Java impact: type initializers, library-run members, field initializer after a comment

[swapnilpaliwal-sd]

Java impact --tests misses measured against a behavioural oracle on 12 Java (Maven, JUnit 4/5) repositories, 50 sampled methods each, every one broken on entry and the suite run (truth = test files that newly fail vs a per-test baseline). Dependencies and the JDK are staged as libraries. 8 repositories were used for tuning, 4 held out and scored once at the end.

Fixes

1. A Java type's initializer is run by whoever uses the type (impact.dl load_hop, rung at load). A call written in a field initializer has no enclosing method, so the engine names the TYPE as its caller; a static / instance block is a <clinit> / <init_block> callable. Nothing calls either, so the upward walk stopped there: a private constructor called only from INSTANCE = new Factory() selected 0 of 17 failing test files. The hop now goes from such an initializer (a type reached as a caller, never a type/field target's own seed) to whoever uses the type: a call to one of its members, a construction of a subtype, a read or write of one of its fields. The SQL fast path declines such a closure.
2. The at-load hop sees a test that reads a static field. KIND_SKIPS hands a method query empty field / faccess facts, so the field clause derived nothing (silently). Both are now loaded for method / param / var; every other rule joining them also needs a relation that stays skipped for those kinds.
3. Parser: a field initializer written after a comment is read. static final F X = // why + newline + new F() { ... }; — the field extractor took the node after = as the initializer, and a comment is a named node, so the whole initializer (and an anonymous class it creates, with its methods) was dropped. Try-with-resources had the same read. Both now skip comments, as local variables already did.
4. The members a library runs on a Java object reach whoever builds it (protocol rung). equals / hashCode / toString, and an override of a library method (iterator of an Iterable, size of a collection), are run by a HashMap, a StringBuilder, a for-each loop — never a written call on a client receiver, and the engine does not fan Object members on purpose. The protocol hop Python already has (whoever constructs the type) gains Java clauses: constructors of the type and its subtypes. Seed-only, as for Python.

Each fix has a case under tests/cases/java/ that fails on the base and passes here, with a control that must not change: type-initializer-runs-for-its-users (+ a type no test touches), field-initializer-after-a-comment (+ the same shape with no comment), library-runs-the-members-it-is-given (+ a type no test constructs). The parser comment-invariance twins gain the three initializer shapes (31 vs 36 rows before, equal after).

Numbers (base = integration tip)

set	recall	macro recall	precision	safe	empty selections
tune (8)	0.930 → 0.982	0.909 → 0.936	0.181 → 0.183	291 → 302 / 340	16 → 10
held out (4)	0.946 → 0.992	0.948 → 0.973	0.105 → 0.105	172 → 182 / 187	5 → 5

An answer that errored counts as an empty selection (fix 3's subject had one mutant whose answer errored on the base: 71 failing files).
No subject loses recall. Per subject on tune, recall gains of +0.25, +0.15, +0.12 come from fixes 1–2, 3–4 and 1 respectively.
path (mutation pairs) 0.849 → 0.853, runtime direct callers 0.754 → 0.754 (585 targets), runtime A→C paths 399/442 unchanged, context recall@5 0.568 unchanged.
Rung precision on tune: at load 0.46 (n=67), protocol 0.10 (n=898, the weakest rung, kept below one of a set).

Cost, warm, interleaved A B A B on two subjects: index 24.5/30.3 s → 28.8/25.7 s and 43.4/41.1 s → 42.1/42.2 s (noise); 20 impact queries 16.4/17.4 s → 22.0/19.4 s and 74.6/71.5 s → 78.9/76.1 s (+6 to +25%, the field facts now loaded for a method query).

Suites

tests/run.py: java 343/343, typescript 269/269, javascript 307/307, python 306/306. Java engine suite with --oracle (JDK 25): 79 passed, 0 failed, on base and on the parser change; torture SKIPs on this machine (its golden is pinned to another platform IR).

Still open (fewer than two repositories, or precision only)

• framework-run members of a web application (controller constructors, binder initialisers, template getters) — one repository
• a callback held by a library object (a cache loader, a class-value computation) — one repository
• assertThat(x.primitiveGetter()).isEqualTo(..): a call argument returning a primitive keeps every overload, so the chained assertion stays unresolved (three repositories; feeds by name rows)
• an implicit super() to a base no-arg constructor is not an edge
• path does not take the at-load hop