sigmap 7.30.0 → 8.0.0

package/CHANGELOG.md

@@ -10,6 +10,29 @@ Format: [Semantic Versioning](https://semver.org/)
 
 ---
 
+## [8.0.0] — 2026-07-04
+
+Major release — **v8.5 "Repo-Context Coverage & Test Discovery" (C1 + C2 + C3).** Marks the v8 milestone: the signature map now reaches beyond functions/classes/routes into the repo's operational surface, impl→test discovery is measured rather than best-effort, and every Evidence Pack file carries a risk label from a richer, precedence-ordered set. All zero-dependency, deterministic, and in-boundary with the North-Star constraints. **No breaking API changes** — the `8.0.0` bump aligns the published version with the roadmap's v8 framing; existing `riskLabel`/`relatedTests` consumers keep working.
+
+### Added
+- **C1 — Repo-context coverage expansion (#402):** four dedicated zero-dep map analyzers under `src/map/` (mirroring `route-table.js`), wired into `gen-project-map.js` and the MCP `get_map` `MAP_SECTIONS` — `env-schema.js` (**Environment variables** — env reads across JS/TS/Python/Ruby/Go + `.env.example` keys), `build-ci.js` (**Build & CI** — npm/pnpm scripts, GitHub Actions workflows, Makefile targets), `config-manifest.js` (**Config & manifests** — package manifests across npm/Python/Rust/Go/Maven/Gradle/Ruby/PHP + notable config files), and `migrations.js` (**Database migrations** — Rails/Alembic/Prisma/Flyway/timestamped-SQL detection). `PROJECT_MAP.md` and `get_map` now surface all four sections.
+- **C2 — Measured test discovery (#402):** `findRelatedTests` now normalizes cross-language test conventions (`test_x.py`↔`x.py`, `x_test.go`↔`x.go`, `XTest.java`↔`X.java`, `x.spec.ts`↔`x.ts`). New reproducible benchmark `scripts/run-test-discovery-benchmark.mjs` (`npm run benchmark:test-discovery`) scores it against an independent canonical-name gold oracle over `benchmarks/repos` — no LLM, pure string math — measuring **F1 98.0%, hit@1 97.4% across 28 repos / 3,701 gold pairs**. The headline number is surfaced in `benchmarks/latest.json` under `test_discovery`.
+- **C3 — Richer risk labels (#402):** `riskLabelFor` now returns the v8.5 set — `migration | payment | auth | security | public-api | config | test | generated | source` — with strict most-specific-risk precedence (a migration touching auth is still `migration`; payment/auth outrank the generic `security` bucket). `test`/`generated`/`config`/`source` semantics are preserved so `findRelatedTests` and the verifier keep working. Extended coverage in `test/integration/evidence-pack.test.js`, `project-map.test.js`, and `benchmark-latest.test.js`.
+
+### Fixed
+- **Comparison-chart correctness multiplier (#399):** corrected the stale answer-correctness multiplier badge in `docs/comparison-chart.svg` (×5.2 → ×6.8) to match the current task-success benchmark.
+
+## [7.31.0] — 2026-07-02
+
+Minor release — **identifier-aware BM25 re-ranker.** Plain exact-token TF-IDF missed queries whose terms live *inside* code identifiers — `component emit` never surfaced `componentEmits` because that is one token sharing no exact term with the query. This was the dominant retrieval-miss cause. The new ranker splits identifiers, stems lightly, boosts path tokens, and scores with length-normalized BM25. Deterministic, zero new dependencies, no LLM/embeddings.
+
+### Added
+- **Identifier-aware BM25 re-ranker (#395, #396):** new zero-dependency `src/retrieval/bm25.js` with (1) identifier-aware tokenization (split camelCase / snake_case), (2) light stemming (`emits` → `emit`, `options` → `option`), (3) path-token boost (filename weighed 3×), and (4) BM25 length-normalized scoring instead of raw TF-IDF. Wired into the core ranker (`src/retrieval/ranker.js`) as the base relevance score — so `sigmap ask`, `sigmap --query`, and MCP `query_context` all benefit — with the existing negative-signal penalty and recency/graph/learned boosts layered on top. Also drives the benchmark runner (`src/eval/runner.js`) and the dev retrieval benchmark.
+- **BM25 unit tests (#396):** `test/integration/bm25.test.js` covers tokenization, stemming, path boost, the `component emit` → `componentEmits` motivating case, and deterministic tie-breaking.
+
+### Changed
+- **Retrieval benchmark refreshed:** on the 18-repo / 90-task suite, hit@5 rose **75.6% → 86.7%** (retrieval lift 5.6× → 6.4×), with rank-1 gains on flask, spring-petclinic, rails, and svelte (60% → 100%). The task-completion proxy also improved (task success 52.2% → 67.8%, prompts/task 1.72 → 1.46) since it retrieves through the same ranker. Residual misses (vapor, serilog) are files whose signatures genuinely lack the query vocabulary — out of scope, they need semantic retrieval.
+
 ## [7.30.0] — 2026-06-23
 
