fix(halstead): vocabulary and decision batch 2026-10-10

.bca-baseline.toml

@@ -279,7 +279,7 @@ value = 42.0
 path = "big-code-analysis-ast/src/parser.rs"
 qualified = "Parser<T>::filters"
 metric = "halstead.effort"
-value = 87390.78947368421
+value = 90363.0
 
 [[entry]]
 path = "big-code-analysis-ast/src/preproc.rs"
@@ -1065,7 +1065,7 @@ value = 6.0
 path = "src/ops.rs"
 qualified = "ops_inner"
 metric = "halstead.effort"
-value = 70903.25755192844
+value = 84452.11438842113
 
 [[entry]]
 path = "src/output/checkstyle.rs"
@@ -1179,7 +1179,7 @@ value = 5.0
 path = "src/spaces/compute.rs"
 qualified = "compute_per_node"
 metric = "halstead.effort"
-value = 63645.7536117776
+value = 64334.51855807396
 
 [[entry]]
 path = "src/spaces/compute.rs"
@@ -1191,7 +1191,7 @@ value = 7.0
 path = "src/spaces/compute.rs"
 qualified = "metrics_inner"
 metric = "halstead.effort"
-value = 121169.85621346734
+value = 141885.98828207294
 
 [[entry]]
 path = "src/spaces/compute.rs"

.claude/rules/testing.md

@@ -651,6 +651,9 @@ tree at `ref` as well (`git archive` into a scratch directory), computes
 which single-language builds compile each test in each, and fails on a
 test that still exists but stopped being built somewhere. The probe set
 comes from the language table, so a new language extends it for free.
+A test's membership includes the `cfg` on the `mod` line that loads its
+file, so narrowing a root `mod` gate in a `tests/*/main.rs` is reported
+too; before that, #1567's drop of `kotlin` from one went unseen.
 
 CI runs it per pull request against the base branch head — the PR's
 `base.sha`, not `git merge-base`, because the job checks out shallow and

.github/workflows/ci.yml

@@ -320,21 +320,29 @@ jobs:
         with:
           fetch-depth: 1
           submodules: recursive
-      # llvm-tools-preview supplies llvm-profdata/llvm-cov for source-based
-      # coverage instrumentation.
+      # Pin the measurement toolchain independently of the stable build jobs:
+      # compiler changes can move coverage regions without changing test coverage.
       - uses: dtolnay/rust-toolchain@02cb101ec7c40f2c49e1d9714d64511d8e1b74de # stable tip
         with:
-          toolchain: stable
+          toolchain: "1.99.1"
           components: llvm-tools-preview
       # Coverage builds carry -C instrument-coverage, so their artifacts are
       # incompatible with the regular build cache; a distinct key keeps the
       # two from thrashing each other.
       - uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2.9.2
         with:
-          key: coverage
+          key: coverage-1.99.1-llvm-cov-0.9.1
       - uses: taiki-e/install-action@9983c65e42da123ff25d1f78505eb6de315aa172 # v2.87.20
         with:
-          tool: cargo-nextest,cargo-llvm-cov
+          tool: cargo-nextest@0.9.146,cargo-llvm-cov@0.9.1
+      - name: Record coverage provenance
+        run: |
+          {
+            git rev-parse HEAD
+            rustc --version --verbose
+            cargo llvm-cov --version
+            cargo nextest --version
+          } > coverage-provenance.txt
       # nextest is the same runner the `test` job uses, so instrumenting it
       # keeps the coverage numbers aligned with what CI actually exercises.
       # --no-report defers reporting so the same profile data can be emitted
@@ -344,6 +352,8 @@ jobs:
       # nightly, and pairing nightly with the workflow-wide `-D warnings`
       # would red-X CI on unrelated new nightly lints. The `test` job still
       # runs doctests for correctness — only their coverage is omitted.
+      - name: Clean coverage data
+        run: cargo llvm-cov clean --workspace
       - name: Run tests with coverage instrumentation
         run: cargo llvm-cov --no-report nextest --all-features --workspace --locked
       # Codecov's native format preserves region coverage (lcov would drop to
@@ -358,6 +368,19 @@ jobs:
       # 103 files that are in fact ~98% covered. Keep the two in step.
       - name: Generate Codecov report
         run: cargo llvm-cov report --workspace --codecov --output-path codecov.json
+      - name: Generate raw LLVM coverage report
+        run: cargo llvm-cov report --workspace --json --output-path coverage-llvm.json
+      - name: Archive coverage reports
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: coverage-${{ github.sha }}
+          path: |
+            codecov.json
+            coverage-llvm.json
+            coverage-provenance.txt
+          if-no-files-found: error
+          retention-days: 30
       # Per-file coverage table in the run's Summary tab, so reviewers can
       # read coverage without leaving GitHub or waiting on the Codecov PR
       # comment.

