docs build *after* a release is already tagged: unbalanced code fences, and `{{ }}` outside a fenced block, which Vue parses as an interpolation (#574, closes #573)
+
+### Fixed
+- Every extractor now **discloses what a ceiling dropped** instead of truncating silently. 20 extractors gained `… +N more` markers on their member and per-file caps (23 now carry them in total). The ceilings themselves are unchanged — raising them is a separate, measured decision (#578, #576)
+  > **Correction (2026-09-13):** "every extractor" was wrong as published. `vue.js` was dead code — `.vue` dispatches to `vue_sfc` — so one of the 20 was unreachable; four reachable extractors (`.tsx`, `.properties`, `.toml`, `.md`) still truncated silently; and `r.js`'s disclosure was defeated by eight inner caps. Corrected in #589 (#582, #583, #584).
+- The retrieval gate reused a gitignored index, so a "regression" could be pure staleness — this produced three separate false investigations, including one re-baseline. It now regenerates every index it scores, including one per JVM repo driven from `benchmarks/config-overrides.json` rather than a hand-written config (#579)
+- The `sigmap lines` CLI example nested a fenced block inside a text fence; the inner fence closed early and a brace expression became a Vue interpolation. This failed the Pages build *after* v8.31.0 was tagged and published (#574)
+
+### Changed
+- The `hard` corpus is now **reported but not enforced** against the previous run — only against its 70% floor. Proven necessary in CI: a probe branch containing one two-assertion test file and no source change scored 75.6% → 74.4% and failed the gate, because `hard` scores sigmap against its own source and BM25 statistics shift with the indexed file set. `mined` and `jvm` keep `--no-regress` (#579)
+- JVM baseline re-recorded at 16.4% hit@5 (from 18.0%) — a single task. `akka:m018` expects `Logging.scala`, which holds a class the 8-member ceiling truncates, so disclosure adds one `… +N more methods` line and BM25's length normalisation drops it from rank 5 to 6. A fix excluding markers from the scored term space did not move the number and was reverted rather than left in as unexplained complexity (#578)
+
 ## [8.31.0] — 2026-09-08
 
 ### Added

package/README.md

@@ -59,12 +59,12 @@ That map is exactly what agentic grep is worst at: reproducible, auditable conte
 
 **Proof it pays off** (full benchmark below):
 <!--SM:whyMetrics-->
-- **81.1% hit@5** — right file in top 5 results (vs 44.0% single-shot grep baseline — 1.73× lift)
+- **78.9% hit@5** — right file in top 5 results (vs 44.0% single-shot grep baseline — 1.73× lift)
 - **96.8% token reduction** — average across 21 real repos
-- **64.8% task-success proxy** — modeled from retrieval tiers, not measured LLM sessions
-- **1.54 prompts per task** — down from 2.84 (45.7% fewer retries, modeled)
+- **62.9% task-success proxy** — modeled from retrieval tiers, not measured LLM sessions
+- **1.58 prompts per task** — down from 2.84 (44.4% fewer retries, modeled)
 <!--/SM:whyMetrics-->
-- **<!--SM:languages-->33<!--/SM:languages--> languages supported** — TypeScript, Python, Go, Rust, Java, R, and more
+- **<!--SM:languages-->32<!--/SM:languages--> languages supported** — TypeScript, Python, Go, Rust, Java, R, and more
 - **No vendor lock-in** — works with any AI assistant or local LLM
 - **No API costs** — use local models (Ollama, llama.cpp, vLLM) with zero token fees
 - **Full privacy** — keep your code and context on your machine
@@ -98,7 +98,7 @@ sigmap verify answer.md --report        # standalone red/amber/green HTML report
 | Without SigMap | With SigMap |
 |---|---|
 | ❌ Non-reproducible agent guesses | ✅ Deterministic map — same input, same output, every time |
-| ❌ "Trust me" AI answers | ✅ Grounded — right file in context <!--SM:hitWhole-->81%<!--/SM:hitWhole--> of the time, every symbol on a real line anchor |
+| ❌ "Trust me" AI answers | ✅ Grounded — right file in context <!--SM:hitWhole-->79%<!--/SM:hitWhole--> of the time, every symbol on a real line anchor |
 | ❌ Embeddings / vector DB required | ✅ Zero deps, no infra, fully offline |
 
 ---
@@ -122,13 +122,13 @@ Ask → Rank → Context → Validate → Judge → Learn
 
 <!--SM:benchmarkBlock-->
 ```
-Benchmark : sigmap-v8.31-main (21 repositories, including R language)
-Date      : 2026-09-08
+Benchmark : sigmap-v8.32-main (21 repositories, including R language)
+Date      : 2026-09-12
 
-Hit@5          : 81.1%   (grep-agent baseline 44.0%  — 1.73× lift)
+Hit@5          : 78.9%   (grep-agent baseline 44.0%  — 1.73× lift)
 Token reduction: 96.8%   (across 21 repos)
-Prompt reduction : 45.7% (2.84 → 1.54 prompts per task, modeled)
-Task success   : 64.8%   (proxy — modeled from retrieval tiers)
+Prompt reduction : 44.4% (2.84 → 1.58 prompts per task, modeled)
+Task success   : 62.9%   (proxy — modeled from retrieval tiers)
 Repos tested   : 21 (JavaScript, Python, Go, Rust, Java, R, C++, C#, Dart, Swift, Ruby, PHP, Scala, Kotlin, and more)
 ```
 <!--/SM:benchmarkBlock-->
@@ -332,7 +332,7 @@ sigmap --health
 | Benchmark methodology | [benchmark.html](https://sigmap.io/guide/benchmark.html) |
 | Config reference | [config.html](https://sigmap.io/guide/config.html) |
 | Roadmap | [roadmap.html](https://sigmap.io/guide/roadmap.html) |
-| <!--SM:languages-->33<!--/SM:languages--> languages | [generalization.html](https://sigmap.io/guide/generalization.html) |
+| <!--SM:languages-->32<!--/SM:languages--> languages | [generalization.html](https://sigmap.io/guide/generalization.html) |
 
 ---
 
@@ -389,7 +389,7 @@ See [.github/PULL_REQUEST_TEMPLATE.md](.github/PULL_REQUEST_TEMPLATE.md) for the
 
 ---
 
-## <!--SM:languages-->33<!--/SM:languages--> languages
+## <!--SM:languages-->32<!--/SM:languages--> languages
 
 TypeScript · JavaScript · Python · Java · Kotlin · Go · Rust · C# · C/C++ · Ruby · PHP · Swift · Dart · Scala · Vue · Svelte · HTML · CSS/SCSS · YAML · Shell · SQL · GraphQL · Terraform · Protobuf · Dockerfile · TOML · XML · Properties · Markdown · R · GDScript