Problem

ConditionalParameter.filter uses a plain substring check, which causes silent mismatches in hybrid-attention models (e.g. FLA / linear-attention architectures).

A filter of "attn" substring-matches both self_attn and linear_attn weights, causing unintended blending of SSM state tensors (A_log, dt_bias, etc.) that must remain intact during merging.

Solution

Add glob metacharacter detection (*, ?, [) in _filter_matches(). When the filter string contains any of these characters, use fnmatch.fnmatch instead of the substring check. Plain strings retain the original substring behaviour for backward compatibility.

# Before: "attn" matches BOTH self_attn AND linear_attn
filter: "attn"   # unintentionally blends SSM parameters

# After: precise control with glob
filter: "*self_attn*"   # matches self_attn only, skips linear_attn

Changes

• mergekit/config.py: add fnmatch import, _GLOB_CHARS frozenset, and _filter_matches() helper; update evaluate_setting() to use it
• tests/test_config_filter.py: 13 new tests covering plain substring (backward compat), glob patterns, and the linear_attn collision case

Tests

pytest tests/test_config_filter.py -v
# 13 passed

Motivation

Discovered while merging FLA-based models (e.g. Jackrong-27B) with standard-attention bases using DARE-TIES. The substring collision corrupted linear-attention SSM parameters, silently degrading model quality. This fix allows users to opt into precise control by writing "*self_attn*" while preserving existing configs that use plain substrings.