test(features): build tests/api under single-language feature sets

[dekobon]

Fixes #1426.

What was broken

Sweeping all 22 single-language subsets of -p big-code-analysis --all-targets under clippy -D warnings found five failing legs, not the
two the issue reported, from four distinct cfg-union defects — every one
of them in tests/api/ast_seam_test.rs, and every one a hand-maintained union
that had drifted from the items it gates.

Leg(s)	Defect	Diagnostic
c	PathBuf / MetricsOptions gated any(rust, python, cpp), used by a c-gated test — union too narrow	E0433
cpp, python	ops_tree_contains_operator gated on three languages with one rust-gated caller — union too wide	dead_code
tcl, irules	two-row Tcl-family fixture table with cfg-gated rows collapses to a singleton	clippy::single_element_loop
(no leg)	LANG / Source carried the same c-shaped drift, but no leg could see it	E0433, only under c,javascript

The cpp / python class is the exact mirror of the reported c one: the
same union, wrong in the other direction.

The library itself always built, so no published configuration was
affected. This was a contributor-facing break, and a costly one, because
--features go is the subset .claude/rules/testing.md tells contributors to
run and the compile error landed in a file unrelated to their change — the
same false-attribution failure as #1220 / PR #1221 and #1286, one layer down.

The fourth defect is the interesting one

LANG / Source was gated:

#[cfg(any(rust, python, c, cpp, tcl, irules, not(feature = "javascript")))]

not(feature = "javascript") is true in every single-language build, so the
union was satisfied without any language arm and none of the six was
load-bearing on any leg. The 22-leg sweep that found the other three
defects reported this line green while it was defective.

The arm exists for a real reason — the two LanguageDisabled tests need those
names with no language enabled — so it is not simply deleted. Those two
tests now import the names themselves and the shared union becomes the bare
language list, which makes every arm load-bearing on its own language's leg.

Measured rather than argued — perturb the c arm out, build --features c:

union shape	result
with not(feature = "javascript")	exit 0 — the leg is blind
bare language list (shipped)	error[E0433]

An earlier revision of this branch instead added a c,javascript CI leg. That
is superseded: it would have gated the c arm alone and left the other five
as blind as before.

Changes

tests/api/ast_seam_test.rs

• PathBuf and MetricsOptions unions gain feature = "c".
• ops_tree_contains_operator narrows to feature = "rust", its sole
  caller's gate.
• LANG / Source gains feature = "c" and drops its
  not(feature = "javascript") arm; the two LanguageDisabled tests import
  locally.
• count_string_agrees_with_find_on_tcl_family_bodies drops the cfg-gated row
  table for two #[cfg] blocks calling a shared check, matching the sibling
  test directly above it and correct at one row or two. The
  #[cfg(any(tcl, irules))] gate on the fn and the ran > 0 non-vacuity
  assertion both stay, per .claude/rules/testing.md.

.github/workflows/ci.yml, feature-matrix

• cargo check --all-targets → cargo clippy --all-targets -- -D warnings.
  single_element_loop is a clippy lint check never runs at all; the other
  two classes are already hard errors here because the workflow sets
  RUSTFLAGS: "-D warnings" top-level. The explicit -- -D warnings is
  insurance against that env var being narrowed later.
• New legs: c, go, python, tcl (lib) and go (ast).

On the leg choice: the issue suggested go alone, which would have caught
none of the defects — it enables none of the five features the shared
items in ast_seam_test.rs are gated on. It is kept (it is the subset the
testing rule names, and it does compile Go's own gated code under src/ —
thirteen feature = "go" gates, including a test), but the discriminating
legs are c, python and tcl.

Verification

• Full 22-leg sweep: 22 pass, 0 fail, against 17/5 before. Same script
  both times, and the before run was not uniform — three distinct error texts
  — so the harness discriminates rather than describing itself.
• Every corrected union confirmed load-bearing by perturbation, restoring
  clean each time via an in-memory backup, never a file-level git checkout
  (.claude/rules/testing.md).
• All ten pre-existing matrix legs verified passing under the stricter command
  before the promotion landed, so no unrelated failure rides in with it.
• The two LanguageDisabled tests run and pass under
  --no-default-features (zero languages — the case the local imports exist
  for) and under --features rust, where the shared and local imports are
  both in scope, with no shadowing warning.
• The Tcl-family tests run and pass under --features tcl,
  --features irules and --all-features, not merely compile.
• make pre-commit: BCA_GATE: pass (gate=pre-commit). make actionlint
  clean.

Scope

Compile failures only. Tests that compile under a partial feature set and
panic at run time are a separate class, tracked in #1285 / #1413 — running
cargo test --no-default-features --features c --test api now that it
compiles gives 24 such failures, all in suppression_test.rs (20) and
derive_eq_hash_ord.rs (4), which parse Rust fixtures with no rust gate.
ast_seam_test itself is clean on that leg. Untouched here.

No public API change, so no STABILITY.md cross-reference is required.