Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@monai/metrics/conformal_risk.py`:
- Around line 59-65: The function `_set_from_threshold` in the file is dead code
that is never called, not exported, and has no tests associated with it. Remove
the entire function definition including the docstring to clean up the codebase
and reduce maintenance burden.

---

Nitpick comments:
In `@monai/metrics/conformal_risk.py`:
- Line 253: Add the `strict=True` parameter to the zip() call that iterates over
self._scores and self._labels in the for loop. This will ensure that both
iterables have matching lengths and raise a ValueError if they don't, helping
catch logic errors early.
- Line 50: The optional_import call for tqdm on line 50 imports both tqdm and
has_tqdm variables, but neither of these variables is used anywhere else in the
conformal_risk.py file. Remove the entire line containing the tqdm
optional_import statement since it introduces dead code that serves no purpose.
- Around line 277-279: The reset() method is missing a docstring as required by
coding guidelines. Add a docstring to the reset() method that describes its
purpose, which is to clear the internal state by resetting the scores list,
labels list, and num_classes attribute back to their initial values. Follow the
project's docstring format conventions.
- Line 76: The labels_flat.clamp operation silently adjusts out-of-range labels
without any logging, which can hide upstream data issues. Before applying the
clamp operation on labels_flat, detect if any values fall outside the valid
range [0, c-1] by checking the minimum and maximum values. If out-of-range
values are detected, log a warning message that indicates clamping is occurring
and optionally includes details about how many or what proportion of labels were
affected. Then proceed with the clamp operation as currently written.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@monai/metrics/conformal_risk.py`:
- Around line 75-76: Instead of silently clamping invalid labels to valid class
indices using the clamp operation, add validation to reject or raise an error
when labels contain invalid values outside the expected range of 0 to c-1. This
needs to be fixed in two locations: the labels_flat.clamp call around line 76
and the similar operation around line 226. Replace the clamping logic with a
check that validates all labels fall within the valid range and either raises an
exception or skips invalid samples, ensuring that corrupted labels do not
silently propagate through the loss, coverage, and lambda_hat calculations.
- Around line 259-262: The code materializes the entire lambda grid evaluation
at once in the line with `sets = scores_i.unsqueeze(0) <= lam_grid.view(-1, 1,
1)`, which creates a tensor of shape (n_lam, P_i, C) that can cause
out-of-memory errors for large 3D volumes. Refactor this block to chunk the
lambda grid into smaller batches and iterate through them in a loop, processing
each chunk separately and accumulating the risk contribution from each chunk
into risk_sum. This prevents materializing the full tensor at once while still
computing the correct cumulative risk.
- Around line 41-48: The `__all__` list in the conformal_risk.py file has items
that are not in alphabetical order, which violates Ruff's sorting requirements.
Specifically, swap the positions of "compute_set_size" and "compute_coverage" in
the `__all__` list so that "compute_coverage" appears before "compute_set_size",
matching alphabetical order.
- Around line 193-197: The validation of lam_grid is incomplete and allows two
problematic cases: empty grids and unsorted grids. In the validation condition
that checks if lam_grid is a 1-D tensor with values in [0, 1], add two
additional checks: verify that lam_grid is not empty (check the length or size),
and verify that lam_grid is sorted in ascending order (you can use
torch.is_nonincreasing or check if the differences between consecutive elements
are non-negative). Update the error message in the ValueError to reflect all
validation requirements. This will prevent crashes at line 271 and ensure
correct behavior at line 273 when finding the infimum.
- Line 262: After calling self.loss_fn with sets_shaped and labels_rep on the
line where risk_sum is updated, add validation to ensure the output is valid.
Validate that the returned loss tensor has the expected shape matching
labels_rep, contains no NaN values, and all values are within the valid range of
0 to 1. If any validation fails, raise a descriptive error to prevent silent
failures in the CRC bound calculation. This validation should occur immediately
after the loss_fn call and before accumulating the result into risk_sum.
- Around line 319-323: The set_threshold method currently accepts tensors of
arbitrary shape, but it should enforce that the lam parameter is a scalar tensor
with a value in the range [0, 1]. After the existing isinstance check for
torch.Tensor, add validation to ensure the tensor has a scalar shape (empty
dimensions) and that its value falls within [0, 1], raising a ValueError if
either condition is violated. This prevents invalid broadcasts over spatial
dimensions that could produce incorrect conformal sets.
- Line 253: In the for loop iterating over self._scores and self._labels using
zip(), add the parameter strict=True to the zip() call to enforce that both
sequences have the same length. This will catch any synchronization issues
between these two buffers if they become out of sync in the future due to
unequal append operations.

---

Nitpick comments:
In `@monai/metrics/conformal_risk.py`:
- Around line 277-279: The reset method in the ConformalRisk class is missing a
Google-style docstring. Add a docstring to the reset method that describes its
purpose of resetting internal state variables (_scores, _labels, and
_num_classes to their initial values). Include a brief one-line summary followed
by a more detailed description if needed. Since the method takes no arguments
and returns None, focus on describing what state is being reset and why this
operation is performed.

In `@tests/metrics/test_conformal_risk.py`:
- Line 187: The variable `probs_out` is unpacked from the predictor call but
remains unused, which triggers the RUF059 linting rule. Replace the unused
variable with an assertion that validates the returned `probs_out` meets the
expected contract, such as checking its shape, type, or value ranges that align
with the test's expectations. This both eliminates the unused variable warning
and improves test coverage by verifying the predictor's output contract.