P2: Fix 'Nothing leaves the runner' claim — scan vs verify modes differ

[rahuldass19]

Summary

The README's Security & Privacy section says "Nothing leaves the runner" — but this is only true for scan modes. The verify mode (LLM output cross-examination) requires a configured QWED backend unless using a self-hosted deployment.

Why this matters

QWED's philosophy is honest, transparent verification. An overclaim in the security README undermines that — a user who reads "nothing leaves the runner" may be surprised that verify mode makes a backend call.

Fix

Replace the bold claim with an accurate, tiered statement:

Scan modes (scan-secrets, scan-code, verify-shell, verify-process) execute entirely on the GitHub runner — no external call.
Verification mode (verify) may call a configured QWED backend unless using a self-hosted or local QWED deployment.

This is honest and matches the README's own note in the Verification Modes section (line ~119) which already acknowledges verify calls the QWED API.

Acceptance Criteria

• The "Nothing leaves the runner" claim is replaced with an accurate tiered statement
• The security section is consistent with the Verification Modes note
• The privacy claim ("Nothing is learned from") remains accurate