terra: add Hermes Agent, wired to local ollama

Points Nous Research's Hermes Agent at terra's own ROCm ollama server
(gemma4:12b) as a custom OpenAI-compatible provider instead of a cloud
key. Native systemd mode via the hermes-agent flake's own NixOS module
— simpler than container mode, avoids the podman-rootful-sudo dance
its docs call out.

Also bumps OLLAMA_CONTEXT_LENGTH (and Hermes' mirrored model.context_length)
from ollama's ~4k default to 131072, load-tested with real multi-ten-
thousand-token prompts rather than just idle `ollama ps` checks — chosen
as the practical ceiling where VRAM headroom and prefill throughput both
start visibly degrading, not just the largest number that technically fit.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

hosts/terra/configuration.nix

@@ -18,6 +18,7 @@ in
     ../../services/desktop/desktop-hyprland.nix
     ../../services/desktop/desktop-apps.nix
     ../../services/desktop/librechat.nix
+    ../../services/desktop/hermes-agent.nix
   ];
 
   networking.hostName = "terra";
@@ -95,6 +96,24 @@ in
     # Sparse activation makes that far less painful than it'd be for a dense
     # model this size, but still expect it to run slower than the two above.
     loadModels = [ "gemma4:12b" "qwen3.6:35b-a3b" ];
+    # Ollama truncates context far below the model's real window unless
+    # told otherwise (the OpenAI-compat /v1 route it's reached through has
+    # no way to set this per-request — see services/desktop/hermes-agent.nix,
+    # which points at this server and mirrors the same number in
+    # model.context_length). 131072 chosen as the practical ceiling after
+    # load-testing with real prompts, not just idle `ollama ps` checks:
+    # 32768 (31.6k-token prompt) and 65536 (40.8k-token prompt) both stayed
+    # 100% GPU with VRAM barely moving (~10.1G / ~10.67G of 16G) — KV cache
+    # cost barely grows with context, likely sliding-window/local attention
+    # on most of gemma4:12b's layers. At 131072 that stopped being true: a
+    # ~108k-token prompt pushed VRAM to ~11.4G/16G (still 100% GPU, no CPU
+    # spillover, negligible GTT) but with visibly shrinking headroom, and
+    # prefill throughput measurably dropped (~490 -> ~460 tok/s) over just
+    # the last 13k tokens — filling the full window would take minutes of
+    # pure prompt processing. Stopped here rather than push further: next
+    # doubling would risk CPU spillover under any concurrent GPU load
+    # (desktop compositor, jellyfin transcode) for diminishing benefit.
+    environmentVariables.OLLAMA_CONTEXT_LENGTH = "131072";
   };
 
   # ---- Dev-data disks — NOT in disko, mounted read-write, never wiped ----