userConfig design

Two tunables, both optional with safe defaults. They are prompted at enable time and exported as CLAUDE_PLUGIN_OPTION_<KEY> env vars the bin scripts read.

key	type	default	effect
gate_mode	string	full	full = the hard gate is active (deny destructive actions until grounded); off = gate disabled. The soft reflex (the ground-first skill) is always available and model-elected regardless.
hard_gate	boolean	true	false makes the gate advisory-only — it stops denying (defers to the normal permission flow) while the skill reflex still works. A kill-switch.

Wiring

key	env var	read by
gate_mode	CLAUDE_PLUGIN_OPTION_GATE_MODE	tmt_enforce.py, tmt_log.py, tmt_reconcile.py
hard_gate	CLAUDE_PLUGIN_OPTION_HARD_GATE	tmt_enforce.py (via tmt_lib.hard_gate_enabled())

CLAUDE_PLUGIN_OPTION_GATE_MODE (the user's plugin config) takes precedence over TMT_ARM (the eval-harness escape hatch). Both fall back to full.

There is no trigger_sensitivity or gate_tier: the soft trigger is the model's judgment (nothing to tune with a knob), and the gate denies by tool reversibility, not a blast-tier setting.