fix: warn on implicit zero-fill and preserve empty dicts in Batch

[Lidang-Jiang]

Summary

Addresses #1088 and #1089:

• Batch: don't just strip off empty entries when creating batches #1089 (empty dict silently dropped): Empty dicts in batch lists were silently dropped during stack_, causing index position misalignment. Now all entries are preserved via _validate_and_convert_batches().
• Batch: don't just set 0 when elements have None entries #1088 (None → 0 without warning): When __setitem__ or stack_ fills missing numeric keys with 0, a UserWarning is now emitted via _warn_numeric_zero_fill() to alert users that the zero value may mask truly missing data.

The fix is backward compatible — the zero-fill behavior itself is unchanged (changing it would break existing code), but users are now explicitly warned when it happens.

Design decisions

• Extracted _validate_and_convert_batches() and _warn_numeric_zero_fill() as module-level helpers to keep stack_() complexity under ruff's C901 limit (max 20).
• Warning is UserWarning (not DeprecationWarning) since the zero-fill behavior is not being deprecated, only made visible.

Before (bug reproduction)

--- Bug #1089: Empty dict silently dropped ---
Input: [{"a": 1}, {}]
Result: Batch(
    info: Batch(
              a: array([1]),   # length 1 — index [1] lost
          ),
)

--- Bug #1088: None replaced by 0 silently ---
Batch(a=[1, 2, 3], b=[{"c": 1}, {}, {"c": 3}])
# b.c = array([1, 0, 3]) — 0 silently inserted at index [1], no warning

After (with fix)

--- Bug #1089: Empty dict now preserved ---
Input: [{"a": 1}, {}]
Result: Batch(
    info: Batch(
              a: array([1, 0]),   # length 2 — both indices preserved
          ),
)
Length: 2

--- Bug #1088: Warning on 0 fill ---
UserWarning: Key 'c' is not present in all batches during stacking
(missing at indices [1]). Filling missing entries with 0 for numeric
type (ndarray), which may mask truly missing values. Consider using
None or np.nan to represent missing data explicitly.

Test results (75 passed)

test/base/test_batch.py ......................................... [ 82%]
test/base/test_batch.py::TestBatchNoneAndEmptyHandling .......... [100%]
======================== 75 passed, 14 warnings in 2.31s ========================

Test plan

• 13 new tests in TestBatchNoneAndEmptyHandling covering empty dict preservation and warning behavior
• All 75 existing batch tests pass
• ruff check and black pass

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ddce36a42e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Reconcile nested schemas before keeping empty stack entries

Keeping top-level empty dicts as Batch() placeholders causes mixed nested schemas to fail in cases that previously worked. For example, Batch.stack([{"info": {"a": 1}}, {}, {"info": {"b": 2}}]) now routes info through the partial-key assignment path; the first value allocates only subkey a, and assigning the third value then raises ValueError("Creating keys is not supported by item assignment.") when subkey b appears. This breaks stacking whenever an empty entry sits between batches whose nested keys differ.

Useful? React with 👍 / 👎.

[Lidang-Jiang]

Good catch. When empty dicts are preserved between batches whose nested keys differ, the element-by-element assignment path fails because __setitem__ doesn't support creating new keys.

Fixed in 0953587: for partial keys containing nested Batch/dict values, we now collect all entries and delegate to recursive Batch.stack instead of assigning one-by-one. This handles differing nested schemas correctly.

Verified locally:

>>> Batch.stack([{'info': {'a': 1}}, {}, {'info': {'b': 2}}])
Batch(info: Batch(a: array([1, 0, 0]), b: array([0, 0, 2])))