Automated evolution PR for issue #1501.

Problem

The patch tool's other error bucket accounts for 68 failures/week (87% of all patch errors). These failures are classified as unknown by classify_error() and other by _classify_failure_reason(), giving the model no actionable recovery path — leading to retry spirals on unknown causes.

Solution

Decompose the other/unknown bucket into actionable sub-categories:

1. old_string_empty — classify_error() now recognizes "old_string cannot be empty" (the error from fuzzy_find_and_replace line 73) and classifies it as old_string_empty instead of falling through to unknown.

2. indentation_mismatch — a secondary classification pass in format_structured_error() that triggers when the primary classifier returns no_match. It finds the best-matching content line, then compares leading whitespace across ALL corresponding lines (the mismatch is often on a subsequent line, not the anchor). When the text matches but indentation differs (tabs vs spaces, different indent levels), it reclassifies as indentation_mismatch.

3. Recovery hints — both new types get specific recovery suggestions: old_string_empty tells the model to provide a non-empty old_string; indentation_mismatch tells it to use read_file to see exact whitespace and update old_string to match.

4. Introspection sub-categories — _classify_failure_reason() in scripts/introspection_extract.py gets three new regex patterns (_PATCH_OLD_STRING_EMPTY_RE, _PATCH_INDENTATION_MISMATCH_RE, _PATCH_AMBIGUOUS_RE) checked BEFORE the generic _NO_MATCH_RE, so specific patch failures are tagged patch-old-string-empty, patch-indentation-mismatch, or patch-ambiguous instead of falling into other.

Files changed

• tools/fuzzy_match.py — new ERROR_TYPES entries, classify_error sub-categories, format_structured_error secondary classification + recovery hints
• scripts/introspection_extract.py — new regex patterns + classification chain entries
• tests/tools/test_patch_self_correction.py — tests for new classify_error and format_structured_error paths

Validation

• Lint: ✓ (ruff check)
• Tests: ✓ (160 passed — test_patch_self_correction + test_introspection_extract + test_fuzzy_match + test_patch_failure_tracking)
• Diff: 128 insertions, 0 deletions (within 200-line merge cap)

Closes #1501