[PIsberg]

Why

AtomicNonAtomicUpdate reported a get followed by a plain store to the same atomic even when the
stored value came from unrelated code. gauge.get() for a log line, then gauge.set(queue.size())
cannot lose an update, but the weaver sent every store to the hook that reports one, so correct code
was flagged (#1187).

What

• The agent reads each woven class's bytes from its loader and runs a per-method taint analysis
  (AtomicStoreTaintAnalysis). For each atomic store it records whether the stored value was derived
  from an atomic read in the same method.
• A store that was not derived is woven through a twin hook named <hook>Unrelated. The twin reports
  the store as an update with no read before it. Twins exist for set, lazySet, setPlain, setOpaque and
  setRelease on AtomicInteger, and for set, lazySet and setRelease on AtomicLong. Each twin keeps the
  original's point and spin-lock release.
• Stores are matched to verdicts by ordinal within their method, using one predicate shared by the
  analysis and the weaver. A store the analysis cannot reach gets the conservative verdict, so later
  stores keep their ordinals.
• A class that cannot be read, or cannot be analysed, keeps every store on the old hook.

Verification

• AtomicUnrelatedStoreWeavingEndToEndTest (CI=true) was red before this change with
  Expected no finding from detector 'AtomicNonAtomicUpdate', but got 1 (75 get+set sequences) and is
  green now. It now includes a lazySet case.
• AtomicDerivedStoreWeavingEndToEndTest, the positive control, is still green.
• AtomicStoreTaintAnalysisTest (7 tests, one built with ASM for a dead store after return),
  AtomicStoreTwinHooksTest, SubstitutionHooksArePointsTest and ControlledScheduleHookPointsTest
  pass.
• Fast tier: 3905 library tests and 106 agent tests, zero failures. ArchitectureTest passes.

Not verified: the lazySet end-to-end case has not been shown red without its twin, and the cost of
re-reading the class bytes has not been measured.

Docs

• docs/CHANGELOG.md: Unreleased, Fixed entry with the known limit.
• Javadoc on the changed code (AtomicStoreTaintAnalysis, FieldAccessWeaver.visitor, the weaver
  branch, the twin hooks, AsyncTestAgent.storeVerdicts).
• Not updated: the @AICore note on AsyncTestAgent still lists the store forms as they were. It is
  generated from the annotation, so it is left for the next vibetags regeneration.

Follow-ups

• [P2] ATOMIC_NON_ATOMIC_UPDATE still reports correct code when the read is passed into a method #1190 (P2): a value passed into a method is treated as derived, so correct code that routes its read
  through a helper keeps the finding.
• [P3] Atomic store ordinal matching has no guard against bytes changed by another transformer #1191 (P3): no guard against bytes changed by another transformer, which would shift the ordinals.
• [P3] Re-reading each woven class for atomic verdicts has no measured cost #1192 (P3): the cost of re-reading each woven class has not been measured.

Fixes #1187

🤖 Generated with Claude Code

CI fixes on this branch

• f702273c4 gives the Gradle agent build the ASM artifacts Maven already declares and relocates them in its shadow jar. Without it, :async-test-agent:compileJava failed in the Gradle Test Suite job. Verified locally: compileJava passes, and the shadow jar has 98 shaded ASM entries and none under org/objectweb.
• a685f3999 removes a PMD UnusedNullCheckInEquals violation in AtomicStoreTaintAnalysis.isAtomicStore, which failed the Maven verify gate in Corpus Eval.
• Corpus Eval failed once on Java 26 with agent_jctoolsHandOff_offererWritesAfterTheOffer silent, the known flake in [P3] Flaky: corpus agent-pairs row agent_jctoolsHandOff_offererWritesAfterTheOffer silent once #1133. Its re-run passed that gate. Other red jobs in this PR were GitHub infrastructure: a checkout that could not reach github.com, a Maven Central plugin that could not be fetched, and a runner that lost communication. Each was re-run and passed.

[codecov]

Codecov Report

❌ Patch coverage is 55.63910% with 59 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...versity/asynctest/telemetry/TelemetryRegistry.java	0.00%	42 Missing ⚠️
...sity/asynctest/agent/AtomicStoreTaintAnalysis.java	86.95%	6 Missing and 3 partials ⚠️
...a/se/deversity/asynctest/agent/AsyncTestAgent.java	45.45%	4 Missing and 2 partials ⚠️
...e/deversity/asynctest/agent/FieldAccessWeaver.java	81.81%	0 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!