[colbymchenry]

Summary

In Ruby, a call on an instance variable, like @writer.flush, now links to the method of the class that the class's initialize gives the variable: @writer = CacheWriter.new(…) makes it a CacheWriter. Before, no call on an instance variable linked at all. On main every such call stays unresolved, for example 2,780 of them in activesupport and 3,239 in actionpack.

T5's triage of #684 found this gap; there is no issue or contributor PR for it. It builds on #3049's receiver code.

Root cause

Both extractors record such a call as @writer.flush. No strategy reads an instance variable's type. matchMethodCall's receiver pattern doesn't even admit the @, so the call ended as a failed reference.

Fix

ruby-constant-paths.ts resolves @name.m the same way #3049 resolves a call on what .new made:

• Where it applies: the call must be in an instance method (not def self.x, not class << self) of a class. A module's methods run on whatever includes it, so they are left out.
• Reading initialize: every assignment to @name in that class's initialize (= or ||=) must start a statement and be a lone Const.new, A::B.new(…) or ::Foo.new …. Arguments in or out of parentheses, a block or a statement modifier may follow. All of them must name one constant.
• When it gives no type: anything else leaves the variable untyped, and the call goes to the other strategies, as before. That covers a conditional, two classes, .tap, writer || Foo.new, @a, @b = …, or a chain continued on the next line.
• The class and method: Ruby's constant lookup from initialize gives the class. The method is its own instance method, else the first ancestor's (rubyAncestry). A class the project doesn't declare (@logger = Logger.new($stdout)) links nothing.
• Only initialize is read, as the triage asked. A reassignment in another method doesn't untype the variable (see the triage).

There is no extraction change, so the kernel is unchanged.

Validation

Tests: __tests__/ruby-ivar-receivers.test.ts runs in both engines. It covers:

• typed calls, including ||= with arguments over several lines, a namespaced class built with a block, and an inherited method;
• untyped cases: a conditional, two classes, .tap, a multiple assignment, and a variable that initialize never assigns;
• a class from outside the project;
• a class method's and a module's variables.

With main's resolver the positive test fails in both engines, and the 6 untyped and outside-class tests pass. With the fix all 8 pass.

A/B: kernel engine, main 3a4cd3bf vs this change, natural-key edge diff. The Rails repos are subtrees of rails 793d744. The branch was then rebased onto 25144678, whose JVM changes (java.rs, kotlin.rs, JVM-gated resolution) don't touch Ruby. The suite ran on the rebased commit.

Repo	Edges added	Removed	Right
sinatra cb22afd	0	0	–
jekyll 97c3bdc	8	0	8
activesupport	19	0	19
actionpack	14	0	14
activerecord	30	0	27

Every edge was checked against its class's initialize. Some examples:

• jekyll's @liquid_renderer.stats_table → LiquidRenderer;
• activesupport's @registry.add → Fanout::Registry, and @lock.* → Concurrency::ShareLock;
• actionpack's @scanner.* → Journey::Scanner, and @router.serve → Journey::Router;
• activerecord's @available.poll → ConnectionPool::Queue (inherited by ConnectionLeasingQueue), and @state.finalized? → TransactionState.

The 3 that are wrong in practice are AbstractAdapter's @pool.* calls. initialize sets @pool = NullPool.new, so they go to NullPool, but pool= later replaces it with the real ConnectionPool.

The controls, requests (Python) and cobra (Go), are byte-identical by dump-graph.

Index time: median of 3 interleaved runs: activesupport 4.72 s vs 4.79 s, activerecord 11.54 s vs 11.54 s.

Full suite: on this commit (e186fd6a, on main 19b74d10), after npm ci, kernel built, --maxWorkers=4. 801 files (795 passed, 6 skipped) and 9,606 tests: 9,520 passed, 0 failed, 86 skipped.

Left out

• A variable typed only through a setter or a lazy getter (def writer; @writer ||= Foo.new; end) stays untyped, since only initialize is read. So does one set in a superclass's initialize.
• A variable that another method reassigns keeps its initialize type. Making any assignment outside initialize untype it would fix the NullPool case, but it would also drop right edges such as activesupport's @registry = @registry.add(…), which is still a Registry.

🤖 Generated with Claude Code