[rmcdaniel]

Part of #325.

Native replayed a live query lease to any later poll from the worker that held it. An independent second poll could therefore receive a task its worker was already running. The query-authority cohort recorded exactly that as a second claim at the deadline boundary (port log, known gaps).

Published PHP replays a live lease only to a poll that repeats the claiming poll_request_id. The installed Python SDK depends on this: it retries a timed-out query poll once with the same ID. Native now does the same:

• a claim records its poll_request_id, which is not delivered to the worker;
• a retry of that request receives the same task, attempt and lease again;
• a poll with a different ID, with no ID, or from another worker is an independent slot;
• a non-string ID is refused with 422, like other invalid poll requests.

Parity probe

A raw-protocol probe ran against published PHP Server 2.8.0 (the query-authority reference image) and against this build, each with its own SQLite database. Each column shows the task leased, or – for none:

step	PHP 2.8.0	this PR	main
A claims the first query	Q1	Q1	Q1
A again	Q1	Q1	Q1
B, nothing else pending	–	–	Q1
A after B	Q1	Q1	Q1
A after the query completed	–	–	–
two concurrent polls with one ID	both Q3	both Q3	both Q3

Three PHP behaviors are still missing, so native keeps advertising query_task_poll_request_idempotency: false:

• Cached empty results. PHP returns a remembered empty result to a retried ID even after a query arrives.
• Supersession. A newer ID ends an older waiting poll at once; native lets it wait out its timeout.
• Status names. PHP reports leased/empty; native reports task/timeout.

Tests

query_lease_replays_only_to_the_poll_request_that_claimed_it covers the same, a newer, a missing and a foreign-worker request ID, the invalid ID and the post-completion poll.

• With no replay at all, as in this PR's first commit, it fails at the same-ID replay.
• With main's same-worker replay, it fails at the newer-ID poll.
• The three existing tests that poll without an ID still expect no replay.

At 6663ab64a the full native suite passes on SQLite locally: 116 library tests (20 ignored) and 277 execution tests (1 ignored).

🤖 Generated with Claude Code

[rmcdaniel]

First qualification at 083646ed: shared run 39333801728 passed 12 of 13 jobs. The query-authority cohort passed on all five configurations it reached: MySQL 8.0, the existing MySQL matrix, MariaDB 10.11 and both PostgreSQL configurations. execution-and-codec failed earlier, at step 8 with exit 101, so its SQLite query-authority step did not run. The same failure appeared on main after #478, before this change. Every local, one-CPU and disposable-host repetition of that step passed. #501 separates dependency fetching so the cause is visible in step metadata. Next: merge #501, refresh this branch onto it, and requalify so the SQLite cohort runs.

[rmcdaniel]

Updated with main (including #509). The full SQLite execution suite passes locally on the merge (278 tests). Its own Shared Server fixtures run is cancelled to save runners: #510 is stacked on it and will be pushed with this head once #516 lands (to resolve that conflict once), so #510's run qualifies both.

[rmcdaniel]

Merged main (with #515's probe retries) as e0eaf8742. #510 stacks on this branch and now contains it, so its parity run 39453102751 qualifies both. I cancelled this branch's duplicate parity run 39453100933.

[rmcdaniel]

#510 merged into this branch (7909e2db9). Its parity run 39453102751 passed all 13 jobs on 5d2bf7e85, which contains #499.

d687d3f0c adds only main's merge (ef7283006): benchmark documents and a smoke script, no Rust or parity changes. I cancelled its duplicate Shared Server fixtures run 39466405568. This head still needs PHPUnit and Sonar before it can merge.

[sonarqubecloud]

Quality Gate passed

Issues
3 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
2.3% Duplication on New Code

See analysis details on SonarQube Cloud