[BUG] Shift-and LIKE kernel returns wrong results for patterns not anchored by %

[rishitesh]

With the default filter_use_like_shift_and = true, LIKE/ILIKE patterns that don't start and end with % return incorrect rows. Found while testing #2 / #3, but it reproduces with plain LIKE on current main.

Repro

Table with a single VARCHAR column containing:
Widget, WIDGET, widget, wIdGeT-Pro, Gadget, gizmo, WidgetX, midget

SELECT name FROM products WHERE name LIKE 'w_dget';

• Expected: widget
• Actual (shift-and path, default): widget, gizmo
• Actual with GQE_FILTER_USE_LIKE_SHIFT_AND=false: widget (correct)

Root cause (three related defects)

1. Strings shorter than the prefix/suffix silently match. In like_fn_shift_and_kernel (src/executor/like.cu), result is initialized to true and the prefix check is skipped when string_len < prefix_len — so the flag is never set to false for strings too short to match (this is how 5-char gizmo matches a 6-char pattern). The suffix check a few lines below has the same hole.

2. Patterns with no unescaped % lose exactness. preprocess_like decomposes 'w_dget' into prefix=w_dget, no suffix, no middles — which the kernel evaluates as 'w_dget%' (prefix-only, no length constraint), so e.g. WidgetX matches ILIKE 'w_dget'.

3. Overlapping prefix/suffix windows. For a pattern like 'ab%ba' on the 3-char string "aba", the suffix window overlaps the prefix, and the length checks are evaluated independently — the correct requirement is string_len >= prefix_len + suffix_len (the % must match zero or more characters between non-overlapping prefix and suffix). This also means string_len - prefix_len - suffix_len, passed to the middle-pattern matcher, can underflow as an unsigned subtraction.

Existing unit tests don't catch these because every LIKE test uses %-anchored patterns.

Proposed fix

Happy to submit a PR:

• Hoist a single length guard before the prefix/suffix checks: if (result && string_len < prefix_len + suffix_len) result = false; — fixes defects 1 and 3 and makes the middle-matcher length non-negative by construction.
• For defect 2, either pass a requires_exact_length flag to the kernel (pattern has no unescaped % ⇒ require string_len == prefix_len; preprocess_like already strips escape chars so prefix_len is the match length), or fall back to cudf::strings::like for such patterns, mirroring what like_utf8 already does for prefix/suffix patterns.
• Parameterized tests covering: no-% patterns with _, no-wildcard ILIKE, the 'ab%ba' overlap cases, strings shorter than prefix, and escaped variants ('w^_dget' escape '^').