tooling(gc): lightweight root/TLS/ratchet audits fail on Windows path and resource assumptions

[proggeramlug]

Summary

Three lightweight GC/performance audit commands fail on a supported Windows development host before they can evaluate repository state:

python scripts/gc_runtime_root_holders.py --self-test   # exit 1
python scripts/gc_runtime_root_holders.py               # exit 1
python scripts/check_thread_locals.py                    # exit 1
python benchmarks/gc_ratchet/gc_ratchet.py validate --scope structural
# ModuleNotFoundError: No module named 'resource'

The first two scanners build repository-relative keys with str(Path.relative_to(...)), which produces backslashes on Windows. Their checked-in inventories/allowlists and synthetic self-test expectations use forward-slash repository paths. Consequently every real declaration appears new and every checked-in entry appears stale; the root-holder self-test also reports every planted fixture as missed.

gc_ratchet.py imports the Unix-only resource module at module import time even though it is unused (peak RSS uses os.wait4) and validate --scope structural does not launch a measured process. Structural artifact validation should not require a Unix process API.

Impact

• Windows developers cannot run two fast gates that protect GC root coverage and hot-TLS policy.
• The ratchet's cheap structural validation cannot run before a heavy build on Windows, despite the harness explicitly separating platform-independent validation from measurement.
• This makes local preflight weaker on a platform Perry advertises and runs in GC native-root CI.

Evidence

• check_thread_locals.py converts relative paths with platform-native separators and again in the live scan
• gc_runtime_root_holders.py does the same for inventory keys
• gc_ratchet.py imports resource unconditionally, while its runner documents and uses os.wait4 instead.

Acceptance criteria

• Normalize all repository-relative identity keys with Path.as_posix() (including call-graph/module-edge keys and synthetic fixture expectations).
• Both scanners' --self-test and live verify modes pass on Windows and Unix.
• Remove the unused resource import or defer any genuinely Unix-only import to the measurement path with a clear platform refusal.
• gc_ratchet.py validate --scope structural and harness unit tests run on Windows without compiling Perry.
• Add a Windows CI step for these structural/self-test commands; it should not run the heavy ratchet measurement.