fix: batch 2026-11-03-2 (8 issues)

.bca-baseline.toml

@@ -259,12 +259,6 @@ qualified = "RubyCode::get_op_type"
 metric = "halstead.effort"
 value = 101217.532578918
 
-[[entry]]
-path = "big-code-analysis-ast/src/lang_helpers/doc_comment.rs"
-qualified = "<file>"
-metric = "loc.ploc"
-value = 642.0
-
 [[entry]]
 path = "big-code-analysis-ast/src/lang_helpers/doc_comment.rs"
 qualified = "bind_leading"
@@ -775,12 +769,6 @@ qualified = "assemble_globals"
 metric = "nargs"
 value = 6.0
 
-[[entry]]
-path = "big-code-analysis-cli/src/walk.rs"
-qualified = "classify_dropped_child"
-metric = "nargs"
-value = 5.0
-
 [[entry]]
 path = "big-code-analysis-cli/src/walk.rs"
 qualified = "expand_seed_paths"
@@ -1267,6 +1255,12 @@ qualified = "RustCode::compute"
 metric = "nargs"
 value = 5.0
 
+[[entry]]
+path = "src/metrics/loc/shared.rs"
+qualified = "credit_bare_continuation_rows"
+metric = "nargs"
+value = 5.0
+
 [[entry]]
 path = "src/metrics/loc/tcl.rs"
 qualified = "TclCode::compute"

.claude/skills/add-lang/SKILL.md

@@ -531,7 +531,7 @@ Create the file, add `mod <lang>;` to `checker.rs`, and write
   `holds_doc_comment_in_body`) — what the language's own
   documentation tool reads (#1597). They feed `dloc`; the defaults
   answer `false`, which is right for a language with no doc
-  convention. Where one exists, `lang_helpers/doc_comment.rs` has the
+  convention. Where one exists, `lang_helpers/doc_text.rs` has the
   shared opener tests.
 - `is_func_space` / `is_func_space_with_code`, `is_func`,
   `is_closure` — forward to the table: `crate::roles::<lang>::opens_space`

.github/workflows/ci.yml

@@ -427,7 +427,18 @@ jobs:
       # runs doctests for correctness — only their coverage is omitted.
       - name: Clean coverage data
         run: cargo llvm-cov clean --workspace
+      #
+      # LLVM_PROFILE_FILE_NAME drops the `%p` cargo-llvm-cov puts in its
+      # default name. nextest runs every test in a process of its own, so
+      # one raw profile per process came to ~8,000 files and 15 GB, and
+      # the job ran the runner out of disk (PR #1804). Without `%p`, every
+      # process of one test binary merges on-line into `%4m`'s pool of
+      # four files: LLVM locks the file and adds the counters, so no count
+      # is lost. Measured on that PR: 101 files and 62 MB, with every
+      # file's covered lines and regions identical to the default naming.
       - name: Run tests with coverage instrumentation
+        env:
+          LLVM_PROFILE_FILE_NAME: big-code-analysis-%4m.profraw
         run: cargo llvm-cov --no-report nextest --all-features --workspace --locked
       # Codecov's native format preserves region coverage (lcov would drop to
       # line-only).
@@ -739,7 +750,8 @@ jobs:
       # SHA256, mirroring the actionlint and checkmake patterns below.
       # Bump RUMDL_SHA256 in lockstep with `RUMDL_VERSION` (and with the
       # `rumdl = …` pin in `mise.toml` and the `Dockerfile`'s
-      # `RUMDL_VERSION` + per-arch SHA256 table). To regenerate the
+      # `RUMDL_VERSION` + per-arch SHA256 table); `make
+      # check-rumdl-lockstep` fails when they disagree. To regenerate the
       # SHA256 after a version bump (swap the triple for
       # `aarch64-unknown-linux-gnu` to get the Dockerfile's arm64 entry):
       #   curl -fsSL https://github.com/rvben/rumdl/releases/download/v${RUMDL_VERSION}/rumdl-v${RUMDL_VERSION}-x86_64-unknown-linux-gnu.tar.gz.sha256
@@ -1013,6 +1025,15 @@ jobs:
       # gate without surfacing the regression in CI.
       - name: check-ruff-lockstep self-tests (explicit)
         run: python3 -m unittest -q utils/check-ruff-lockstep-test.py
+      # Defensive twin for the rumdl-lockstep gate (#1779): mise.toml,
+      # the `Install rumdl` step above and the Dockerfile must pin one
+      # rumdl version, and the Dockerfile's x86_64 hash must equal
+      # `RUMDL_SHA256`. Comments were the whole mechanism before, and
+      # #1653's bump nearly left the Dockerfile behind.
+      - name: check-rumdl-lockstep (explicit)
+        run: python3 utils/check-rumdl-lockstep.py
+      - name: check-rumdl-lockstep self-tests (explicit)
+        run: python3 -m unittest -q utils/check-rumdl-lockstep-test.py
       # Defensive twin for the enums-codegen-drift gate (#405):
       # running any grammar regen previously regenerated
       # `src/c_langs_macros/*.rs` to a pre-optimization form

