Intelligence Bar write tools — the trust boundary

The trust boundary is unconditional (GATE_IB_UI_CONFIRM is retired and
ignored); the emergency stop is IB_WRITES_DISABLED=true. An IB write tool
never executes from the model loop. The flow is:

1. The tool call returns a preview (no side effects).
2. The route persists a pending action in ib_pending_actions
   (actor-bound, 10-minute expiry, payload hash, single-use).
3. The client renders a Confirm/Cancel card (PendingActionsCard).
4. Only the operator's click commits, via /confirm-action — never a model
   tool; the pending id is never model-visible.

Adding a new write tool — checklist

1. Implement the tool in the context module
   (server/services/intelligence-bar/{context}-tools.js) following the
   TOOLS + executeTool pattern (see
   server/services/intelligence-bar/README.md).
2. Register it in server/services/intelligence-bar/write-gates.js.
   A write tool missing from the gate set executes directly from the model
   loop — that's a trust-boundary breach, not a style nit.
3. Update the mirror test
   server/tests/intelligence-bar-write-gate-contract.test.js — it exists to
   make step 2 unforgettable; it should fail until the registration is right.
4. The preview must show the operator everything the commit will do
   (recipient, amount, date — no hidden fields).
5. Tech-portal contexts stay read-only. Never add a write tool to
   tech-tools.