voxtral + kyutai-tts: fix wrong image tag / wrong endpoint paths; fish-s2: env-selectable model variant

Three fixes from the second-wave deploy attempts:

* voxtral: vllm/vllm-omni doesn't publish a `latest` tag — pull
  failed with "manifest unknown". Pinned VOXTRAL_VLLM_TAG to v0.18.0
  (released 2026-03-29, the day after the Voxtral 4B TTS release —
  first cut with Voxtral support).

* kyutai-tts: NillPointer wrapper exposes ONLY /health (root) and
  POST /v1/audio/speech. No /v1/models, no /v1/audio/voices —
  those return 404. Verified by /openapi.json against the live
  container. Compose healthcheck + playbook wait + verify steps
  all repointed at the actual paths. POST /v1/audio/speech is now
  smoke-tested with a RIFF WAV assertion (same pattern as fish-s2).

* fish-s2: added FISH_S2_MODEL env var so the model variant is
  swappable via .env without rebuilding. Both s2-pro (default) and
  s1-mini are pre-pulled into the bind-mount; LLAMA_CHECKPOINT_PATH
  + DECODER_CHECKPOINT_PATH now use ${FISH_S2_MODEL:-s2-pro}.
  s1-mini was originally gated on fishaudio's HF org (401), but
  niobures/OpenAudio-S1 mirrors the same files openly — pulled
  from there via a one-shot snapshot_download.

playbooks/deploy-kyutai-tts.yaml

@@ -73,26 +73,32 @@ steps:
   - name: docker compose up -d
     shell: cd {{ compose_dir }} && docker compose up -d
 
-  - name: Wait for /v1/models to respond (allow ~10 min for first download + warmup)
+  - name: Wait for /health to respond (allow ~10 min for first download + warmup)
     shell: |
       for i in $(seq 1 120); do
-        curl -sf -o /dev/null --max-time 3 http://localhost:{{ host_port }}/v1/models && exit 0
+        curl -sf -o /dev/null --max-time 3 http://localhost:{{ host_port }}/health && exit 0
         sleep 5
       done
       exit 1
     changed_when: "false"
 
 verify:
-  - name: /v1/models returns valid JSON
-    shell: |
-      curl -sf http://localhost:{{ host_port }}/v1/models \
-        | python3 -c "import json,sys; json.load(sys.stdin)"
+  - name: /health returns 200
+    shell: curl -sf -o /dev/null http://localhost:{{ host_port }}/health
     changed_when: "false"
 
-  - name: /v1/audio/voices returns valid JSON
+  - name: /v1/audio/speech returns a real WAV (POST with text body)
+    # NillPointer wrapper exposes ONLY /health and POST /v1/audio/speech
+    # (no /v1/models, no /v1/audio/voices). Smoke by POSTing and
+    # asserting a RIFF WAV comes back.
     shell: |
-      curl -sf http://localhost:{{ host_port }}/v1/audio/voices \
-        | python3 -c "import json,sys; json.load(sys.stdin)"
+      out=$(mktemp --suffix=.wav)
+      curl -sf -X POST http://localhost:{{ host_port }}/v1/audio/speech \
+        -H 'Content-Type: application/json' \
+        -d '{"model":"tts-1.6b-en_fr","input":"Verify."}' \
+        -o "$out" --max-time 30
+      file -b "$out" | grep -q '^RIFF.*WAVE'
+      rm -f "$out"
     changed_when: "false"
 
   - name: Container is running

stacks/fish-s2/.env.example

@@ -34,6 +34,15 @@ FISH_S2_GPU_DEVICES=1
 # checkpoint.
 FISH_S2_COMPILE=1
 
+# Model variant. Both checkpoints are pre-pulled by the deploy
+# playbook into /worktank/fish-s2/checkpoints/. Swap by editing this
+# value + `docker compose up -d --force-recreate` (no rebuild needed).
+#   s2-pro   — 4B-class, highest quality, ~7-8 s TTFB on a long phrase
+#   s1-mini  — lighter, ~3-5× smaller model.pth, expected ~2-3 s TTFB
+# fishaudio/s1-mini on HF is gated, but niobures/OpenAudio-S1 mirrors
+# the same files openly — playbook pulls from there.
+FISH_S2_MODEL=s2-pro
+
 # ── persistent storage on the host ───────────────────────────────────
 # Model checkpoints — Fish auto-downloads s2-pro on first run (~9 GB
 # at BF16) and caches under here. Persistent across container

stacks/fish-s2/compose.yaml

@@ -63,6 +63,13 @@ services:
       # checkpoint.
       - COMPILE=${FISH_S2_COMPILE:-1}
       - API_PORT=8080
+      # Model selection. Override Fish's Dockerfile defaults so we can
+      # swap variants via .env without rebuilding. Both checkpoints are
+      # pre-pulled by the deploy playbook into bind-mounted checkpoints/.
+      #   s2-pro   — 4B class, ~17 GB VRAM, slow but highest quality
+      #   s1-mini  — lighter, ~3-5× smaller model.pth, much faster
+      - LLAMA_CHECKPOINT_PATH=checkpoints/${FISH_S2_MODEL:-s2-pro}
+      - DECODER_CHECKPOINT_PATH=checkpoints/${FISH_S2_MODEL:-s2-pro}/codec.pth
       # Hugging Face cache for model weights — first start pulls
       # fishaudio/s2-pro (~9 GB BF16) into this dir.
       - HF_HOME=/app/hf_cache

stacks/kyutai-tts/compose.yaml

@@ -39,10 +39,12 @@ services:
       - ${KYUTAI_TTS_CACHE_DIR}:/app/hf_cache
       - ${KYUTAI_TTS_VOICES_DIR}:/app/voices:ro
     healthcheck:
-      # The wrapper exposes /v1/models for OpenAI-compat — same shape
-      # as Voxtral / Qwen3-TTS. Use that as the readiness signal.
-      # 127.0.0.1 explicit to dodge IPv4/IPv6 localhost race.
-      test: ["CMD-SHELL", "python3 -c \"import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/v1/models', timeout=5).status==200 else 1)\""]
+      # NillPointer wrapper's actual endpoints (verified at runtime via
+      # /openapi.json): just /health (root) and /v1/audio/speech (POST).
+      # No /v1/models, no /v1/audio/voices — those 404. /health is the
+      # liveness signal. 127.0.0.1 explicit to dodge IPv4/IPv6 localhost
+      # race.
+      test: ["CMD-SHELL", "python3 -c \"import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5).status==200 else 1)\""]
       interval: 30s
       timeout: 10s
       retries: 3

stacks/voxtral/.env.example

@@ -3,9 +3,11 @@
 
 # ── image pin ────────────────────────────────────────────────────────
 # vLLM-Omni image tag (Mistral's partner serving stack for Voxtral).
-# Use a specific version rather than `latest` — vLLM moves fast and
-# Voxtral has version-specific compatibility.
-VOXTRAL_VLLM_TAG=latest
+# Pin a specific version — vllm/vllm-omni does NOT publish `latest`;
+# `:latest` 404s with "manifest unknown". v0.18.0 was released
+# 2026-03-29, one day after the Voxtral 4B TTS release, and is the
+# first vLLM-Omni cut with Voxtral support.
+VOXTRAL_VLLM_TAG=v0.18.0
 
 # Voxtral model on Hugging Face. The 4B variant is the only released
 # checkpoint as of 2026-04. Default BF16 weights are ~8 GB.