 Minor release — **v8.0 E2 + E4 (the "Pivot"):** completes v8.0 by repositioning every public surface to the chosen framing — *"the deterministic, verifiable grounding layer for AI code work"* — and framing coding agents as **consumers, not competitors**. The Evidence Pack code (E1/E3/D3 + `mcp install`) already shipped in 7.27–7.29; this is the positioning half. Docs/strings only — no runtime behaviour change, zero new dependencies.

package/README.md

@@ -57,10 +57,10 @@ That map is exactly what agentic grep is worst at: reproducible, auditable conte
 
 **Proof it pays off** (full benchmark below):
 <!--SM:whyMetrics-->
-- **75.6% hit@5** — right file found in top 5 results (vs 13.6% baseline)
+- **86.7% hit@5** — right file found in top 5 results (vs 13.6% baseline)
 - **97.0% token reduction** — average across 21 real repos
-- **52.2% task success rate** — up from 10% without context
-- **1.72 prompts per task** — down from 2.84 (39.4% fewer retries)
+- **67.8% task success rate** — up from 10% without context
+- **1.46 prompts per task** — down from 2.84 (48.8% fewer retries)
 <!--/SM:whyMetrics-->
 - **<!--SM:languages-->33<!--/SM:languages--> languages supported** — TypeScript, Python, Go, Rust, Java, R, and more
 - **No vendor lock-in** — works with any AI assistant or local LLM
@@ -74,7 +74,7 @@ That map is exactly what agentic grep is worst at: reproducible, auditable conte
 | 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-->76%<!--/SM:hitWhole--> of the time, every symbol on a real line anchor |
+| ❌ "Trust me" AI answers | ✅ Grounded — right file in context <!--SM:hitWhole-->87%<!--/SM:hitWhole--> of the time, every symbol on a real line anchor |
 | ❌ Embeddings / vector DB required | ✅ Zero deps, no infra, fully offline |
 
 ---
@@ -98,13 +98,13 @@ Ask → Rank → Context → Validate → Judge → Learn
 
 <!--SM:benchmarkBlock-->
 ```
-Benchmark : sigmap-v7.30-main (21 repositories, including R language)
-Date      : 2026-06-23
+Benchmark : sigmap-v8.0-main (21 repositories, including R language)
+Date      : 2026-07-04
 
-Hit@5          : 75.6%   (baseline 13.6%  — 5.6× lift)
+Hit@5          : 86.7%   (baseline 13.6%  — 6.4× lift)
 Token reduction: 97.0%   (across 21 repos)
-Prompt reduction : 39.4% (2.84 → 1.72 prompts per task)
-Task success   : 52.2%   (baseline 10%)
+Prompt reduction : 48.8% (2.84 → 1.46 prompts per task)
+Task success   : 67.8%   (baseline 10%)
 Repos tested   : 21 (JavaScript, Python, Go, Rust, Java, R, C++, C#, Dart, Swift, Ruby, PHP, Scala, Kotlin, and more)
 ```
 <!--/SM:benchmarkBlock-->