fix: cap wildcard import expansion to avoid token explosion

[mashraf-222]

Problem

Wildcard imports like import org.jooq.* expand to 870+ types, causing 5 minutes of disk I/O per function before discovering the 4000-token skeleton budget is exceeded. In jOOQ, 89% of functions (70/79) were skipped due to token overflow from wildcard imports.

The expand_wildcard_import() function globs all .java files in the package directory unconditionally, and the token budget check in get_java_imported_type_skeletons() only fires after reading each file and parsing its skeleton — by which point hundreds of files have already been read from disk.

Root Cause

context.py:933-940: Wildcard expansion happens without any count limit or early bailout.
import_resolver.py:223-252: expand_wildcard_import() returns all types unconditionally.

Fix

import_resolver.py

• Added max_types parameter to expand_wildcard_import() for early termination
• Added filter_names parameter to only include types matching a given set

context.py

• Added MAX_WILDCARD_TYPES_UNFILTERED = 50 constant
• When a wildcard expands to >50 types:
  • If target code references specific types → re-expand with filter_names=priority_types (only referenced types)
  • If no target types available → cap at 50 (first 50 found)
• Small wildcards (<50) are expanded fully as before

This turns a 5-minute failure into <1 second resolution with only the relevant types included.

Test Coverage

New test test_large_wildcard_is_filtered_to_referenced_types:

• Creates 70 types in a package (exceeds cap of 50)
• Target code references only Type000 and Type001
• Verifies only referenced types appear in result, not the full 70

All 4 existing edge case tests pass unchanged.

Closes CF-1085

🤖 Generated with Claude Code

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[mashraf-222]

Review

Bug premise verified — real. expand_wildcard_import() in codeflash/languages/java/import_resolver.py on main globs *.java unbounded (lines 223-252). The token-budget guard in codeflash/languages/java/context.py:1005 is checked AFTER per-file read + skeleton extraction + tokenization — so for org.jooq.* (~870 types) the pipeline still does 870 disk reads even though only a handful of tokens will fit the budget. The 89%-skip / 5-min-per-function observation on jOOQ is consistent with that.

Fix is well-scoped. Two-phase expansion (count-capped probe → filter-by-referenced-types OR truncate to first 50) matches the neighboring IMPORTED_SKELETON_TOKEN_BUDGET = 4000 guard style; test in test_context.py covers the happy path (70 types, 2 referenced).

CI blockers addressed in the last two commits:

1. prek / prek was failing on 9 pre-existing mypy errors in context.py that the prek hook surfaced because this PR touches the file — none caused by the PR itself. Fixed all of them:
   • Import Language from language_enum directly (the re-export via base.py is implicit and strict mypy rejects attr-defined).
   • Annotate _extract_class_declaration, _import_to_statement, get_java_imported_type_skeletons, and resolved_imports: list[ResolvedImport].
   • Guard None start/end_line in _extract_function_source_by_lines and find_helper_functions; guard None resolved.file_path in the import skeleton loop.
   • Drop the unreachable if not node: continue in _extract_public_method_signatures (JavaMethodNode.node is non-nullable).
   • Plus -> None on all 94 tests in test_context.py and one int | None comparison fix.
     880/880 Java tests pass locally.
2. unit-tests (windows-latest, 3.13) — test_help_banner.py was failing on main (Rich box-drawing characters undecodable under cp1252). Added encoding="utf-8" to both subprocess.run calls — same fix as PR fix: decode help-banner test subprocess output as UTF-8 #2120.

Non-blocking gaps worth following up on: no direct unit tests for expand_wildcard_import(max_types=…) / filter_names=… on the resolver (only integration-tested); no boundary test at exactly 50/51; the priority_types-empty + overflow fallback path at context.py:953 is not tested. Worth a small follow-up.

Ready for re-review.