fix(analyzers): HIGH SC8 when skill ships __pycache__ or .pyc

[ajlennon]

Summary

Closes the silent bytecode skip in #356: __pycache__ is excluded from discovery and .pyc is treated as binary, so a skill can ship decoy sources plus malicious bytecode and still score SAFE.

This adds SC8 (Shipped Python Bytecode) — a HIGH finding whenever __pycache__/ or *.pyc/*.pyo is present under the skill tree (still skipping .git / node_modules / venvs). Presence alone is enough to fail closed; full .pyc disassembly can follow later.

Test plan

• uv run --with pytest pytest tests/nodes/analyzers/test_sc8_shipped_bytecode.py -q (2 passed)
• uv run ruff check on touched files

AI disclosure

This pull request was AI-assisted (Cursor) on behalf of Alex Lennon, with human review before open.

Fixes #356

Made with Cursor

[nedlir]

Like you mentioned, I think this solution is not comprehensive enough (we still skip node_modules, git files, etc) and need a more agnostic solution

[rng1995]

[Automated SkillSpector Review]

Requesting changes. SC8 detects the files, but it does not deliver the claimed fail-closed behavior: one .pyc scores 23 (MEDIUM / CAUTION) and the CLI exits 0; per-rule scoring caps three SC8 occurrences at 41. Add explicit blocking semantics and an end-to-end verdict/exit regression. Also run and commit Ruff formatting; ruff format --check fails on the touched analyzer file.

[Spectorian]

An explicit acceptance criterion for bytecode-poisoning scenario: shipped .pyc, .pyo, or pycache content must produce an explicit SC8 finding and must not complete as a passing or SAFE scan.
Please add a full end-to-end scan regression using a harmless poisoned-style fixture. It should verify that analysis completes successfully, SC8 identifies the shipped artifact, the finding is preserved in the report, and the final recommendation and exit behavior are appropriately blocking.
I also reproduced an integration failure where the absolute skill_path is passed to the inspection ledger, which rejects it because ledger paths must be relative. This leaves SC8 absent and changes the result only because analysis failed. A generic degraded-analysis result is not sufficient; the bytecode finding itself must be emitted.