Two kinds of Ruby call linked nothing. A call to a method whose name ends
in `!` or `?` on a receiver (`writer.flush!`, `writer.ready?` after
`writer = CacheWriter.new`, `CacheWriter.create!(1)`) failed the method-call
pattern, whose method part took no `!` or `?`. A method called straight on
a new object (`CacheWriter.new(key).write_cache`, `Billing::Ledger.new.charge!`)
was recorded by both extractors with the receiver's raw text, arguments and
blocks included, which no strategy matched.
Both extractors now record a call on `Const.new(...)`, `A::B.new` or
`::Foo.new` as `new Const().m`, the constructed-receiver shape TS/JS, Java
and C# already use; the wasm extractor and the kernel's ruby.rs emit the
same refs. ruby-constant-paths resolves that shape before the name
pre-filter: the class by Ruby's constant lookup from the namespaces around
the call, the method as the class's own instance method or the first its
ancestry declares, and nothing when either is missing, so rack-protection's
`Request.new(env)` (rack's Rack::Request) links nothing. The Ruby dot pattern
in matchMethodCall takes a trailing `!` or `?`; the receiver-word scoring
among several project methods of the name stays off for such names on a
value receiver, where it was wrong on 55 of 74 new edges (`options.key?`
on a Hash -> InheritableOptions#key?). rubyAncestry no longer reads a
module a call returns (`include RouteSet.new.tap {...}.url_helpers`) as a
mixin of the constant the call starts at.
Validation, kernel engine, main 37052b18 vs this change:
- sinatra cb22afd +12, jekyll 97c3bdc +53, activesupport +398,
actionpack +97 (2 removed: `ActionController::Base.with` had reached
RoutesHelpers.with through a misread mixin), activerecord +1012
(rails 793d744 subtrees). About 21 of 1,572 added edges are wrong, all
from the class-name and single-method guesses that plain names get too.
- Controls requests (Python) and cobra (Go): dump-graph byte-identical.
- kernel-parity.mjs: fixtures 10/10, sinatra 147/147, jekyll + rails
1044/1044 files byte-parity, none deferred.
- New tests: 10 fail with main's kernel and TS, all pass with the fix, in
both engines. Dropping the scoring gate or the mixin fix fails its guard
test.
- Index time, median of 3 interleaved: activesupport 5.35 vs 4.63 s,
jekyll 2.66 vs 2.73 s (noise).
Reimplements PR #684 by @ricobl, who found both gaps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
