fix(core-engine): args counts declared parameters, not call sites (objective-c, typescript)

[squid-protocol]

Closes #2773.

args sat at 73% consistency across the control corpus. The documented low tail is the
no-parameter-surface family; the high tail had never been examined, and two languages turned
out to be counting the arguments at call sites alongside the parameters at declarations — so
their args total scaled with how many calls a file happens to make.

The contract is stated in the new docs/args_rule_contract.md, in the shape #2730 established for
api:

args matches the parameters a callable declares.

How the numbers were measured

Every regex match is classified against a tree-sitter parse of the same code stream the engine
scans (prism.split_streams(...)["code_stream"] — prism blanks comments in place, so byte
offsets stay aligned). A match whose parenthesised span resolves to formal_parameters /
parameter_list / method_definition is a declaration; one that resolves to arguments /
argument_list / message_expression is a call. That is the rule-17 hand-checkable oracle, and
unlike #2753's PyYAML oracle or #2752's CSS walker it generalises to any language in
tree_sitter_language_pack.

objective-c — 378 → 101 on the crucible, 21 → 13 on the control shells

Only 101 of 378 hits were declarations (26.7% precision).

• The colon-selector arm matched every Objective-C message send. [store setVersion:V],
  [list objectAt:i] — a send is lexically identical to the untyped label:name parameter shape
  objective-c args extraction: untyped keyword-message params undercount to 0, and body content can leak into the count #1335 added. 120 of the 277 false hits, plus 3 ternaries (cond ? MIN_HEIGHT : maxY) and
  both @interface Anchor:Object superclass colons.
  Now requires a real -/+ method-declaration head (with func_start's own optional macro /
  __attribute__ prefix and an optional return type) — the one context a keyword-message
  signature can appear in and a send never can, since a send always leads with [.
  This also closes the string-literal hole the issue asked for a negative test on: prism strips
  comments but not strings (prism/detector: string-literal shielding is selective by signal family (if/try/open count inside a literal, eval does not) #2535), so @"status: ok" really was an args hit.
• The plain C arm accepted any name(...) followed by {/;. free(conn);, abort();,
  printf("...", x); each scored an argument — 146 of the false hits. Now demands the same
  structural proof c/cpp demand (the list must open with a type token) plus a parameter name
  after that token, without which StrAllocCopy(Address, tag); still reads as typed via the
  PascalCase typedef fallback all three rules share.

After: 100 of 101 hits are declarations — the 101st is a method definition tree-sitter itself
mis-parses (see Claim 15 below). The only real declaration dropped is page_width(), an empty
parameter list, which c/cpp already treat the same way.

typescript — 16844 → 11025 on the crucible, 15 → 13 on the control shells

• The class-member arm had no declaration anchor at all. ^[ \t]*IDENT(...) is equally the
  shape of a bare call statement: 6133 call sites against 3026 declarations, and
  describe(kit); / expect(kit); in the control shell were two of them. javascript — the same
  rule one file over — already anchors this with a (?=[ \t\n]*\{) body lookahead; typescript,
  which also has return-type annotations, needs { or a : return type. Measured: keeps 95.8%
  of declarations, drops 98.5% of calls.
  A bodyless constructor overload (constructor(runner: () => void, timeout: number);) has
  neither by grammar, and all 125 declarations the anchor would otherwise lose are constructors —
  so constructor is named explicitly rather than readmitting a bare ;, which would bring back
  4906 call statements.
• The arrow arm's gap between ) and => was [^=;{]* — unbounded and newline-crossing, so
  any parenthesised expression matched if a => turned up later (384 casts, if conditions and
  grouped operands). Only a return-type annotation may legally sit there, so the gap is now
  whitespace or a bounded :-led annotation carrying one level of balanced parens (a curried
  arrow's return type is itself a function type — ): ((...a: A) => Either<E, B>) =>).

Precision 61.3% → 99.0%, and recall rises (10326 → 10910 real declarations): the
over-reaching arrow arm used to swallow real parameter lists inside an over-long match.

The audit — all 46 languages

docs/args_rule_contract.md carries the contract, three corollaries, the fallback-family table for
languages with no formal parameter list, and a per-language table with both corpora. Four
violations found beyond the two fixed here, each filed with its own evidence:

#2782	groovy +46%	the return-type run is {0,3}, so the method arm degenerates to ^IDENT(...)
#2783	apex +38%	the return type is optional, same degeneration
#2784	m4 +38%	counts $1 references; shell counts the highest position (_args_findall_max_groups)
#2785	haskell	an arrow-less :: signature (region :: IORef Int) is a value, not a callable

css counting calc()/var()/url() arguments is a call site too, but it is settled and ledgered
(args-no-parameter-surface-morphology) — documented as the one knowingly-approximate fallback,
the same way matlab is in docs/api_rule_contract.md, not re-litigated here.

The compliant rules cluster into four anchors, and the doc names them so the next rule does not
have to rediscover them: a declaration keyword (19 languages), a mandatory return type
(java/csharp), a terminator lookahead (javascript/dart, now typescript), or a typed parameter list
(c/cpp, now objc's C arm).

Also in this PR

• gitgalaxy/standards/how_to_add_a_language.md — the one-line contract next to args in the
  output schema, matching what The api rule has no stated contract: 7 languages' own visibility idiom is invisible to it, and 5 others count any public token #2730 did for api.
• docs/why_gitgalaxy_beats_ast_here.md Claim 15 — tree-sitter-objc cannot parse
  - keyDown:(NXEvent*)theEvent followed by #ifdef TRY1
  (language-crucible/data/objective-c/worldwideweb/HyperText.m:1426) and loses the region to
  error recovery -- no method node covers that line at all. GitGalaxy
  reads it correctly. Logged per CLAUDE.md's standing instruction, with its own "where this does
  NOT apply" section — the same parse was the oracle that found GitGalaxy's own 277 false
  positives.

Verification

• Golden masters blessed: 25 differences, every one of them Function Parameters, on
  typescript and objective-c files only (plus javascript/react/BabelPlugin.ts, a .ts file).
  Zero off-target languages, zero topological drift.
• rosetta_audit.py — 46 languages, 2 moved: objective-c and typescript, both onto the
  planted 13 exactly. The other 44 are byte-identical. That is the mechanical check that the
  anchors are right and not merely narrower.
• tree_sitter_accuracy_audit.py --ci --all — 30 languages, no regressions. typescript's
  args_exact_match holds at 2881/2915 and objective-c's at 153/153. An earlier spelling of the
  arrow-arm gap that excluded ( outright cost 6 curried fp-ts functions here; that is what the
  balanced-paren allowance and its regression test exist for.
• tri_comparison_chart.py --all --ci — 3 languages, all OK.
• Scaling sweep over both rules: linear at n=2000…16000 on every adversarial shape (macro-prefix
  runs, selector-shaped text with no lead, unterminated parameter lists, return-type gaps that
  never reach an arrow).

Corpus pairing (keyword-rosetta, after this merges)

No authoring change is needed — the fix lands both shells on 13 exactly. What is owed is the
re-bless of data/objective-c and data/typescript expected_signals.json, and re-scoping
deviation_ledger.json's ts-callparen-args: retire its objective-c and typescript arms, keep
apex and groovy and re-point them at #2783 / #2782.

🤖 Generated with Claude Code

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab