[kinyoklion]

Update, 2026-11-13

• Rebased onto feat/overrides, which now includes feat: Add the overrideAffected indicator to EvaluationReason #218 and feat: Count override-affected evaluations separately and send no individual events for them #219 and has main merged into it. From main it gets the 7.18.0 release, CI on the latest v3 contract tests, flag change listener support in the contract test service, and the java-sdk-internal 1.12.1 and 1.12.2 fixes. A common 2.6.0 and internal 1.13.0 release from this branch needs those fixes.
• No changes of its own.
• Verified at the top of the stack, with common and internal published locally from feat/overrides:
  • every PR in the stack compiles, with Javadoc, at its own head;
  • the server unit tests pass (2285, with the 72 existing skips);
  • the full v3 contract tests pass (6321 tests). That includes the flag overrides suite, with the deleted-entry tests of feat: Test that an override file with a deleted entry fails the reload sdk-test-harness#599, and both flag listener suites.

Requirements

• I have added test coverage for new or changed functionality
• I have followed the repository's pull request submission guidelines
• I have validated my changes against all supported platform versions

Related issues

This resolves Finding 1 of the Windows stability assessment for the Java overrides work: plans/java-overrides-windows-test-stability.md in sdk-scratchpad.

It is stacked on #259, the top of the overrides stack, so the PRs below it do not change.

Describe the solution you've provided

#220 added FileSynchronizerReloadBehaviorTest. It pins how the legacy file data synchronizer reloads: on every file system event, reporting every failure, without retrying. On Windows, two of its tests fail every time; on Linux they pass:

• everyFailedReloadIsReportedAndLoggedAsThePlainDescription
• failedReloadIsNotRetriedWithoutAFileChange

Why they fail on Windows:

• Both tests rewrite the file with Files.write, which truncates the file and then writes it.
• On Windows that produces two WatchService keys, and the legacy synchronizer reloads after each.
• The reload after the truncation reads an empty file. The legacy loader picks the parser by content, so the empty file goes to the YAML parser, which reads it as an empty document. The synchronizer then delivers a change set with no flags.
• The reload after the write then reports the malformed content.
• The tests asserted that the first result after the rewrite is that STATUS, so they received the change set instead.

A new helper, awaitStatus, consumes results until the first STATUS arrives, within the existing timeout. The three assertions that waited on the next result now use it. On Linux, the first result is still the STATUS. The pins still describe the legacy behavior they set out to pin: the failure is reported, each failure is logged as the plain description, and nothing retries without a file change.

To reproduce the Windows sequence on Linux, I made a scratch copy of the test that empties the file, waits 300 ms, and then writes the malformed content. Both pinned tests fail with the old assertions and pass with this change. The whole class passes on Linux as written. The file data, file synchronizer and file override test classes pass (144 tests), and checkstyleTest is clean. I did not run this on Windows. java-core CI runs on Linux only.

Describe alternatives you've considered

Writing the malformed content to a temporary file and moving it into place, which is one event on both platforms. That would pin a different writer than the one the tests describe.

Additional context

Not in this PR:

• Finding 2. That reload after the truncation also shows a defect on main, outside the overrides work. On Windows, every in-place rewrite of a legacy file data source's file briefly applies an empty data set: all flags are gone and defaults are served until the next reload. The assessment recommends a minimal fix: treat an empty file as a failed load, which keeps the last data. That changes the legacy data sources, so it belongs in its own PR against main.
• Finding 3. Build hygiene, all on main:
  • the compile tasks set no source encoding, so JDK 17 on Windows reports the box-drawing characters in three source files as unmappable;
  • the server pins common 2.6.0 and internal 1.13.0, which are not released yet.

_{Opened from termag}

---

Note

Overview
Adds an awaitStatus helper in FileSynchronizerReloadBehaviorTest that drains synchronizer futures until the first STATUS result, within the existing timeout.

everyFailedReloadIsReportedAndLoggedAsThePlainDescription and failedReloadIsNotRetriedWithoutAFileChange now use awaitStatus instead of assuming the next result after a malformed rewrite is immediately a failure status. On Windows, in-place writes can emit an extra reload (empty file → change set) before the malformed-content STATUS, which was breaking those pins; Linux behavior is unchanged when STATUS is already first.

^{Reviewed by Cursor Bugbot for commit d63a9a0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Stale comment

Risk: low.

Automated risk review (dry run): Low

This PR would be automatically approved.

Why: The diff only updates FileSynchronizerReloadBehaviorTest.java so two reload-failure pins wait for the first STATUS after a Windows double WatchService event, with no production, infra, or CI workflow changes.

If you disagree with this review, share this PR in #feedback-agent-approvals or submit a changes-requested review explaining why.

  

_{Sent by Cursor Approval Agent: Dry-Run PR Approver}

[cursor]

Automated risk review (dry run): Low

This PR would be automatically approved.

Why: The only change is in FileSynchronizerReloadBehaviorTest.java, a test helper that waits for the first STATUS result so Windows dual WatchService events do not fail existing pins. No production, infrastructure, auth, or CI files are modified.

  

_{Sent by Cursor Approval Agent: Dry-Run PR Approver}