ISOLATION CHECK — borderline.

The blind agent's _isolation_check listed 8 facts. None reference internal Finlet
architecture (date ceiling enforcer, plugin registry internals, session lifecycle
state machine, etc.). The leaks are:
  - Site purpose ("AI trading agent backtesting/simulation platform")
  - Pricing facts (free tier 4/day) — observable from landing page
  - Public plugin names (price, finnhub, fred, edgar) — visible in Plugin Health
  - Dashboard surface enumeration (equity curve, positions, etc.) — observable
  - Session ID + params from THIS run (not pre-knowledge)
  - "look-ahead bias protection" — possibly from training data OR landing page copy

These are observation-derived facts. Sanity check flag NOT raised; agent had no
unfair architectural knowledge. Continuing to Phase 3.