.pre-commit-config.yaml

@@ -7,7 +7,8 @@ repos:
     -   id: check-symlinks
     -   id: check-yaml
     -   id: check-json
-        exclude: enums/templates/json.json
+        # Pyright accepts JSONC comments; py-typecheck validates its configuration.
+        exclude: ^(enums/templates/json\.json|big-code-analysis-py/pyrightconfig\.json)$
 -   repo: https://github.com/marco-c/taskcluster_yml_validator
     rev: v0.0.12
     hooks:
@@ -559,31 +560,17 @@ repos:
       # because pre-commit's isolated venv cannot see the project's
       # `maturin develop`'d `_native` extension or the rest of the
       # dev environment — under that isolation, `--strict`'s whole-
-      # program analysis is silently meaningless. The system hook runs
-      # against the contributor's PATH-resolved `mypy`, same pattern
-      # as the clippy/udeps/test hooks above.
-      #
-      # pyright is intentionally NOT a pre-commit hook: it bundles a
-      # Node runtime that pre-commit cannot hermetically isolate and
-      # suffers the same isolated-venv problem mypy does. The CI
-      # `python-lint` job covers it.
+      # program analysis is silently meaningless. The Makefile resolves
+      # mypy and pyright from the bootstrapped project venv before PATH,
+      # matching the canonical gate without requiring manual activation.
       - id: py-mypy
         name: py-mypy
         language: system
         # Match both `.py` source and `.pyi` stubs — editing
         # `_native.pyi` must also re-run mypy so the stub stays in
         # sync with the Rust-side ABI it describes.
-        files: '^big-code-analysis-py/.*\.pyi?$'
-        # Use --config-file + absolute-ish paths so the invocation is
-        # cross-platform — `bash -c "cd X && ..."` fails on native
-        # Windows pre-commit (no bash on cmd/PowerShell PATH).
-        entry: mypy
-        args:
-          - --strict
-          - --config-file=big-code-analysis-py/pyproject.toml
-          - big-code-analysis-py/python
-          - big-code-analysis-py/tests
-          - big-code-analysis-py/examples
+        files: '^big-code-analysis-py/(.*\.pyi?|pyrightconfig\.json|pyproject\.toml)$'
+        entry: make py-typecheck
         pass_filenames: false
 
 default_language_version:

AGENTS.md

@@ -681,7 +681,9 @@ stranded `.snap.new` drift that blocks CI on every subsequent change.
 The limits `bca init` scaffolds are defined once, in
 `big-code-analysis-cli/src/default_thresholds.rs`. They are derived
 from published thresholds plus a 20-language corpus measurement (#1140),
-and each carries its rationale inline. Change them there, not in a
+and each carries its rationale inline. `halstead.effort` was re-measured
+in #1569 on a rebuilt corpus recorded in
+[`docs/development/threshold_corpus.md`](docs/development/threshold_corpus.md). Change them there, not in a
 copy: the scaffolded `bca.toml` renders its `[thresholds]` block from
 that table, so there is no second copy of the numbers to drift.
 
@@ -709,8 +711,8 @@ Three things to know before quoting a number at anyone:
 - **These are language-agnostic defaults, and no language is
   agnostic.** Measured 97.5th-percentile `cognitive` runs from 4 (C#)
   to 50 (C). C, Tcl, Bash, Lua, Perl, and Go want roughly double the
-  defaults; Java, Ruby, Kotlin, C#, and Elixir sit far under them. The
-  per-language table lives in
+  `cognitive` default; Java, Ruby, Kotlin, C#, and Elixir sit far
+  under it. The per-language table lives in
   [Choosing thresholds](big-code-analysis-book/src/recipes/thresholds.md).
 - **The right limit depends on the job.** A blocking CI gate, an agent
   edit loop, a legacy triage pass, and a safety-critical build want