[devdanzin]

Summary

• Health analysis of ~/runs/jit_run12 (session-fuzz) found 37.5% of ignored_crash events (2,150 of 5,738) were children that fail at compile() in the executor and never run — 0 coverage, 0 JIT trace, pure wasted executions that also count toward sterility/coldness.
• Root cause: the mutation pipeline validated children only via ast.parse/ast.unparse round-trips, which do not catch match-statement semantic errors — only compile() does. Dominated by alternative patterns bind different names (from _add_pattern_complexity wrapping a name-binding pattern in case P | [_]:) and mapping pattern checks duplicate key (from a constant mutator flipping a key in case {'x': _, 'y': _}:).
• Fix 1 (general): prepare_child_script now compile()s the generated core and skips the mutation on SyntaxError, logging a core_code_syntax_error health event. compile() is far cheaper than the wasted subprocess run it prevents, and it kills the entire class (both variants + any future semantic-SyntaxError generator). SyntaxWarnings are suppressed during this side-effect-free validation.
• Fix 2 (targeted): _add_pattern_complexity now skips patterns that capture any name, since the alternatives it adds bind none.

Test plan

• TestCompileGate: a child that round-trips through ast.parse but fails compile() is rejected (returns None, records the health event); a valid child passes and the assembled child compiles.
• _pattern_binds_names detection (capture / star / mapping-rest); _add_pattern_complexity skips name-binding patterns and still wraps non-binding ones — both outputs compile.
• Full suite: 2182 passed, 492 subtests (no new warnings).
• mypy lafleur clean (38 files); ruff check / ruff format clean.

Closes #891

🤖 Generated with Claude Code