feat: split mixed-pattern groups before CRX inference

Add --split-mixed flag that detects groups with diverse first symbols
and splits them before CRX. This separates distinct calling patterns
that CRX would otherwise merge into flat bags.

Results:
- FastAPI: 22→33 grammars, avg score 0.47→0.57
- RAGSAK: 5→8 grammars, avg score 0.45→0.71
- dependency_testing: bag→clean 'return.q?.commons?.skip?.limit?' (1.00)

212 tests pass

bex/tag_preprocessor/analyze.py

@@ -283,7 +283,33 @@ def analyze_clusters(file_paths, extension, project_root="", min_coverage=DEFAUL
     return [("(all methods)", result, len(sequences), meta)]
 
 
-def _infer_group(label, group_seqs, group_files, project_root, min_coverage, prefer, kmax, N, include_kore=False, include_idregex=False, method='langsize', min_methods=3, crx_method='standard', min_structure=0.0):
+def _split_by_first_symbol(symbol_seqs, min_subgroup=3):
+    """Split sequences by first symbol to separate mixed patterns.
+
+    Returns:
+        dict mapping first_symbol → list of sequences,
+        or None if all sequences share the same first symbol (no split needed).
+    """
+    if not symbol_seqs:
+        return None
+
+    first_symbols = {}
+    for seq in symbol_seqs:
+        if seq:
+            first_symbols.setdefault(seq[0], []).append(seq)
+        else:
+            first_symbols.setdefault('(empty)', []).append(seq)
+
+    # Only split if 2+ sub-groups have enough methods
+    viable = {k: v for k, v in first_symbols.items() if len(v) >= min_subgroup}
+
+    if len(viable) <= 1:
+        return None  # no useful split
+
+    return viable
+
+
+def _infer_group(label, group_seqs, group_files, project_root, min_coverage, prefer, kmax, N, include_kore=False, include_idregex=False, method='langsize', min_methods=3, crx_method='standard', min_structure=0.0, split_mixed=False):
     """Infer grammar for one package group. Module-level for ProcessPoolExecutor."""
     filtered = frequency_filter(group_seqs, min_coverage=min_coverage)
     imports = _extract_imports(group_files)
@@ -303,6 +329,34 @@ def _infer_group(label, group_seqs, group_files, project_root, min_coverage, pre
         meta = {"files": group_files, "imports": imports, "arg_patterns": arg_patterns, "packages": packages, "skip_reason": "too_diverse"}
         return (label, None, len(filtered), meta)
 
+    # Split mixed-pattern groups before CRX
+    if split_mixed:
+        splits = _split_by_first_symbol(symbol_seqs, min_subgroup=min_methods)
+        if splits is not None:
+            # Infer each sub-group, pick the best
+            best_result = None
+            best_score = -1
+            best_label = label
+            total_count = 0
+            for first_sym, sub_seqs in splits.items():
+                sub_label = f"{label} [{first_sym}]"
+                sub_result = infer_ensemble(sub_seqs, kmax=kmax, N=N, prefer=prefer, min_coverage=min_coverage, include_kore=include_kore, include_idregex=include_idregex, method=method)
+                if sub_result and sub_result.get('best') and sub_result['best'].get('grammar'):
+                    g = sub_result['best']['grammar']
+                    ok, _ = validate_sore(g)
+                    if ok:
+                        score = grammar_structure_score(g) if min_structure > 0 else sub_result['best']['mdl_score']
+                        if score > best_score:
+                            best_score = score
+                            best_result = sub_result
+                            best_label = sub_label
+                total_count += len(sub_seqs)
+
+            if best_result is not None:
+                meta = {"files": group_files, "imports": imports, "arg_patterns": arg_patterns, "packages": packages, "split": True, "n_splits": len(splits)}
+                return (best_label, best_result, total_count, meta)
+            # fall through to unsplit inference
+
     if crx_method == 'refined':
         from ..crx_refined import crx_with_confidence
         info = crx_with_confidence(symbol_seqs)
@@ -329,7 +383,7 @@ def _infer_group(label, group_seqs, group_files, project_root, min_coverage, pre
     return (label, result, len(filtered), meta)
 
 
-def analyze_by_package(file_paths, extension, project_root="", min_coverage=DEFAULT_COVERAGE, prefer=None, kmax=2, N=3, min_pkg_size=3, include_kore=False, include_idregex=False, method='langsize', min_methods=3, crx_method='standard', min_structure=0.0):
+def analyze_by_package(file_paths, extension, project_root="", min_coverage=DEFAULT_COVERAGE, prefer=None, kmax=2, N=3, min_pkg_size=3, include_kore=False, include_idregex=False, method='langsize', min_methods=3, crx_method='standard', min_structure=0.0, split_mixed=False):
     """Preprocess and group by package directory, infer per group.
 
     Groups methods by their file's relative directory path, merging
@@ -364,7 +418,7 @@ def analyze_by_package(file_paths, extension, project_root="", min_coverage=DEFA
             gs = [sequences[i] for i in indices]
             gf = set(seq_files[i] for i in indices)
             f = ex.submit(_infer_group, label, gs, gf, project_root,
-                           min_coverage, prefer, kmax, N, include_kore, include_idregex, method, min_methods, crx_method, min_structure)
+                           min_coverage, prefer, kmax, N, include_kore, include_idregex, method, min_methods, crx_method, min_structure, split_mixed)
             futures[f] = label
 
         for f in as_completed(futures):
@@ -594,6 +648,7 @@ def analyze_directory(
     min_structure=0.0,
     context_strategy="dir",
     reduce_threshold=0.15,
+    split_mixed=False,
 ):
     """Scan a directory and run analysis for each language found.
 
@@ -633,6 +688,7 @@ def analyze_directory(
                 min_methods=min_methods,
                 crx_method=crx_method,
                 min_structure=min_structure,
+                split_mixed=split_mixed,
             )
         elif slice == "reduce":
             results[ext] = analyze_by_reduce(
@@ -844,6 +900,10 @@ def _parse_args(argv=None):
         "--min-structure", type=float, default=0.0,
         help="Minimum structure score (0.0-1.0) to keep grammar. Flat bags of symbols below this are dropped (default: 0, keep all)",
     )
+    parser.add_argument(
+        "--split-mixed", action="store_true",
+        help="Split groups with mixed first symbols before CRX inference (produces tighter grammars)",
+    )
     return parser.parse_args(argv)
 
 
@@ -870,6 +930,7 @@ def main():
         min_structure=args.min_structure,
         context_strategy=args.context_strategy,
         reduce_threshold=args.reduce_threshold,
+        split_mixed=args.split_mixed,
     )
 
     if args.json_flag or args.format == "json":