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

Inline comments:
In `@pyspectrafuse/commands/build_cluster_db.py`:
- Around line 272-275: The groupby().idxmin() call can fail when a cluster's
posterior_error_probability is all NaN; before computing best_idx, coerce
posterior_error_probability to numeric on the enriched DataFrame and replace
NaNs with np.inf (e.g., set enriched['_pep'] =
pd.to_numeric(enriched['posterior_error_probability'],
errors='coerce').fillna(np.inf) or call enriched['_pep'].fillna(np.inf,
inplace=True)) so that enriched.groupby('cluster_id')['_pep'].idxmin() returns
valid positions; then compute best_idx and select best_rows as before (symbols:
enriched, '_pep', posterior_error_probability, idxmin, best_idx, best_rows).

In `@pyspectrafuse/commands/quantmsio2mgf.py`:
- Around line 46-53: The worker pool is being created with
Pool(processes=os.cpu_count()) which ignores the --task_parallel option; change
the Pool creation to use the task_parallel value (i.e.
Pool(processes=task_parallel)) so that Parquet2Mgf.convert_to_mgf_task actually
runs with the requested number of worker processes, and keep the existing
chunksize parameter for batching (or compute a separate chunksize variable) when
calling pool.imap to control task batching rather than concurrency.

In `@pyspectrafuse/commands/spectrum2msp.py`:
- Line 198: The code currently calls ParquetPathHandler(...).get_item_info()
with the first parquet file path (path_parquet_lst[0]), which extracts the
accession from the last path segment and therefore yields the parquet filename
instead of the project basename; change the call to pass the project directory
(e.g., the parent directory of the parquet files or the existing project_dir
variable) into ParquetPathHandler so get_item_info() reads the directory name as
the accession; update the reference at the call site (where basename is set) to
construct the handler with the folder containing the parquet files rather than
the file path itself.

In `@pyspectrafuse/common/parquet_utils.py`:
- Around line 58-68: The loop building parquet_path_lst currently appends any
glob hit (parquet_file) without checking whether it's a file, so partitioned
datasets stored as directories ending in .parquet get returned; update the loop
in this helper to skip directories by calling parquet_file.is_file() (or
equivalently ensure pathlib.Path(parquet_file).is_file()) before applying the
existing suffix/name/parts filters and appending to parquet_path_lst, leaving
all other filtering (checks against _metadata_suffixes, _exclude_names,
_exclude_dirs) unchanged.

In `@pyspectrafuse/incremental/resolve_clusters.py`:
- Around line 99-103: The current filtering uses rep_mgf_basename and
str.contains, which yields false positives; instead compute exact basenames
using pathlib.Path and do an equality match: derive rep_mgf_basename from
Path(rep_mgf_path).name, compare against cluster_df['mgf_path'].map(lambda p:
Path(p).name) (or add a temporary column of basenames) to build rep_mask, then
set rep_rows/new_rows accordingly; ensure pathlib.Path is imported if not
already.

In `@pyspectrafuse/mgf_convert/parquet2mgf.py`:
- Around line 155-163: _get_mgf_path currently always pulls instrument from
sample_info_dict, ignoring the skip_instrument option; update _get_mgf_path to
check the skip_instrument flag (the variable/attribute used by Parquet2Mgf for
that option) and, when True, use a single shared instrument bucket (e.g., a
fixed name like "all_instruments" or an empty instrument segment) instead of
info[0]; modify the function where mgf_group_df['mgf_file_path'] is set so
_get_mgf_path reads skip_instrument and constructs the path accordingly
(reference function name _get_mgf_path, the sample_info_dict lookup, and
Parquet2Mgf.skip_instrument or the local skip_instrument variable).

In `@tests/test_incremental.py`:
- Around line 150-155: The fixture string assignments for the variable
tsv_content use unnecessary f-string prefixes (e.g., f"reps.mgf\t0\tNEW_A\n"
blocks) which trigger Ruff F541; remove the leading f from each quoted literal
(including the second occurrence around lines 193-196) so they become plain
string literals while preserving the content and escape sequences.

