fix: classify str_replace and insert tools as mutating edits

[MiltonHeYan]

What's wrong

classify_tool_call identifies the tool family from the tool name using a set of regex patterns. The pattern for the "edit" family is:

if re.search(r"write|edit|patch|apply|create|delete|rename", name):
    return "edit", True

"replace" and "insert" are missing. As a result, tools like str_replace, replace_in_file, insert_text, and insert_at_line fall through every check and return ("unknown", False) — classified as non-mutating, unknown family.

Impact

For any agent that edits files via str_replace:

• distinct_mutation_targets is always empty → min_distinct_mutation_targets requirement always fails
• read_before_write_ratio is 1.0 for the wrong reason — no mutations are detected, so the denominator collapses to 1
• "edit" never appears in distinct_families → required_families always reports it as missing
• Trajectory score is silently wrong without any error

Fix

Extend the edit pattern:

# before
if re.search(r"write|edit|patch|apply|create|delete|rename", name):
# after
if re.search(r"write|edit|patch|apply|create|delete|rename|replace|insert", name):

Tests

Added two new tests in test_trajectory.py:

• test_replace_and_insert_tools_are_classified_as_edit — unit test for classify_tool_call directly
• test_str_replace_mutation_is_detected_in_trajectory — end-to-end trajectory test verifying that mutations via str_replace are properly detected

All 7 tests pass.