feat(gmw): route all LLM traffic through 9router

GMW moves off omniroute (100.121.180.82:20128) and off the direct NVIDIA
vision endpoint onto 9router, which runs on the same host as both services
(127.0.0.1:4014) — loopback avoids the TLS/proxy hop and localhost calls
bypass 9router's remote-key guard.

- gateway + backend: AI_LLM_BASE_URL default -> http://127.0.0.1:4014/v1
- drop stale 'omniroute' router references from comments/docs now that the
  active router is 9router (llmClient, llmCaller, ARCHITECTURE, AGENTS)

Verified against 9router before wiring: model 'text' -> gemini-3.5-flash-lite
(SSE, as the pipeline expects), 'multimodal' -> nemotron-3-nano-omni answers
image input, and gemini/gemini-embedding-001 returns 3072 dims — matching the
existing Qdrant collections (no reindex needed). The GMW key is already
registered in 9router's apiKeys table.

typecheck + lint + tests green (gateway 138, backend 37 excluding e2e).

services/backend/src/shared/config/index.ts

@@ -96,10 +96,10 @@ export const configSchema = z
       .transform((v) => v === "true")
       .default(false),
     AI_LLM_API_KEY: z.string().optional(),
-    AI_LLM_BASE_URL: z
-      .string()
-      .url()
-      .default("http://100.121.180.82:20128/api/v1"),
+    // 9router — OpenAI-compatible router on this host (127.0.0.1:4014).
+    // Loopback on purpose: backend runs on the same machine as 9router, so no
+    // TLS/proxy hop is needed.
+    AI_LLM_BASE_URL: z.string().url().default("http://127.0.0.1:4014/v1"),
     AI_LLM_MODEL: z.string().default("text"),
     AI_LLM_VISION_MODEL: z.string().optional(),
     AI_LLM_EMBEDDING_MODEL: z.string().optional(),

services/discord-gateway/AGENTS.md

@@ -54,7 +54,7 @@ src/
    **Never** reintroduce regex/heuristic content classification.
 2. **Discord tokens sanitized** before reaching LLM (`discordTokens.ts`).
 3. **Semantic cache is batched** — one embed call + one Qdrant batch search.
-4. **Streaming is mandatory** against the omniroute base URL.
+4. **Streaming is mandatory** against the router base URL.
 
 ## AI moderation pipeline
 

services/discord-gateway/ARCHITECTURE.md

@@ -194,5 +194,5 @@ pipeline gauges registered by `app/metrics-collector.ts` —
   numeric snowflake IDs never trigger false positives.
 - **Semantic cache is batched** (one embed call + one Qdrant batch search),
   not N sequential round-trips. `ensureQdrantCollection` is memoized.
-- **Streaming is mandatory** against the omniroute base URL (non-stream waits for
+- **Streaming is mandatory** against the router base URL (non-stream waits for
   the full body and times out). `llmClient` aggregates SSE chunks.

services/discord-gateway/src/modules/ai-moderation/llmCaller.ts

@@ -92,7 +92,7 @@ export async function callModerationLLM(
             jsonResponse: { type: "json_object" },
             retries: 0,
             signal,
-            // Router (omniroute) always streams SSE even when the
+            // Router always streams SSE even when the
             // request omits `stream`. In non-stream mode the OpenAI SDK waits
             // for the FULL body before parsing, so slow/long upstream streams
             // hit the 30s/60s timeout and abort mid-generation. Streaming mode

services/discord-gateway/src/modules/ai-moderation/llmClient.ts

@@ -131,7 +131,7 @@ type LLMResponseChunk = {
  * `delta.content`; falls back to reasoning fields so reasoning-only models
  * still produce usable aggregated text. Providers differ in the field name:
  * - DeepSeek-style / Cloudflare gemma → `delta.reasoning_content`
- * - mimo (via omniroute) streams reasoning in `delta.reasoning` +
+ * - mimo (via the router) streams reasoning in `delta.reasoning` +
  *   `delta.reasoning_details[].text` (content:"") — without these fallbacks
  *   vision aggregation came back empty ("Vision API null response").
  * Exported for unit tests.
@@ -267,7 +267,7 @@ export function buildLlmParams(
       reasoning: { enabled: false },
       // vLLM / Qwen / litellm
       chat_template_kwargs: { enable_thinking: false },
-      // Anthropic / Claude-format (omniroute exposes thinkingFormat
+      // Anthropic / Claude-format (router exposes thinkingFormat
       // "claude-adaptive" / "claude-budget" on its reasoning models)
       thinking: { type: "disabled" },
     } as Record<string, unknown>);

services/discord-gateway/src/shared/config/index.ts

@@ -149,10 +149,12 @@ export const configSchema = z
       .transform((v) => v === "true")
       .default(false),
     AI_LLM_API_KEY: z.string().optional(),
-    AI_LLM_BASE_URL: z
-      .string()
-      .url()
-      .default("http://100.121.180.82:20128/api/v1"),
+    // 9router — the OpenAI-compatible router on this host (127.0.0.1:4014).
+    // Loopback on purpose: the gateway runs on the same machine as 9router,
+    // so no TLS/proxy hop is needed (and localhost bypasses 9router's
+    // remote-key guard). Public alias https://9router.asepharyana.my.id/v1
+    // works too but requires the key for every call.
+    AI_LLM_BASE_URL: z.string().url().default("http://127.0.0.1:4014/v1"),
     AI_LLM_MODEL: z.string().default("text"),
     // Vision uses the SAME router/base URL as text moderation
     // (AI_LLM_BASE_URL) but a different model alias. The dedicated NVIDIA