fix(qoder): report usage to all clients and stop inlining large attachments

- Coalesce Qoder's empty finish-in-delta frame with the later choices:[] usage
  frame so OpenAI and Claude clients receive prompt_tokens, completion_tokens
  and cache-hit tokens (the dashboard already saw them)
- Upload inlined images through /api/v2/image/upload like qodercli, and stub
  oversized non-image files instead of stuffing 30MB+ data URIs into
  agent_chat_generation
- Emit response.completed -> response.usage for chat-native upstreams so
  /v1/responses clients (Codex CLI, sub2api) no longer log 0/0/0
- Keep Claude message_delta.usage working when usage arrives without choices[0]
- Escalate to the smallest advertised Qoder context tier (200K/400K/1M) when
  the estimated prompt no longer fits max_input_tokens
- Pass apiKey for PAT connections and list hidden enable:false catalog keys
  from /v1/models

open-sse/providers/capabilities.js

@@ -222,9 +222,9 @@ export const PROVIDER_CAPABILITIES = {
   // windows (GLM-5.3 / Kimi-K3 / Qwen3.8-Max claim 180K but accept more).
   // max_output_tokens arrives as 0 for every model, so outputs are
   // best-guess from the real model family. Vision tags below follow the
-  // upstream is_vl flag per explicit request, even though the executor
-  // currently sends image_urls:null (image pass-through over the agent_chat
-  // SSE protocol is unverified). reasoning:true on all of them — every model can
+  // upstream is_vl flag. The executor uploads inlined images to
+  // /api/v2/image/upload and leaves image_urls/chat_context.imageUrls null
+  // (same as qodercli). reasoning:true on all of them — every model can
   // reason; the upstream is_reasoning flag only drives model_config selection.
   // thinkingFormat keeps the true-model family for documentation/UI, but
   // thinkingCanDisable:false everywhere: the executor only forwards