feat(deploy): DGX standalone compose (ai-server only, 24855) + confirmed-topology runbook

.dockerignore

@@ -57,10 +57,19 @@ build
 # Large source materials (kept locally, never shipped)
 references
 
-# Docs — not needed at runtime
+# Docs — not needed at runtime, EXCEPT the versioned agent prompts, which the
+# ai-server image bakes in (apps/ai-server/Dockerfile: COPY docs/ai/prompts
+# /app/prompts + ENV PROMPTS_BASE_DIR). Without these negations the COPY fails
+# at build time ("not found" — the whole docs tree would be absent from the
+# build context).
 docs
+!docs/ai/prompts
+!docs/ai/prompts/**
 README.md
 
+# Archived legacy materials — segregated for traceability, never shipped
+_archive
+
 # Local data
 *.sqlite
 *.db

docs/ai/deployment_integration_plan.md

@@ -29,12 +29,37 @@
 
 Deferred (P2 backlog): idempotency guard, request_id propagation (REV-001 majors), narrative activation (gated), FHIR $validate external option.
 
-## 3. Deployment runbook (Docker)
+## 3. Deployment runbook (Docker) — DGX topology confirmed 2026-07-20
 
-1. Local: `docker compose -f infra/deploy/docker-compose.yml build ai-server` → container smoke (boot, /health, route smoke).
-2. Transfer: `docker save` → `ssh DGX docker load` (or private registry).
-3. AI server: place `.env` (operator), `docker compose up -d`; in-compose DB via `postgres:5432`, external DB via 28881 if split later — env-DSN only.
-4. Verify: `/health` 200, DB preflight, prompt SHA pins, one F1-turn + one F5-report smoke.
+Confirmed topology (user, 2026-07-20): one DGX Spark host at `223.194.33.26` runs the DB
+server and the AI server as separate containers with distinct external ports.
+
+| Service | External | Internal | State |
+|---|---|---|---|
+| DB server (postgres, admin-managed container) | `223.194.33.26:28881` | 5432 | running |
+| ai-server (this deployment) | `223.194.33.26:24855` | 8001 | to deploy |
+
+The dev workstation (`192.168.68.62`) has no docker and needs none — deployment executes on
+the DGX. Dedicated standalone compose file: `infra/deploy/docker-compose.dgx.yml`
+(ai-server only; defines no postgres/api, distinct project+container name, production CMD,
+no source mounts, `PROMPTS_BASE_DIR` pinned to the baked `/app/prompts` so the dev `.env`'s
+workstation-relative value cannot leak in via env_file). Never combine it with the dev
+`docker-compose.yml`.
+
+On the DGX, as the account that runs the DB container:
+
+```bash
+git clone https://github.com/Neuro-AI-Lab/neurosync.git && cd neurosync   # or git pull
+# place apps/ai-server/.env (operator copies from the dev workstation; never committed)
+docker compose -f infra/deploy/docker-compose.dgx.yml up -d --build
+curl -s localhost:24855/health
+```
+
+Verify: `/health` 200 → route smoke (`POST /ai/survey/plan`, `/ai/temporal/analyze`,
+`/ai/handoff/report`) → DB preflight. If the container cannot reach the DB through the
+host's external IP (hairpin), change the DSN host in `.env` to
+`host.docker.internal:28881` (extra_hosts maps it) — env edit only.
+Backend base URL: `http://223.194.33.26:24855`.
 
 ## 4. Gates
 qa (CI-mirror + route contract tests + container build/boot smoke) → clinical-validator quick pass on /ai/survey/plan semantics (safety-net/SI-supplement decisions must match the validated behavior) → critic wording/regression check. Local commits only; publish on user's word.

infra/deploy/docker-compose.dgx.yml

@@ -0,0 +1,51 @@
+# DGX Spark production deployment — ai-server ONLY.
+#
+# ⚠️  STANDALONE file, NOT an override of docker-compose.yml — run with ONLY
+#     `-f infra/deploy/docker-compose.dgx.yml`. Never combine the two files.
+#
+#   - The DB server already runs on this DGX as a separately managed container
+#     (external 28881 → postgres 5432). This file deliberately defines NO
+#     postgres and NO api service — never start a second DB from here.
+#   - Distinct compose project name + container name → zero state collision
+#     with any compose project or container already on the DGX.
+#   - External port 24855 is the allocated ai-server port. Backend base URL:
+#     http://223.194.33.26:24855  (container listens on 8001, Dockerfile CMD).
+#   - All secrets come from apps/ai-server/.env via env_file at RUN time; the
+#     image never contains .env (root .dockerignore excludes it). The operator
+#     places .env on the DGX before `up`.
+#
+# Run from repo root on the DGX:
+#   docker compose -f infra/deploy/docker-compose.dgx.yml up -d --build
+#   curl -s localhost:24855/health
+#
+# DB reachability: DATABASE_URL in .env points at 223.194.33.26:28881. If the
+# container cannot hairpin through the host's external IP, change only the DSN
+# host to host.docker.internal:28881 (mapped via extra_hosts below) — an .env
+# edit, no code change.
+
+name: neuro-sync-ai-dgx
+
+services:
+  ai-server:
+    build:
+      context: ../..
+      dockerfile: apps/ai-server/Dockerfile
+    container_name: ns-ai-server-dgx
+    # No `command:` override and no src/tests volume mounts — the baked image
+    # is the deployment artifact; the Dockerfile's production CMD runs
+    # (no --reload).
+    env_file:
+      - ../../apps/ai-server/.env
+    environment:
+      LOG_LEVEL: info
+      # `environment:` outranks env_file in compose precedence. The dev .env
+      # carries the workstation-relative PROMPTS_BASE_DIR=docs/ai/prompts,
+      # which does not exist inside the container — pin the baked-in copy
+      # (Dockerfile: COPY docs/ai/prompts /app/prompts) so env_file injection
+      # can never break prompt loading.
+      PROMPTS_BASE_DIR: /app/prompts
+    ports:
+      - "24855:8001"
+    extra_hosts:
+      - "host.docker.internal:host-gateway"
+    restart: unless-stopped