Declare Qwen-3.5-27B model so its benchmarks ingest + display

[aistackdev]

Problem

Benchmarks for the dense Qwen3.5-27B (model-prefix qwen3.5-27b, vLLM + SGLang on 1× H100) force-ingested "successfully" but every row was dropped. The ingest log:

=== ingesting …/runs/26717611277 ===
  bmk_qwen3.5-27b_1k1k_fp8_vllm_…   (all 18 artifacts read)
  Skipped: 18 rows
    qwen3.5-27b

resolveModelKey returned null because qwen3.5-27b isn't a known model key. It's a new prefix, deliberately kept distinct from the 397B qwen3.5 so the two don't conflate on the dashboard — but it was never declared. (gemma4 ingested fine because gemma4 is declared.)

Fix — declare the model in the three registries

• constants/models.ts DB_MODEL_TO_DISPLAY: qwen3.5-27b → Qwen-3.5-27B. Single source of truth — DB_MODEL_KEYS and DISPLAY_MODEL_TO_DB derive from it, so the ETL now resolves the prefix instead of skipping.
• app/src/lib/data-mappings.ts: new Model.Qwen3_5_27B enum member + MODEL_CONFIG entry (label: 'Qwen3.5 27B', prefix: 'qwen3.5-27b') for frontend display/labeling.
• db/src/etl/normalizers.ts MODEL_TO_KEY: map Qwen/Qwen3.5-27B[-FP8] → qwen3.5-27b as a fallback when the prefix field is absent (e.g. eval artifacts).

Notes

• No migration: configs.model is free-form text with a lowercase CHECK; qwen3.5-27b already satisfies it.
• Distinct display name (Qwen-3.5-27B, separate from the 397B Qwen-3.5-397B-A17B) → its own dashboard bucket, not point-release grouped.
• Only exhaustive Record<Model,…> is MODEL_CONFIG (updated); MODEL_ARCHITECTURES is Partial<…>. Local tsc --noEmit passes.

After merge + deploy

Re-ingest the two skipped runs so their rows land:

• vLLM: …/InferenceX/actions/runs/26717611277
• SGLang: …/InferenceX/actions/runs/26717831054

gh workflow run auto-ingest.yml -R vngcloud/InferenceX-app -f run_url=<run-url>   # one at a time

🤖 Generated with Claude Code