fix: batch 2026-08-05 (#1204, #1209, #1212)

[dekobon]

Three low-priority follow-ups from #1197, #1200 and #1208, each closing a
gap where something was unguarded, undocumented, or dead. No metric value
moves and no public API changes, so there is no snapshot churn and no
big-code-analysis-output submodule bump.

What's here

Commit	Issue	Change
8ab58c3d	#1212	Remove Search::first_occurrence and its #[allow(dead_code)]
dd34b1c1	#1209	Document the K&R wrapped-return-type parse limitation, pinned by a drift-marker test
1c18af87	#1204	Add §7's fourth cross-walk pair and a _with_code subsection to grammar-dispatch.md, with a fixture sweep
303509f5	#1204	Guard that sweep against a bare root (found by /audit-tests)
cb4945b0	—	Consolidated CHANGELOG entry

#1212 — remove Search::first_occurrence

#1208 moved the four C-family get_func_space_name impls — its only
production callers — onto the shared declarator walk in
src/c_declarator.rs, leaving the method alive only under an
#[allow(dead_code)] and guarded by three tests that protected nothing
shipped. Search is pub(crate), so there is no public API impact and
STABILITY.md needs no change.

The one remaining utility use in src/spaces_tests.rs located a nested
function_definition, so first_child was not a drop-in (it scans one
level). It now uses root.preorder().find(…), which is the same
leftmost-first, root-inclusive walk — Preorder has its own order pin in
preorder_matches_recursive_document_order.

The child_scan_cursors consumer count drops six to five in both places
that state it. docs/development/benchmarking.md's historical #1112
measurement is re-attributed to act_on_node rather than restated with an
invented number — and that re-attribution was checked by reverting
act_on_node to children(), which reproduces exactly the recorded
"50 cursors over 50 nodes".

#1209 — K&R functions with a wrapped return type

A pre-ANSI K&R definition whose return type wraps the declarator opens no
function space under C or Objective-C:

int  krplain(a, b) int a; int b; { if (a) { return 0; } return 1; }  /* space, nargs 2 */
int *krptr(a, b)   int a; int b; { if (a) { return 0; } return 1; }  /* no space at all */

The report's root-cause section was partly wrong and has been corrected on
the issue. The parse is re-associated, not merely reclassified, and
produces no ERROR node: the declaration's wrapping declarator
swallows the first parameter declaration, int b; becomes a second
top-level declaration, and the body is orphaned as a bare
compound_statement. In tree-sitter-c 0.24.2 the old-style function
definition nests its declarator inside the old-style declarator, so an
outer pointer_declarator is unreachable.

The impact is not uniform invisibility — the orphaned body's decisions are
charged to the file's unit space (cyclomatic.sum 2 with
nom.functions 0), a file-level number no function-level row accounts for.

Two facts not in the original report: LANG::Objc shares the defect, and
Cpp / Mozcpp open no space for either K&R form, so .h / .mm
routing is not a mitigating factor.

Resolution is option 1 — document, do not work around. src/checker/c.rs
already enumerates both function_definition aliases correctly; the node
it would match is never produced, so there is no dispatch arm to add.
Re-implementing K&R declarator recognition would be a slice of the grammar
rewritten for a form C23 removed. The grammar pin is untouched, and
nothing was filed upstream.

The paired test is a drift marker: it pins cyclomatic.sum == 2 —
where the decision currently lands — so a grammar bump that fixes the
parse fails loudly rather than shifting metrics silently. Its doc comment
says explicitly that the values are a bug-lock, not an endorsement.
A control fixture (char **kr2, same defect, no decision) measures 1,
proving the pinned 2 tracks the orphaned if rather than being a
parse-failure constant.

#1204 — grammar-dispatch §7 and the _with_code spelling

§7 listed three predicate pairs and not is_func_space ↔
get_space_kind. That pair is worth its own entry because the three
existing ones disagree into a wrong count, which a snapshot diff
shows you, while this one disagrees into an absent key — a promoted
node the getter left Unknown is not a member scope, so its space
serializes no npm / npa block, which reads exactly like a language
with no containers.

The new subsection states why the _with_code trap is silent: the
defaults forward to the byte-less form, so both spellings compile and
behave identically everywhere except where a language overrides — which
is precisely where the answer matters. Elixir is that language, and this
has already landed once: ops_inner used the byte-less form and returned
a bare file-level space for every Elixir input (#1130).

Note for reviewers: the issue argued from metrics::opens_member_scope
(src/metrics/mod.rs:106), which #1203 deleted in ac437bb1 the day
before. The hazard survived — it now runs through
spaces::compute::open_func_space — so this fixes it on a corrected
premise, and the issue body has been updated. The issue's prescribed test
shape was also not buildable: Checker / Getter methods are static and
monomorphised per parser, and AstInner offers no LANG-generic way to
invoke them against a node, so building it meant growing production macro
surface to host a test. The sweep asserts the same implication one step
downstream, on the serialized space kind.

Verification

Every measured claim was checked by perturbation rather than assertion:

• Reverting open_func_space to the byte-less get_space_kind fails
  exactly 12 lib tests, all Elixir. containers_emit_npm_and_npa and
  function_spaces_emit_neither are among them, so the new sweep is
  not uniquely responsible for that case — its independent value is a
  partial-degradation selector hole, and the doc comment says so rather
  than overclaiming.
• /audit-tests found the new sweep had the vacuity hole its three
  siblings already guard against: it iterated every emitted space without
  checking any existed below the always-Unit root. Neutering the Elixir
  fixture to x = 1 made the unguarded sweep pass. Fixed in
  303509f5.
• Two independent /code-review passes returned no findings, each
  re-measuring the K&R claims against real parser output.

make pre-commit → BCA_GATE: pass (gate=pre-commit).

Coverage: the batch adds 135 lines of Rust, of which 123 are in test-only
files cargo llvm-cov does not place in the denominator and 12 are
comment lines in src/node.rs (98.79%, above the 96.62% project figure),
with zero overlap against that file's uncovered set. No measured
production line ships, so patch coverage cannot sit below project
coverage.

Fixes #1204
Fixes #1209
Fixes #1212