[adarshsm]

Fixes #10296.

Problem

A parameter named size or nfiles is invisible to the run-cache, so changing it does not
re-run the stage and the outputs from the previous run are checked out instead. Reproduced
on current main:

$ cat params.yaml
size: 30
count: 30

$ dvc stage add -n gen -p size,count -d gen.py -o out.txt "python gen.py"
$ dvc repro
$ cat out.txt
size=30 count=30

# changing an ordinary param re-runs the stage
$ sed -i '' 's/count: 30/count: 40/' params.yaml && dvc repro
Running stage 'gen': ...

# changing 'size' does not
$ sed -i '' 's/size: 30/size: 40/' params.yaml && dvc repro
Stage 'gen' is cached - skipping run, checking out outputs

$ cat out.txt
size=30 count=40     # stale: the new value never took effect

The failure is silent, which is what makes it nasty — the stage reports as cached and the
outputs look valid.

Cause

_get_cache_hash() excludes size and nfiles because they are file metadata rather than
part of the stage definition, but it passes the exclusion to dict_sha256() for the whole
lockfile, and dict_filter() is recursive:

return dict_sha256(cache, exclude=[Meta.PARAM_SIZE, Meta.PARAM_NFILES])

The lockfile has four top-level sections — cmd, deps, params, outs — and metadata
only ever appears under deps and outs. params holds arbitrary user keys, so a
parameter literally named size or nfiles was stripped along with the metadata and never
reached the hash.

This matches @skshetry's reading on the issue: "We are recursively excluding nfiles and
size before hashing for stage cache, which is incorrect. Most likely, we'll be able to
remove size and nfiles only from outputs that are not parameter dependencies."

Fix

Apply the exclusion to the deps and outs sections only.

Existing run-cache entries stay valid. For a stage without a parameter of either name,
the filtered dictionary is identical to what the old code produced, so the hash is
unchanged. Only stages that were mis-hashed get a new hash, which is the point.

Verified directly:

case	result
ordinary stage, old hash vs new hash	identical → run-cache preserved
size: 30 vs size: 40, old code	identical → the bug
size: 30 vs size: 40, new code	different → fixed
size under deps/outs changed	identical → metadata still excluded

And end to end, on the reproduction above: size changes now re-run the stage and
out.txt tracks the parameter, while an unchanged pipeline still reports
Data and pipelines are up to date.

Tests

Two unit tests in tests/unit/stage/test_cache.py:

• test_cache_hash_tracks_params_named_like_meta — the regression. Fails on main
  (the two hashes are equal), passes here.
• test_cache_hash_still_excludes_deps_and_outs_meta — guards the behavior that must not
  change; passes both before and after.

pytest tests/unit/stage/                                                # 84 passed
pytest tests/func/test_run_cache.py tests/func/test_stage.py \
       tests/func/params/ tests/func/repro/                             # 150 passed

tests/unit/stage/ also reports two failures in my environment
(test_fill_from_lock_use_appropriate_checksum, test_dump_nondefault_hash), both
ImportError: s3 is supported, but requires 'dvc-s3' to be installed. They are unrelated:
the failing test ids are identical with and without this change.

ruff check, ruff format --check and mypy dvc/stage/cache.py are clean.

Note on scope

@TimCosemans reported the same symptom on the issue for parameters named time_dummies and
method. Those names are not excluded anywhere in the hashing path, so I believe that is a
different problem and have deliberately left it alone rather than widen this change; happy
to look at it separately if it is still reproducible.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.98%. Comparing base (2431ec6) to head (21facb2).
⚠️ Report is 220 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #11079      +/-   ##
==========================================
+ Coverage   90.68%   90.98%   +0.30%     
==========================================
  Files         504      505       +1     
  Lines       39795    41177    +1382     
  Branches     3141     3263     +122     
==========================================
+ Hits        36087    37466    +1379     
- Misses       3042     3072      +30     
+ Partials      666      639      -27     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[skshetry]

Thanks for contributing! LGTM.