Improve tool-call friction signals

[hamzamerzic]

Summary

• correlate project-local skill reads with later authoritative skill reads so avoidable indirection becomes visible as one causal chain
• exclude path-only reads and edits from exact-repeat claims because retained transcripts omit their ranges or mutations
• surface successful shell no-op calls as a separate high-confidence waste signal

Why

The existing privacy-safe command families introduced in #29 make recurring mechanics interpretable, but they cannot connect two differently shaped reads or distinguish repeated targets from repeated calls. This follow-up builds on #15 and #29 by adding bounded causal signals without retaining raw command text.

Verification

• 5 focused Python tests passed
• app validation passed with 0 errors and 0 warnings
• the detector reports skill indirection, no-op calls, and exact-repeat candidates separately

[miljanm]

Reviewer: all clear

I didn't find a concrete issue in this revision. This is a QA second look, not a maintainer approval.

Reviewed revision f8741895e5fe.