prompts: verify cross-function resource cleanup

[OllieinCanada]

Fixes #124

Summary

Require Sashiko to inspect cleanup helper implementations before reporting a cross-function resource leak.

Root cause

The resource-management prompts required ownership tracking across function boundaries, but the final false-positive verification did not explicitly require opening cleanup, unwind, destroy, unregister, or release helpers. A call site could therefore appear to leak a resource even when a conditional helper released it internally.

Changes

• inspect cleanup helper bodies and relevant callees rather than inferring behavior from names
• evaluate the condition guarding a cleanup call on the exact candidate path
• track the same object, field, reference, descriptor, or registration from acquisition to release
• account for nested helpers, managed resources, ownership transfer, and subsystem unregister behavior
• retain real findings when cleanup is bypassed, partial, applies to another resource, or cannot be verified
• require an auditable acquisition-to-release verdict

Validation

• confirmed the guide is always loaded during Stage 10 false-positive verification
• cargo fmt --all -- --check
• git diff --check
• exact one-file cumulative diff
• signed-off commit verified

No model-backed regression or paid API call was used.

[rgushchin]

This look reasonable to me, but do you have any examples or stats on how it helps in practice? Any specific false positives it helps to avoid? I strongly prefer to make prompts changes based on data.

[OllieinCanada]

I don't have aggregate stats. The source is #124, but that report doesn't include the affected patchset, code, or exact false-positive output. I agree that isn't enough evidence for a confident prompt change, so I've moved this back to draft and asked for a concrete reproducer before taking it further.