Add reproducible 90-sample security benchmark

[tiltom]

What changed

• adds a pinned 90-sample benchmark corpus and deterministic runner
• adds 25 metadata-only OpenSSF OSV malicious-package samples
• records canonical pre-fix and post-fix results
• adds manual review notes for every severe reference finding
• refreshes the bundled signature snapshot to 2026.07.12.003
• adds benchmark pin/run scripts and regression coverage

Why

Flagrix needed reproducible evidence for malicious recall and false-positive launch gates. The runner never clones, installs, or executes benchmark samples.

Results

• overall: 37/90 → 90/90 passed
• critical fixtures: 9/10 → 10/10
• malicious metadata: 5/25 → 25/25
• reference High verdicts: 27/55 → 0/55
• reference Critical findings: 16 → 0

Rollout dependency

The benchmark uses scanner-core from the adjacent local checkout. Before this PR is made ready or the CLI is released, update the npm dependency and lockfile from scanner-core 0.2.0 to the published 0.2.2 from flagrix-scanner-core#1.

Validation

• 15 CLI tests passed
• TypeScript type-check passed
• all four benchmark gates pass