[rmcdaniel]

Cuts the Shared Server fixtures' remaining long pole, and bounds every readiness probe as Cloud's #756 review asked.

Why. After #756 the corpus part sets the wall time alone: 18 to 20 min per database, while every other part takes 16 min or less (#756 comment 6394082981). Almost all of it is one step, 1031 s on mysql-8.0. That step:

• records and compares the corpus: about 8 min;
• runs the published patch, repeated-patch, legacy-marker and range-repair profiles: about 4 min;
• runs the native restart and PHP takeover-refusal cases: about 5 min.

The profiles and restart cases use their own workflows and only the nodes the step starts.

What changes

• A sixth part, restarts, in the MySQL, PostgreSQL and SQLite jobs (134d822).
  • A shared step, "Start the published PHP and native nodes the corpus parts share", runs in both the corpus and restarts parts.
  • The corpus part records and compares the corpus.
  • The restarts part runs the profiles and restart cases.
  • Every check still runs exactly once per database.
  • The SQLite job scopes its resources per step, so its RESOURCE_SCOPE env now follows each of the three steps. The Docker isolation policy caught this missing in a first draft; without it the later steps would have started unscoped containers.
• Bounded readiness probes (41f4655). Every /api/ready probe now passes --connect-timeout 2 --max-time 5, so one accepted but silent connection can't hold a curl until the 60-minute job limit (Cloud's review of 058861c).

Tests

• ServerParityPartsTest.mjs now requires:
  • six parts;
  • the shared node start in exactly the corpus and restarts parts;
  • the profiles and restart cases only in restarts, and none of them in the corpus comparison;
  • a bounded request on every readiness probe.
• The new cases fail against the previous workflow (3 of 4 parts tests fail).
• Locally: parts test 5/5, actionlint, the Docker isolation policy, and the public boundary check.

Expected effect. The corpus part drops to about 11 to 12 min and the restarts part takes about 12 min, so the slowest part becomes sdk-directions or native-tests at about 16 min, down from about 20. I'll post the measured per-part times from this PR's run.

🤖 Generated with Claude Code

[rmcdaniel]

Cloud's bounded read-only wrapper review accepts exact head bc95d005514e8278c66942abd60914c89ff3df5e within scope: .github/workflows/server-parity.yml resource startup, step guards, readiness, evidence and cleanup. No native/shared/frozen implementation, stopped inspector evidence or fixture execution was inspected or run.

All122 docker run command bodies match the accepted#756 workflow exactly after normalizing the added --connect-timeout 2 --max-time 5 flags. All21 readiness probes now have those request bounds. The shared node-start step is enabled for both corpus and restarts, each part keeps its own runner/DB/network, SQLite retains the explicit step-level resource scope, and the existing always-run cleanup/artifact naming covers the new part. Corpus result assertions remain explicit after comparison; the new profile step still fails on command failure.

No additional wrapper finding or Cloud hold. Native behavior, full CI success and measured throughput improvement remain the owner's separate qualification gates. Any later workflow/head change needs its own scoped recheck.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[rmcdaniel]

Merged as 7ede756, pinned to the reviewed head a0ba83a95716174c8a9ae2e97cb64e7ba672d5fc. Main's tree 3f003815 is that head's tree.

Checks at the head. Required PHPUnit Feature, Public Boundary, SonarQube and Shared Server fixtures all passed. The fixtures run (40169120328) needed one rerun of two failed jobs. Both failures were reference-side PHP repair-pass races, recorded on #754 (comment 6394576240), and both jobs passed on attempt 2.

Measured on attempt 1 (job start to finish):

Part	Before #757	After
corpus (all backends)	~20 min, the slowest part	6–12 min
restarts (new part)	inside corpus	11–13 min

The slowest part is now native-tests on the existing PHP MySQL matrix, at 18 min, followed by sdk-directions at 14–16 min. The whole attempt took 24 min of wall time from first job start to last job end, queueing included.

The next CI change goes after native-tests. Further fixture parts stay on the one-activation-at-a-time queue.