.pre-commit-config.yaml

@@ -295,6 +295,27 @@ repos:
         entry: python3 -m unittest -q utils/check-ruff-lockstep-test.py
         pass_filenames: false
 
+      # rumdl version lockstep (#1779). mise.toml, the CI lint job and
+      # the Dockerfile each pin rumdl, and the Dockerfile's x86_64 hash
+      # must equal CI's. #1653's bump nearly left the Dockerfile behind.
+      # .rumdl-lockstep-baseline.txt records the last accepted version and
+      # hash, so a bump that keeps the old hash fails too.
+      - id: check-rumdl-lockstep
+        name: check-rumdl-lockstep
+        language: system
+        files: '^(mise\.toml|Dockerfile|\.github/workflows/ci\.yml|\.rumdl-lockstep-baseline\.txt|utils/check-rumdl-lockstep\.py)$'
+        entry: python3 utils/check-rumdl-lockstep.py
+        pass_filenames: false
+
+      # Self-tests for the rumdl-lockstep gate. Runs on edits to the
+      # script or its test file.
+      - id: check-rumdl-lockstep-test
+        name: check-rumdl-lockstep-test
+        language: system
+        files: '^utils/check-rumdl-lockstep(-test)?\.py$'
+        entry: python3 -m unittest -q utils/check-rumdl-lockstep-test.py
+        pass_filenames: false
+
       # Publish metadata for the crates.io uploads (#1224). The four
       # top-level crates pin internal deps at `=<version>`, so none can
       # be `cargo publish --dry-run`-ed before the tag that publishes

.rumdl-lockstep-baseline.txt

@@ -0,0 +1,13 @@
+# rumdl lockstep baseline: the rumdl release and its x86_64 tarball
+# SHA256 that check-rumdl-lockstep.py last accepted (#1779).
+# Maintained by `python3 utils/check-rumdl-lockstep.py --update`; see the
+# script's docstring.
+#
+# The gate fails when the version pinned in mise.toml, ci.yml and the
+# Dockerfile has moved but the hash in ci.yml and the Dockerfile is still
+# the one recorded here: the bump kept the old release's hash. A bump
+# that pastes the new release's hash into both files fails too, until
+# `--update` records the new pair; commit this file with that bump.
+#
+# One record: `<version> <sha256>`.
+0.3.2 f527d1374765d826c1627ceab3d1f46e548268731a495b5b12a3af529b3d79e8

AGENTS.md

@@ -97,6 +97,7 @@ and `cargo run -p big-code-analysis-web --`.
   `check-grammar-marker-sync.py`, `check-enums-codegen-drift.sh`,
   `check-grammar-crate.py`, `check-grammars-crates.sh`,
   `check-excluded-manifests.py`, `check-ruff-lockstep.py`,
+  `check-rumdl-lockstep.py`,
   `check-publish-metadata.py`, `check-test-lang-gates.py`,
   `check-role-census.py`, `check-role-coverage.py`,
   `check-role-bypass.py` (the three role-table gates, all reading

CHANGELOG.md

@@ -217,6 +217,18 @@ for historical reference.
   (`EISDIR` on unix) rather than a skip. The Python
   `Ast.from_path` `ValueError` names the reason.
 
+- **`make check-rumdl-lockstep`** (`utils/check-rumdl-lockstep.py`,
+  #1779) fails when the rumdl pin in `mise.toml`, the CI lint job's
+  `RUMDL_VERSION` / `RUMDL_SHA256` and the `Dockerfile` disagree on the
+  version, naming the file that differs, or when the `Dockerfile`'s
+  x86_64 SHA256 differs from CI's. `.rumdl-lockstep-baseline.txt`
+  records the version and hash it last accepted, so a bump that moves
+  the version but keeps the old hash fails by name, and an honest bump
+  fails until `python3 utils/check-rumdl-lockstep.py --update` records
+  the new pair in the same commit. It is wired into `make lint`, `make pre-commit`, `make ci`,
+  the pre-commit hooks (which also run on an edit to the baseline) and
+  the CI lint job.
+
 ### Changed
 
 - **The role-bypass baseline is empty** (#1701, part of #1690). The
@@ -901,6 +913,14 @@ for historical reference.
   whose predicates call it (metrics, `ops`, `find` and `count`) all
   lend the cache. (#1785)
 
