Add stdDevPop, stdDevSamp window operator and setWindowFields stage tests

[Cassie-2025]

Add compatibility tests for:

• stdDevPop window operator (11 test files, 146 tests)
• stdDevSamp window operator (11 test files, 146 tests)
• setWindowFields stage error validation (5 test files, 33 tests)

Coverage includes: documents/range/time-range frame modes, partitioning, sort order, numeric precision, non-numeric handling, NaN/Infinity behavior, field paths, pipeline stage interactions, and frame validation errors.

All 325 tests verified passing.

[documentdb-triage-tool]

🤖 Auto-triaged by documentdb-triage-tool.

Applied: compatibility test, enhancement
Project fields suggested: Component test-coverage · Priority P2 · Effort XL · Status Needs Review
Confidence: 0.92 (mixed)

Reasoning

component from path globs (test-coverage); effort from diff stats (12099+0 LOC, 24 files); LLM: Adds 325 new compatibility tests across multiple files for stdDevPop/stdDevSamp window operators and setWindowFields stage, spanning documents/range/time-range frame modes and various edge cases.

If a label is wrong, remove it manually and ping @patty-chow so the rules can be tuned. The bot will not re-label items that already have component labels.

[drk-knght]

lgtm

[eerxuan]

Summary

This is the first substantive window-operator PR — 325 tests across 24 files for $stdDevPop, $stdDevSamp, and $setWindowFields stage errors, where previously every window// folder and setWindowFields/ held only smoke tests. Because it sets the template for ~26 remaining window operators, the structural decisions matter more than the individual assertions. The stdDev
math itself is correct (population vs sample formulas, single-element pop=0 vs samp=null, deterministic sorts — no cross-contamination), but the PR bakes in significant duplication, omits the reviewer-requested TEST_COVERAGE.md spec, and deviates from the repo's near-universal test conventions.

Findings

1. 🔴 Window-frame / partition / sort mechanics are tested per-operator and duplicated across both operators — they belong once at the stage level.
   documents_mode_frames, range_mode_frames, time_range_mode_frames, partitioning, sort, field_paths, and pipeline_stages (≈7 of 10 files per operator) test $setWindowFields stage behavior — how documents/range/time bounds select the frame, how partitionBy/sortBy work, pipeline composition. These are spec-consistent across all ~30 window operators, and the files are
   byte-parallel between stdDevPop and stdDevSamp (identical test-name sets: whole_partition, cumulative, sliding_centered, trailing, …; verified 0 error assertions — value-only). Per TEST_COVERAGE.md §18/§19 ("test once, representative operator") and REVIEW.md §8 ("duplicate coverage"), this matrix should live once under stages/setWindowFields/ driven by one
   representative operator. Left as-is, it becomes the precedent that every future window operator re-copies the full ~48-case frame matrix ×N. Keep per-operator only the genuine stdDev delta (numeric_precision, non_numeric_handling, special_floats + 1–2 wiring cases).

2. 🟠 docs/testing/TEST_COVERAGE.md not updated, though this introduces the window-operator feature category.
   REVIEW.md §9 (line 85): "TEST_COVERAGE.md updated if the PR introduces coverage rules for a new feature category (e.g. date operators, window functions)" — window functions is the literal example, and the PR touches the doc 0 times. As the first window-operator PR, it must add a "Window Operator Coverage" section stating what's tested once at the $setWindowFields stage
   level (frame modes, partitionBy, sortBy, frame-validation errors — representative operator) vs per-operator (the operator's own math: formula, empty/single-element result, non-numeric handling, special floats, precision). Without this written rule, finding #1's duplication becomes the standing precedent.

3. 🟠 Test-format deviation — 322 hand-written def test_* functions instead of the parametrized-dataclass pattern.
   TEST_FORMAT.md "Parametrized Tests" and REVIEW.md §4 mandate @DataClass(frozen=True) + pytest_params() for operators with many cases. This PR uses 0 pytest_params/parametrize and 322 standalone def test_* with a copy-pasted local _run_stdDev* helper — while 1603 existing files under tests/core/operator/ use pytest_params and none use this style. This is the lone
   deviation, and it's the template the next 26 window operators would copy. (Converting to parametrized dataclasses also collapses the 800-line files, so I'm not flagging length separately.)

4. 🟠 Shared helper + corpus duplicated per file instead of a window/utils/ module.
   Every file re-defines its own _run_stdDevPop/_run_stdDevSamp pipeline-builder and BASIC_DOCS/numeric corpus (byte-identical across the two operators bar the operator name). REVIEW.md §5/§6 and TEST_FORMAT.md require shared helpers/data in utils/; sibling accumulators/utils (AccumulatorTestCase), stages/lookup/utils, etc. establish the convention. A window-pipeline
   change would need editing ~19 times.

5. 🟡 Raw integer error codes in the stage-error files.
   REVIEW.md §4: "Error codes from framework.error_codes used (not raw integers)." The three setWindowFields_*_errors files make 29 assertFailureCode(result, ) calls (code 9 twenty times, despite FAILED_TO_PARSE_ERROR = 9 existing) with 0 error_codes imports — so REVIEW.md §8 ("test the right error code") can't be verified at a glance.

6. 🟡 Inline float("nan")/("inf")/("-inf") literals instead of the existing constants.
   TEST_COVERAGE.md:613 "All test constants … defined in framework/test_constants.py. Import from there — do not duplicate values." FLOAT_NAN/FLOAT_INFINITY/FLOAT_NEGATIVE_INFINITY already exist; the special_floats/precision files hardcode them ~80×. (Note: the files do correctly use the established assertSuccessNaN idiom for NaN comparison — that part is compliant.)

7. 🟡 No # Property [Name]: comments in any of the 24 files.
   REVIEW.md §4 / TEST_FORMAT.md require per-group # Property [...] comments; the diff has 0. As the category template, this drops the property-documentation convention wholesale.