fix: recalibrate FG-002/FG-003 high band to match real exploitability

[north-echo]

Summary

Triaging a mid-tier scan surfaced that the high-severity band was inflated by two low-actionability classes, drowning out the genuinely attacker-controllable findings.

• FG-002 Script Injection — inputs.* / github.event.inputs.* are populated only by a write-gated manual dispatch or an internal workflow_call; they're never reachable directly from an untrusted trigger. Those drop a tier (high → medium, with a note to verify reusable-workflow callers). Attacker-controllable expressions (pull_request.title, head_commit.message, comment.body, head_ref, …) stay high.
• FG-003 Tag-Based Pinning — branch pins (@main/@master/@dev) are supply-chain hygiene, not an acute exploit. Capped at medium (down from high) with a "mutable branch ref" note. Version tags stay medium; first-party actions/* stay info.

Measured impact (277-repo mid-tier corpus)

• High-severity findings: 989 → 171 (−83%)
• FG-002 highs: 467 → 14 (the 14 survivors are all attacker-controllable expressions; 270 moved to medium)
• FG-003 highs: 347 → 0
• The two real pwn-request criticals: unchanged

The high band now means "an untrusted actor can influence this," which is the point.

Testing

• Positive-control fixtures/tests: pull_request.title injection stays high; third-party @main pin is medium.
• Full go test ./..., gofmt, vet, and self-scan clean.

🤖 Generated with Claude Code