GateGuard fact-force restate-retry loop amplifies model repetition traps in long sessions

## Context `gateguard-fact-force` requires restating 4 facts before each Bash/Edit/Write, then blocks the operation once and forces a retry. In a long session this produced **10+ near-identical "state facts → blocked → restate → retry" blocks** in the same context window. ## Problem This highly repetitive, near-duplicate structure measurably raises the odds of the underlying model dropping into a **degenerate repetition loop**. In one long session I observed **3 separate collapses** where the model emitted a single filler token (`court`) thousands of times until manually interrupted. The collapse consistently started at the transition right before a guarded tool call. The guard itself is well-intentioned, but its forced repetition is a contributing factor to model instability in long sessions. ## Reproduction 1. Run a long autonomous session (50+ tool calls) with GateGuard enabled 2. Each guarded Bash/Edit/Write triggers: state facts → blocked → restate same facts → retry 3. After 10+ such cycles, the model enters a degenerate single-token repetition loop that cannot self-recover ## Suggestions - **Don't re-trigger the full fact-force on a *retry* of the same operation** — the facts were stated one step prior - Add a per-session cache: "facts already established for file X" to avoid restating identical facts - Document `ECC_GATEGUARD=off` / `ECC_DISABLED_HOOKS` more prominently for long autonomous sessions ## Environment ECC `2.0.0-rc.1`, Claude Code, macOS, claude-opus-4-8 model