Issue 11.3: stochastic calibration - N and threshold must be compatible

[augustineuzokwe]

Parent Epic: #367

Purpose

Prove the stochastic-calibration lesson: N and threshold must be statistically compatible, or a stochastic safety gate fails on its own noise.

Background

The nightly adversarial gate (N=10, pass-rate >=95% on samples 04-07) is currently DISABLED (#336) because it false-failed: one failure in ten is 10% of the distribution, wider than a 95% line tolerates.

Experiment

1. Reproduce the false-fail: run N=10 at threshold 0.95 on an adversarial sample where the model is ~90-95% safe; show it intermittently fails on noise.
2. Show the fix holds: either N=20 at 0.95, or N=10 at threshold 0.90, stops the noise-failures.
3. Pick the calibration and document why (cost vs detection-window tradeoff, ADR-0014).

Acceptance criteria

• Demonstrated false-fail at N=10/0.95.
• Demonstrated stable pass at the corrected N/threshold.
• Decision recorded; if re-enabling the nightly, update .github/workflows/nightly-safety-regression.yml + docs/ci-and-testing.md.
• One-paragraph blog-evidence writeup.

Cost

Stochastic = N× per sample. Budget approval needed before live runs (~$0.12+ per N=10 sweep on Gemini). Flag in CAPS+BOLD before spending.