fix(cli): thin-client pre-connectEngine routing seam for CLI-only commands

[2anoubis]

Problem

gbrain graph-query, gbrain salience, gbrain anomalies, gbrain think, gbrain recall, gbrain forget, gbrain whoknows, and gbrain founder all crash on thin-client installs with:

No database URL: database_url is missing from config

gbrain status works because it has a dedicated pre-connectEngine() thin-client guard. gbrain get works because it routes through the op-layer dispatch path which has its own thin-client routing seam.

Root Cause

Each of these commands has isThinClient() → callRemoteTool() branches in their handlers (added in v0.31.1, commit 6ba7feb), but handleCliOnly() calls connectEngine() at line 1568 BEFORE the switch statement can dispatch to them. The handlers never reach their thin-client code.

Fix

Add pre-connectEngine() thin-client guards in handleCliOnly(), matching the exact pattern already used by status (v0.41.19.0) and eval whoknows (v0.33.1.3). Each passes null for the engine argument — the handler detects isThinClient and routes through the remote MCP op instead.

Commands fixed

Command	MCP Op	Scope
graph-query	traverse_graph	read
salience	get_recent_salience	read
anomalies	find_anomalies	read
think	think	write
recall	recall	read
forget	forget_fact	write
whoknows	find_experts	read
founder	find_trajectory	read

Verification

Tested on a thin-client install (remote MCP mode, no database_url):

• gbrain graph-query people/bob ✅ — produces graph traversal output
• gbrain salience ✅ — produces salience table
• gbrain recall people/bob ✅ — no crash (empty results, correct)
• gbrain status ✅ — still works (no regression)