chore(vllm): retire LFM2.5-2.6B permanently; audit finds nevermore on the harmful reranker

Operator directive: lfm2.5-2.6b goes down permanently.

  - stacks/vllm/compose.yaml   vllm-lfm25 service removed (replaced by a
                               tombstone comment), pushed live to ana-ml2
  - ana-ml2                    container docker rm -f'd, 8,721 MiB freed on GPU1
                               (95,388 -> 86,667 of 97,887)
  - litellm config             lfm2.5-2.6b alias deleted, live + canonical,
                               28 -> 27 models

It was an EVAL-ONLY bake-off seat against granite-4.1-8b that never received
the operator ruling it was pending; the comparator was retired from the roster
on 2026-08-15; it was deliberately never wired into any default or fallback
routing chain; and spend logs show 0 calls in the 4-day window to 2026-08-21.
Weights stay in the shared HF cache -- nothing deleted from disk.

The gateway restart that makes the alias deletion take effect is HELD so it can
batch with a pending reranker change. Until then the name is still routable
in-memory and will error against a dead backend.

Auditing the three reranker seats while answering "why do we have three" turned
up a real problem. The design is one production, one rollback, one fallback --
but the traffic is backwards:

  :8013 A3 bge-v2-m3      PRODUCTION, backs `reranker`     0 calls / 4 days
  :8002 Qwen3-Reranker    RETIRED incumbent, rollback only 7 calls, 12-hourly
  :8014 A4 gte-modernbert "fallback"                       no alias at all

nevermore is hard-wired to the incumbent by name (NEVERMORE_RERANK_MODEL=
qwen3-reranker), so the R43 cutover never moved it -- the cutover repointed the
`reranker` alias and correctly left `qwen3-reranker` naming the Qwen model.
Brokkr R43 measured that model harming 80/90 fleet queries, so nevermore's
twice-daily rerank pass is likely degrading its own briefing.

Fix is one line in nevermore's .env plus a nevermore restart, and it must land
before :8002 is retired. Recorded in persistent-memory with the A4 alias also
noted as absent (global CLAUDE.md names reranker-a4-gte-modernbert; it does not
exist).

stacks/vllm/compose.yaml

@@ -342,78 +342,15 @@ services:
       - homepage.description=Qwen2.5-Coder-1.5B FIM code-completion (ana-ml2, Zed edit-predictions)
       - homepage.href=http://10.250.50.54:${CODER_PORT}/docs
 
-  vllm-lfm25:
-    image: vllm/vllm-openai:${VLLM_VERSION}
-    container_name: vllm-lfm25
-    restart: unless-stopped
-    ipc: host
-    ports:
-      - "${LFM25_PORT}:8000"
-    volumes:
-      - /tank/aimodels/huggingface:/hfcache
-    environment:
-      - HF_HOME=/hfcache
-      - HF_HUB_CACHE=/hfcache/hub
-      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
-      - VLLM_API_KEY=${API_KEY:-}
-    command:
-      # LiquidAI/LFM2.5-2.6B (BF16, Lfm2ForCausalLM). NON-PRODUCTION bake-off alias
-      # vs granite-4.1-8b on structured extraction / classification / tool-call
-      # formatting (brokkr R-target, 2026-08-10). LFM Open License v1.0 (<USD 10M-rev
-      # commercial, not OSI) — EVAL-ONLY pending an operator ruling on production; NOT
-      # wired into any default/fallback routing chain. Reasoning model: the chat
-      # template opens <think>; deliberately NO --reasoning-parser, so the full
-      # generation (thinking + answer) lands in `content` non-empty — brokkr's explicit
-      # requirement (an empty content with the answer stranded in reasoning_content
-      # reads as a degenerate model). Vendor sampling (temp 0.1 / top_k 50 / rep_pen 1.1)
-      # is the LiteLLM alias default, not a launch arg.
-      - ${LFM25_MODEL}
-      - --served-model-name
-      - ${LFM25_SERVED_NAME}
-      - --host
-      - 0.0.0.0
-      - --port
-      - "8000"
-      - --gpu-memory-utilization
-      - ${LFM25_GPU_MEM_UTIL}
-      - --max-model-len
-      - ${LFM25_MAX_MODEL_LEN}
-      - --max-num-seqs
-      - ${LFM25_MAX_NUM_SEQS}
-      - --dtype
-      - auto
-      - --kv-cache-dtype
-      - ${LFM25_KV_CACHE_DTYPE}
-      - --enable-prefix-caching
-      # Split the <think> reasoning (delimited by </think>; the opening tag is
-      # injected by the chat template into the prompt) into reasoning_content, so
-      # `content` is the clean post-</think> answer — scoreable JSON for brokkr's
-      # structured-extraction bake-off (raw-served, reasoning prepended, is not).
-      - --reasoning-parser
-      - deepseek_r1
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              device_ids:
-                - "${LFM25_GPU_ID}"
-              capabilities:
-                - gpu
-    healthcheck:
-      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
-      interval: 30s
-      timeout: 10s
-      retries: 3
-      start_period: 300s
-    networks:
-      - tnet
-    labels:
-      - homepage.group=AI - Inference
-      - homepage.name=vLLM LFM2.5-2.6B (bake-off)
-      - homepage.icon=mdi-flask-outline
-      - homepage.description=LiquidAI LFM2.5-2.6B non-prod bake-off vs granite (ana-ml2)
-      - homepage.href=http://10.250.50.54:${LFM25_PORT}/docs
+  # vllm-lfm25 (LiquidAI/LFM2.5-2.6B, :8021) — RETIRED PERMANENTLY 2026-08-20 by
+  # operator directive. It was an EVAL-ONLY bake-off seat against granite-4.1-8b
+  # (brokkr R-target, 2026-08-10) that never received the operator ruling it was
+  # pending. Its comparator is gone (granite retired from the roster 2026-08-15),
+  # it was deliberately never wired into any default/fallback routing chain, and
+  # LiteLLM spend logs showed 0 calls in the 4-day window ending 2026-08-21.
+  # Freed 8,772 MiB on ana-ml2 GPU1. The `lfm2.5-2.6b` gateway alias was removed
+  # in the same pass so the name 404s cleanly rather than erroring against a dead
+  # backend. Weights remain in the shared HF cache; nothing was deleted from disk.
 
 networks:
   tnet: