[act.research.ρ.D8+] supervise the authorized local Qwen endpoint

The governed loop had a real JSON-schema provider path but no durable host
boundary for the explicitly authorized RTX 5080 deployment.  Add a loopback-only
user service with bounded memory, private writable caches and telemetry disabled;
make the backend consume an optional operator-owned environment file so provider
URLs and the shared-GPU grant remain outside Git.  Document that this grant does
not promote model output to evidence or silently authorize physical FEP, and
regenerate the architecture twin after the runtime topology changed.

Reality checks: the cold service reached /health with zero restarts, vLLM reported
Qwen3.5-2B and held 10,534 MiB on the RTX 5080, a real ai.research.propose Job
persisted a schema- and semantics-accepted proposal Artifact in one validation
attempt, and all 21 repository gates passed.

Author: ivanicu

backend/README.md

@@ -114,6 +114,11 @@ profiles remain fail-closed unless the host operator sets the exact
 proposals still pass the frozen schema, current-action and scientific-evidence validators.
 The production loop requires forward migrations 049 and 050.
 
+The canonical workstation's optional local endpoint is supervised by
+`deploy/systemd/dirac-qwen.service` on loopback port 8930. Provider profiles and the
+host-scoped grant live in gitignored `deploy/ai/providers.local.json` and
+`deploy/ai/dirac-ai.env`; no provider URL or credential comes from a request payload.
+
 ```bash
 bash scripts/gates.sh research
 

deploy/README.md

@@ -8,6 +8,7 @@ The canonical local deployment has one current topology:
 | `dirac-fields.service` | 8901 | unified application, Command, Method, Job and artifact service |
 | `dirac-ops.service` | 1355 | read-only operations projection |
 | `dirac-digital-twin.service` | — | source watcher and architecture-twin regeneration |
+| `dirac-qwen.service` | 8930 (loopback) | optional local Qwen3.5 Research Loop inference |
 
 The standalone `dirac-physics.service` on 8902 is superseded and retained only under
 `deploy/systemd/_archive/`. Physics implementations are reached through the unified 8901
@@ -43,6 +44,22 @@ systemctl --user daemon-reload
 systemctl --user enable --now dirac-web dirac-fields dirac-ops dirac-digital-twin
 ```
 
+Local Qwen is intentionally a separate opt-in unit. Copy the provider example to the
+gitignored `deploy/ai/providers.local.json`, mark the workstation profile
+`shared_dirac_gpu`, and place its paths plus the exact
+`DIRAC_ALLOW_SHARED_GPU_AI=1` grant in gitignored `deploy/ai/dirac-ai.env`. Then:
+
+```bash
+systemctl --user link "$PWD/deploy/systemd/dirac-qwen.service"
+systemctl --user enable --now dirac-qwen
+systemctl --user restart dirac-fields
+curl --fail http://127.0.0.1:8930/health
+```
+
+The grant authorizes this host configuration; it does not turn model output into
+evidence or make physical FEP schedulable. Keep physical GPU execution disabled until a
+single lease owner coordinates model residency and FEP admission.
+
 Verify state, socket ownership and HTTP responses rather than treating installation as
 proof:
 

deploy/systemd/dirac-fields.service

@@ -24,6 +24,9 @@ Environment=PYSCF_TMPDIR=/home/ivan/.cache/dirac/scratch
 # CPU methods remain local inside the bounded controller pool.
 Environment=DIRAC_EXECUTOR=kubernetes
 Environment=DIRAC_KUBERNETES_EXCHANGE_HOST=/home/ivan/dirac/.runtime/pv/exchange
+# Optional operator-owned provider URLs, credentials and explicit shared-GPU grant.
+# The ignored file is absent on hosts that intentionally run without AI inference.
+EnvironmentFile=-/home/ivan/dirac/deploy/ai/dirac-ai.env
 ExecStart=/home/ivan/dirac/backend/env/bin/python backend/field_server.py
 Restart=on-failure
 RestartSec=5

deploy/systemd/dirac-qwen.service

@@ -0,0 +1,40 @@
+# dirac-qwen.service — loopback-only Qwen3.5 inference for the Research Loop.
+# The shared RTX 5080 profile remains unusable unless dirac-fields also receives
+# DIRAC_ALLOW_SHARED_GPU_AI=1 through its operator-owned environment file.
+
+[Unit]
+Description=Dirac local Qwen3.5 inference (:8930)
+After=network.target
+StartLimitIntervalSec=300
+StartLimitBurst=5
+ConditionPathExists=/home/ivan/.local/share/dirac-qwen-vllm/bin/vllm
+ConditionPathExists=/mnt/e/data.ai-models.local-model-store.storage.xl.private.readonly/Qwen3.5-2B/config.json
+
+[Service]
+Type=simple
+WorkingDirectory=/home/ivan/dirac
+Environment=CUDA_VISIBLE_DEVICES=0
+Environment=VLLM_NO_USAGE_STATS=1
+# FlashInfer's sampler JIT searches /usr/local/cuda on this CUDA-13 wheel; the
+# transformer kernels remain CUDA FlashAttention/Triton when only this sampler is off.
+Environment=VLLM_USE_FLASHINFER_SAMPLER=0
+Environment=XDG_CACHE_HOME=%C/dirac-qwen
+Environment=TORCH_HOME=%C/dirac-qwen/torch
+Environment=VLLM_CACHE_ROOT=%C/dirac-qwen/vllm
+Environment=TRITON_CACHE_DIR=%C/dirac-qwen/triton
+Environment=FLASHINFER_WORKSPACE_BASE=%C/dirac-qwen
+CacheDirectory=dirac-qwen
+ExecStart=/home/ivan/.local/share/dirac-qwen-vllm/bin/vllm serve /mnt/e/data.ai-models.local-model-store.storage.xl.private.readonly/Qwen3.5-2B --host 127.0.0.1 --port 8930 --served-model-name Qwen3.5-2B --max-model-len 16384 --gpu-memory-utilization 0.65 --language-model-only
+Restart=on-failure
+RestartSec=5
+TimeoutStopSec=45
+NoNewPrivileges=true
+PrivateTmp=true
+ProtectSystem=strict
+ProtectHome=read-only
+ReadWritePaths=%C/dirac-qwen
+MemoryHigh=24G
+MemoryMax=32G
+
+[Install]
+WantedBy=default.target