[vatsrahul1001]

Follow-up to the deserialization guard backported in #75745 (manual backport of #75263).

Problem

#75745 backported the crash guard from #75263 but dropped the starved_dag_runs dimension, because the full #75263 was entangled with #70307 ("Fill scheduler task batches after concurrency rejection"), which is not on v3-4-test.

Without starved_dag_runs, the guard skips a task instance whose Dag fails to deserialize but never removes that dag run from the candidate set. When such a run sorts to the top of a batch limited by max_tis:

1. _build_schedulable_tis_query fetches only the highest-priority rows (the bad run's TI).
2. _task_concurrency_allows_execution skips it (deserialization fails → return False), but nothing is starved.
3. found_new_filters is False → the selection loop breaks.
4. Lower-priority runnable TIs behind the bad run are never queued that tick — and the same thing repeats every tick.

So the backport guarded against the crash but reintroduced the starvation that #75263 exists to prevent. This surfaced as a branch-wide failure of test_find_executable_task_instances_skips_dag_that_cannot_be_deserialized (shipped by #75745), which fails the CI passed gate on every open PR into v3-4-test.

Fix

Port only the self-contained starved_dag_runs slice from main's #75263 onto the selection machinery already present on v3-4-test:

• mark a dag run as starved when its Dag fails to deserialize (and short-circuit an already-starved run);
• exclude starved runs from the re-query in _build_schedulable_tis_query;
• count a newly starved run in found_new_filters so the loop re-queries and makes forward progress.

This makes the backport behave identically to main without pulling in #70307. The found_new_filters refill loop and the query-exclusion builder it relies on are already on v3-4-test.

Testing

The previously-failing test now passes, along with the sibling DB-error-propagation test and the full test_find_executable_task_instances* / query-builder / critical-section suite (40 passed) locally on SQLite:

test_find_executable_task_instances_skips_dag_that_cannot_be_deserialized PASSED
test_find_executable_task_instances_reraises_db_errors_from_dag_load PASSED

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 4.8)

Generated-by: Claude Code (Opus 4.8) following the guidelines

[vatsrahul1001]

Closing in favour of 75802