fix(checks): descendants JSONPath returns scalar vs list depending on match count

[Qalipso]

Fixes #2633.

Problem

_is_list_expression handles Descendants in the same recursive branch as Child:

if isinstance(expression, Child | Descendants):
    return _is_list_expression(expression.right) or _is_list_expression(expression.left)

For trace..ctx both children are single-field Fields, so this returns False and resolve() falls back to its match-count heuristic (len(matches) > 1 or _is_list_expression(expression)). A .. expression then yields a bare scalar when the data holds exactly one match and a list when it holds two or more.

This is the defect #2605 fixed for dot-wildcards and multi-field selectors. That fix added a Fields branch; .. compiles to Descendants and was not covered. .. is a supported expression — test_extraction.py already asserts trace..outputs validates.

It breaks ComparisonCheck's match=any/all/none, which requires a list/set/tuple:

Equals(key="trace..ctx", expected_value="doc-1", match="any")

one interaction carries ctx : FAIL | Expected a list, set, or tuple at key 'trace..ctx'
                                     when match is 'any', but got str.
two interactions carry ctx  : PASS | At least one value in ['doc-1', 'doc-2'] is equal to 'doc-1'.

Identical check configuration; the outcome depends on incidental data shape.

Fix

Give Descendants its own branch returning True, and leave Child recursing. A .. searches a whole subtree, so its match count is a property of the data, not of the expression — the same reasoning the bracket-wildcard and Fields branches already encode.

Behavior change worth flagging

trace..missing now returns [] instead of NoMatch. That matches how list-expressions already behave — trace.last.metadata.* on empty metadata returns [] today — while ordinary paths such as trace.last.nope keep returning NoMatch. Covered by a test.

Verification

• uv run pytest libs/giskard-checks -m "not functional" → 761 passed, 4 skipped (757 passed, 4 skipped on main before this change; the 4 added tests are the difference).
• TDD red→green: reverting only extraction.py reproduces 'doc-1' == ['doc-1'] (bare scalar) and fails 2 of the new tests; reapplying passes all 5 in the class.
• ruff check / ruff format --check on both changed files: clean.
• basedpyright on extraction.py: 0 errors, 28 warnings — byte-identical to the count on main, so nothing new was introduced.

Python 3.13, jsonpath-ng from the pinned range.

Separate observation, not addressed here

Trace.model_dump() includes a last key duplicating the final interaction, so trace..outputs on a single-interaction trace returns ['hello', 'hello']. Different problem, different blast radius — noted in #2633 and left alone.

[davidberenstein1957]

Suggested change

	# `..` searches a whole subtree, so how many values it finds is a property
	# of the data, not of the expression. Treat it as a list-expression the way
	# bracket wildcards are, or resolve() falls back to its match-count
	# heuristic and returns a bare scalar whenever the data happens to hold one
	# match.

[davidberenstein1957]

Opened a same-repo follow-up with the same fix plus format fixes so CI lint can pass (fork PRs cannot receive the auto-format push): this supersedes the branch here.

Thanks for the clear repro and tests — credited as co-author on the fix commit.

[davidberenstein1957]

Opened a formatted follow-up so CI check-format can pass: #2645

Same change as this PR, plus the markdown/ruff format fix. Please use #2645 going forward — we can close this one once that is green.

[Qalipso]

Thanks — happy to have #2645 carry it, and no objection to closing this once that lands.

One heads-up: #2645 currently shows as conflicting with main (mergeStateStatus: DIRTY), same for #2652. Keeping this branch open until yours is green, as you suggested; say the word and I'll close it, or rebase this one instead if that turns out to be easier.

[kevinmessiaen]

Duplicate of #2644