---

Nitpick comments:
In `@pyspectrafuse/commands/parquet2dat.py`:
- Around line 35-39: The parquet discovery helper find_target_ext_files is being
imported from pyspectrafuse.commands.spectrum2msp causing a cross-command
dependency; move or reuse the generic helper from
pyspectrafuse.common.parquet_utils instead: remove the import of
find_target_ext_files from pyspectrafuse.commands.spectrum2msp in parquet2dat.py
and import the equivalent function (or add a new helper) from
pyspectrafuse.common.parquet_utils, then call that helper to produce
parquet_files and delete any duplicated discovery logic so both spectrum2msp and
parquet2dat rely on the single common helper (update exports in parquet_utils if
needed).

In `@pyspectrafuse/commands/spectrum2msp.py`:
- Around line 40-53: glob.glob() can return files in nondeterministic order;
make the discovery deterministic by sorting the found paths. Change the loop
over glob.glob(...) to iterate over sorted(glob.glob(...)) or sort the collected
list `res` (e.g., call `res.sort()` before returning) so that the order of
`path_str` items and the resulting `res` list is stable for downstream
`file_idx` increments and output naming; update code around the for-loop that
builds `res` (referencing `path_str`, `file`, and `res`) to ensure the sorted
behavior.

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

Inline comments:
In `@Dockerfile`:
- Around line 39-44: The Docker build fails because pyproject.toml declares
readme = "README.md" but README.md is not copied before running RUN uv pip
install --system --no-cache . Update the Dockerfile so that README.md is copied
into the image before the install step (i.e., place a COPY README.md . between
the COPY pyproject.toml . and COPY pyspectrafuse/ ./ lines) so the package
installation can read the declared README during build.
- Line 34: Replace the non-reproducible image reference used as a build stage
source in the COPY statement (currently "ghcr.io/astral-sh/uv:latest" in the
COPY --from=... line) with an explicit version tag and its SHA256 digest (e.g.,
ghcr.io/astral-sh/uv:<version>@sha256:<digest>); obtain the correct tag and
digest using docker manifest inspect (or the uv Docker integration docs) and
update the COPY --from=... source to that pinned reference to ensure
reproducible, verifiable builds.

In `@pyspectrafuse/maracluster_dat.py`:
- Around line 147-149: The code defines a columns list including 'rt' but still
writes Spectrum records with retentionTime=0.0; update the Spectrum/dict
construction to use the parsed rt value instead of the hard-coded 0.0 (e.g.,
retentionTime=rt or retentionTime=row['rt']), and make the same change at the
other spot flagged (the second Spectrum/dict creation). Ensure the 'rt' value is
cast to float if necessary before assigning to retentionTime so the parquet
persists the actual retention time.
- Around line 196-218: The code converts mz_array and intensity_array to numpy
arrays then calls _bin_spectrum but doesn't verify they have the same length;
when they differ _bin_spectrum may index out-of-bounds. Before computing
charge/prec_mz/etc., add a validation step comparing lengths of mz_arr and
int_arr (and that both lengths > 0) and skip the spectrum (increment skipped) if
they differ or are zero-length; reference the variables mz_arr, int_arr and the
called function _bin_spectrum when adding this guard so only well-formed arrays
reach _bin_spectrum.
- Around line 264-279: The converter is currently grabbing any ".parquet" files
(via find_target_ext_files) and produces .dat outputs for non-spectrum files;
restrict it to spectrum inputs by changing the finder/filter to only match
".psm.parquet" (e.g., call find_target_ext_files(parquet_dir, '.psm.parquet') or
post-filter parquet_files to keep only names ending with '.psm.parquet'), and
keep the FileNotFoundError when the filtered list is empty; reference the
existing parquet_files variable and the find_target_ext_files call so the filter
change is localized and prevents producing .dat for cluster metadata or
membership files.

In `@README.md`:
- Line 162: The fenced code block at the README's project tree is missing a
language identifier (triggering markdownlint MD040); update the opening fence
from ``` to ```text (or another appropriate language like ```bash) so the block
becomes labeled (e.g., change the block around the pyspectrafuse tree to start
with ```text) and leave the closing ``` as-is to satisfy the linter and improve
rendering.