feat(security): SEC014/SEC015/SEC016 — second B-class batch

[wei9072]

Summary

Three more B-class rules following the α/β/γ pattern split from
PR #7. Each closes a real false-negative class observed in the
agent comparison study, all kept under FP rate < 1% via narrow
disambiguator filters.

Pattern	Rule	Description
α text-level	SEC014	hardcoded -----BEGIN ... PRIVATE KEY----- PEM block
β multi-language	SEC015	silent broad except — Python except Exception: pass, JS catch (e) {}
β multi-language	SEC016	SSRF marker — HTTP call where URL arg is request input

Commits (review in order)

1. feat(security): SEC014 (α) — hardcoded PEM private key —
   text scan; public keys excluded.
2. feat(security): SEC015 (β) — silent broad except / empty catch —
   AST on except_clause (Python) / catch_clause (JS / TS).
   Body-emptiness logic mirrors the existing empty_handler_count
   signal in signals/smells.rs. Bare except: stays SEC013's job.
3. feat(security): SEC016 (β) — SSRF marker on HTTP call —
   per-language receiver gating filters dict.get / cache.get
   noise. URL-arg pattern list covers Flask / Django / FastAPI /
   Express / Koa request shapes.
4. docs(readme): bump SEC rule count 13 → 16 + name new rules —
   EN + ZH README tables.

Skipped (deferred or out of scope)

These were on the backlog but excluded from this PR because they
can't hit FP rate < 1% without dataflow analysis:

• Path traversal (../ in path-like contexts) — too many legit
  relative imports / paths.
• IDOR (/api/x/<id> without owner check) — needs route + auth
  context.
• N+1 query (loop with DB call inside) — needs DB call awareness.
• Float for money (variable named price/amount) — heuristic
  too broad, lots of legit non-money use.

If we later add a dataflow / type pass, those become viable.

Test results

cargo test --workspace — 130 / 130 pass (was 116 baseline; +14
new tests covering positive and must-NOT-fire negatives per rule).

Live MCP sanity-check on a single Python file with all 3 patterns:

$ python3 aegis_validate.py /tmp/b2-test.py
FINDINGS ...: 20 total (security=3, signal=16, workspace=1)
  [security] SEC014 @line 3   hardcoded PEM private key ...
  [security] SEC015 @line 13  broad exception caught and silently swallowed ...
  [security] SEC016 @line 8   outbound HTTP call with URL derived from request input ...

All three patterns were silent before this PR.

Test plan

• cargo test --workspace — 130 / 130 pass
• cargo install --path crates/aegis-mcp --force succeeds
• Live MCP run fires SEC014 + SEC015 + SEC016 on test file
• False-positive guards verified by negative tests:
  • -----BEGIN PUBLIC KEY----- → no SEC014
  • except ValueError: pass → no SEC015 (specific type, narrow)
  • except Exception as e: log(e); raise → no SEC015 (body has work)
  • requests.get("https://api.example.com/data") → no SEC016 (static URL)
  • data.get(req.params.url) → no SEC016 (dict, not HTTP client)
• CI green on push

🤖 Generated with Claude Code