+- **`bca check` measures its ignored-file summary on the `--jobs`
+  thread budget** (#1307). Walked directories were re-listed serially
+  after the walk; they are now claimed per directory by scoped worker
+  threads, which cuts a 65,720-directory tree from 340 ms to 143 ms.
+  The `--report-skipped` listing is byte-identical, including which
+  spelling of a path reached by overlapping seeds is printed (the
+  first-named seed's).
+
 ### Deprecated
 
 - `FromPathError::Unreadable`: superseded by `Skipped(SkipReason)`;
@@ -1187,9 +1207,8 @@ for historical reference.
   than an identifier walk, because `[temp.pre]/5` allows a template
   declaration at most one declarator and a walk would read an array
   bound, an initialiser's callee or a function pointer's parameter
-  name. Cpp and Mozcpp alike. A pointer-to-member-function data member
-  (`static int (S::*fp)(T);`, and the plain `int (S::*pmf)(int);`)
-  still reads as a method — #1718.
+  name. Cpp and Mozcpp alike. A templated pointer-to-member-function
+  data member (`static int (S::*fp)(T);`) is left to #1718, below.
 
 - **Ruby's `npm` counts a `def` wherever the class body's statements
   reach it** (#1401). The member walk read the body's direct children
@@ -1332,6 +1351,52 @@ for historical reference.
   stdin, so the after side applied no exclusions and reported the files
   they should have hidden as added.
 
+- `utils/perturb.py` refuses a `cargo test` argv that can run more than
+  one test target without `--no-fail-fast`, instead of reporting the
+  guards in later targets as uncaught (#1787). It sees through a wrapper
+  in front of cargo (`timeout`, `env`, `nice`, `rustup run`), counts a
+  `-p` inside a short-option cluster (`-qpa`), and refuses a second
+  `--target`.
+
+- `npa` / `npm` (C++, Mozcpp): a templated pointer-to-member data
+  member such as `template<class T> static int (S::*fp)(T);` is counted
+  as an attribute instead of a method, and the methods
+  `template<class T> int (*S::*getPp())(T);` and
+  `template<class T> static int S::*getPd();` as methods instead of
+  attributes. A method returning a member-function pointer stays a
+  method, and the parenthesised non-templated `int (S::*pmf)(int);` was
+  already correct (#1718). A member function template whose every
+  parameter parses as an expression, such as
+  `template<class T> void connect(T* obj, void (T::*slot)());`, is
+  counted as a method instead of an attribute: both grammars read it as
+  a variable with a parenthesised initialiser, which a class member
+  cannot have. Still misread by the grammars' error recovery, as before:
+  the unparenthesised `int S::*getPd();` as data, and
+  `static constexpr int (S::*fp)(int) = nullptr;` as a method.
+
+- Ruby `npa` counts `attr_*` macros and `@x` / `@@x` class-scope
+  assignments wherever the class body runs them: under a conditional,
+  loop, `begin` / `rescue` / `ensure`, `case` arm, statement modifier,
+  parentheses, an assignment's right side, or a call's argument. A bare
+  `private` / `public` / `protected` shadowed by a local of the same
+  name — bound by an assignment, `for` or `rescue =>` the walk reads —
+  no longer changes attribute visibility, and a marker inside a
+  wrapper or in expression position now sets it, as it already did for
+  `npm`. The attribute and method walks share one traversal and agree
+  on a body's visibility (#1719).
+
+- `loc` (Bash, Tcl, iRules): a line-continuation row holding only the
+  `\` counts as code instead of blank, or as comment where it continues
+  a Tcl or iRules comment (`# note \`), in its enclosing function, proc
+  or event handler as well as the file (#1445). One that continues
+  nothing counts as the line it starts once joined to the next row: a
+  comment where that row is one (`\` + `# note`), blank where it is
+  empty, code otherwise. Where CRLF line endings
+  reach the parser unchanged (the library's `analyze`, Python's
+  `Ast.parse`; `bca` converts them first), the Tcl and iRules grammars
+  still lex such a row as a word (#1802). The C family, Ruby and Elixir
+  splice the row the same way and still count it as blank (#1801).
+
 ## [2.4.0] - 2026-10-21
 
 ### Added

Dockerfile

@@ -251,6 +251,8 @@ EOF
 
 # ---------------------------------------------------------------------------
 # rumdl (Markdown linter/formatter) — SHA256-verified release tarball.
+# `make check-rumdl-lockstep` holds RUMDL_VERSION and the x86_64 hash to
+# the CI lint job's pin, and the version to mise.toml (#1779).
 # ---------------------------------------------------------------------------
 RUN <<'EOF'
 set -eux