Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/snakemake/dag.py`:
- Around line 1357-1360: The current logic adds files to
_evicted_checkpoint_outputs just because out.exists(), which can let stale files
satisfy downstream inputs; change the guard so we only record outputs for
checkpoints that have actually completed previously—for example, require that
the output is present in the checkpoint's recorded created_output (or otherwise
verify checkpoint completion via Checkpoint.get()/a checkpoint.completed flag)
before adding to self._evicted_checkpoint_outputs; update the block around
job.is_checkpoint/job.output/out.exists() to check that authoritative completion
evidence (e.g. out in checkpoint.created_output) instead of plain filesystem
existence.

In `@tests/test_uncreatable_checkpoint_input/Snakefile`:
- Around line 1-2: The Snakefile currently mutates the checkpoint output by
opening and writing "a.txt" during parsing (the two-line block that writes
"a\nb\n"), which updates its mtime every Snakemake run; remove that file-writing
from the Snakefile and instead perform the setup in the test harness or an
explicit preparation step (e.g., test setup function or fixture) that
creates/updates "a.txt" before invoking Snakemake so checkpoint behavior is
modeled by external changes rather than by Snakefile evaluation.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/tests.py`:
- Around line 1499-1500: Update test_uncreatable_checkpoint_input to reproduce
the rerun regression: perform the first successful run via
run(dpath("test_uncreatable_checkpoint_input")) while preserving the workdir
(use the existing preserve_workdir helper or avoid cleanup), then update the
timestamp on the checkpoint input file with os.utime(<path_to_a.txt>, None), and
then invoke a second dry-run/run (e.g., run(..., dry_run=True) or run(...)
again) and assert that the downstream rule is NOT rescheduled by checking the
run output/return (e.g., assert "Input files updated by another job" not in
output or ensure no downstream job appears in the scheduling output). Ensure you
reference the test function name test_uncreatable_checkpoint_input, use run and
dpath to locate the test tree, and touch a.txt with os.utime.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/snakemake/dag.py`:
- Around line 1357-1363: The async method
self.workflow.persistence.incomplete(job) is being called without awaiting,
causing is_incomplete to be a coroutine and the condition to behave incorrectly;
update the checkpoint handling in the block that checks job.is_checkpoint (which
iterates over job.output and calls out.exists()) to await the incomplete call
(i.e., use await self.workflow.persistence.incomplete(job)) and ensure has_meta
is evaluated before awaiting incomplete if you want to avoid unnecessary awaits;
this will allow the subsequent if not is_incomplete: branch to run and correctly
add outputs to self._evicted_checkpoint_outputs.

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@src/snakemake/dag.py`:
- Around line 1357-1363: The code currently only calls
self.workflow.persistence.incomplete(job) when has_metadata(job) is true, which
allows checkpoint outputs without metadata to be treated as complete; instead
call await self.workflow.persistence.incomplete(job) unconditionally for each
out in job.output (i.e., remove the has_metadata gate and compute is_incomplete
= await self.workflow.persistence.incomplete(job)), and only add outputs to
self._evicted_checkpoint_outputs when that unconditional incomplete check is
false so incomplete-marker files are not promoted into
checkpoints.created_output (this touches job.is_checkpoint handling and uses
self.workflow.persistence.incomplete(job) and self._evicted_checkpoint_outputs).