[BUG] Batched application failures can use another task's retry budget

Describe the issue

FailTaskAppFailure can retry a task after its own application retry budget is exhausted when another task in the same failure batch has a larger budget. A task configured with retries=0 can therefore run again.

The query updates v1_task FROM tasks_to_steps without joining the target task to its matching row. The comparison tasks_to_steps.retries > v1_task.app_retry_count can use another task's budget. The missing join also allows a task excluded from locked_tasks because it has no active runtime to be updated.

Environment

• Observed with self-hosted engine v0.106.3 and Python SDK 1.35.1; the persisted Step had retries=0, while its task reached app_retry_count=1 and internal_retry_count=0.
• Deterministically reproduced through the repository query on upstream main at 929f9d2c5d0fc65640efce885c33bde76809baa9, using PostgreSQL 15.6 and Go 1.26.5.
• The same query is present in v0.107.4. This is separate from the SDK default-resolution issue #4727.

Expected behavior

Each active attempt should use its own Step's budget; non-retryable failures and attempts without a runtime should not be retried.

Reproduction

Create tasks with the following stored state, then submit all five to one FailTaskAppFailure call, using their current IDs, insertion timestamps and retry counts:

Task	Step retries	App retries used	Internal retries used	Active runtime	Non-retryable	Expected retry	Actual retry
A	0	0	0	yes	no	no	yes
B	1	1	0	yes	no	no	yes
C	2	1	2	yes	no	yes	yes
D	3	0	0	yes	yes	no	no
E	3	0	0	no	no	no	yes

Proposed fix

Carry inserted_at through tasks_to_steps and join it to the update target on (id, inserted_at). A repository regression test fails before this change and passes afterward; the existing queued-task replay test also passes.

I would like to contribute the fix and regression test. Could a maintainer triage this and assign it to toliv if accepted?

AI Disclosure

• An LLM was used in preparing this issue, in accordance with Hatchet's AI_POLICY.md.
• Details: OpenAI Codex investigated the query, searched for existing reports, drafted this issue, and prepared and ran the fix and regression test.