docs(hermes): align training with Codex-first routing

services/hermes/skills/master-hermes-on-atlas/SKILL.md

@@ -11,7 +11,7 @@ substitute a lecture for a lab.
 
 ## Select the training model
 
-Use `openai-codex/gpt-5.4` for assessments, multi-reference labs, incident
+Use `openai-codex/gpt-5.6-terra` for assessments, multi-reference labs, incident
 grading, and skill evaluation. If the active provider is the local
 `gpt-oss:20b`, stop before reading the reference files and ask Brad to select
 Codex in Models or start a Codex-backed session. Continue locally only when Brad

services/hermes/skills/master-hermes-on-atlas/references/architecture.md

@@ -18,23 +18,25 @@ through the chat Service or Ingress.
 ## Inference path
 
 ```text
-operator Hermes ─┐
+operator Hermes ───> independently authenticated openai-codex/gpt-5.6-terra
-                 ├─> hermes-model-gate ─> Ollama gpt-oss:20b on titan-24 GPU
+consumer Hermes ───> independently authenticated openai-codex/gpt-5.6-terra
-consumer Hermes ─┘             │
+       │
-                               └─ 503 while Wolf owns the GPU
+       └─ provider error or manual selection ─> hermes-model-gate
-
+                                                │
-each instance independently ─────> openai-codex/gpt-5.4 fallback
+                                                ├─ Hermes owns GPU ─> Ollama gpt-oss:20b
+                                                └─ Wolf owns GPU ───> HTTP 503
 ```
 
 - The configured context length is 64,000 tokens.
 - `hermes-model-gate` is the stable OpenAI-compatible endpoint.
 - A coordination Lease stores whether Hermes or Wolf/Moonlight owns the GPU.
   Ariadne has narrowly scoped permission to update that Lease during handoff.
-- Wolf ownership makes the gate return a deliberate unavailable response. The
+- Wolf ownership makes the local gate return a deliberate unavailable response.
-  Hermes gateway stays alive and can use its configured fallback.
+  Normal interactive chat remains on its Codex primary and does not consume
+  titan-24 GPU resources.
 - The two agent pods do not consume titan-24 GPU memory. Ollama does. Agent pods
   are ARM gateway workloads with persistent state on separate PVCs.
-- Fallback credentials are deliberately per-instance. Never copy Brad's Codex
+- Codex credentials are deliberately per-instance. Never copy Brad's Codex
   credential store into the consumer PVC.
 
 Verify rather than memorize:

services/hermes/skills/master-hermes-on-atlas/references/curriculum.md

@@ -4,7 +4,7 @@ Complete labs by evidence, not elapsed time. A focused pass can establish
 operational competence in several days; mastery requires repeating real triage
 and recovery work over multiple incidents.
 
-Run the curriculum on `openai-codex/gpt-5.4`. Use the local model only for a
+Run the curriculum on `openai-codex/gpt-5.6-terra`. Use the local model only for a
 deliberate comparison lab; it is not the default coach for multi-reference work.
 
 ## Phase 1: orientation and control
@@ -22,8 +22,8 @@ agent to model, including the fallback branch and the consumer boundary.
 ### Lab 2 — Models, context, and fallback
 
 Inspect `hermes status`, `hermes fallback list`, deployment placement, and GPU
-owner state. Explain why a 32K model was rejected, why the gateway remains up
+owner state. Explain why a 32K model was rejected, why normal Codex-backed chat
-when Wolf owns the GPU, and which credentials a fallback consumes.
+remains available when Wolf owns the GPU, and when the local fallback can run.
 
 Success evidence: predict outcomes for local healthy, local slow, gate 503,
 invalid local response, and expired Codex authorization without changing state.

services/hermes/skills/master-hermes-on-atlas/references/two-hour-proof-sprint.md

@@ -9,12 +9,13 @@ Brad commits to a result.
 
 ## 0–15 minutes — control surface and request path
 
-1. Have Brad select `openai-codex/gpt-5.4` for the training session.
+1. Have Brad start a new session and confirm it uses
+   `openai-codex/gpt-5.6-terra`.
 2. In the web UI, have him locate Chat, Sessions, Files, Models, Logs, Skills,
    Plugins, MCP, Channels, Webhooks, Cron, Profiles, Config, Keys, System, and
    Documentation.
-3. Have him explain browser → operator agent → model gate → Ollama, including
+3. Have him explain browser → operator agent → Codex primary, including the
-   Codex fallback and the separate consumer instance.
+   model-gate/Ollama fallback branch and the separate consumer instance.
 
 Evidence: Brad can identify persistent, shared, and instance-local state and
 can name the GPU owner check without changing it.