feat: Qwen VL grounding endpoint on Modal (OpenAI-compatible, scale-to-zero)

[abrichr]

Summary

Stands up a self-hosted Qwen VL endpoint on Modal for openadapt-flow's OpenAICompatibleGrounder, as qwen_endpoint/:

• Model: Qwen/Qwen2.5-VL-7B-Instruct pinned to HF revision cc594898137f460bfe9f0759e9844b3ce807cfb5, served bf16 by vLLM 0.10.1.1 on one A10G (24 GB), --max-model-len 16384. Upgrade path to Qwen3-VL-8B (needs vllm>=0.11) documented in the runbook, gated on a verified deploy.
• Auth: every request needs the bearer token from the Modal secret qwen-endpoint-token (vLLM --api-key; unauthed => 401). The token is read only inside the serving container; the client launcher with_token.sh sources it from the macOS Keychain at runtime. No secret material anywhere in the repo.
• Scale-to-zero: scaledown_window=120, no container floor, no warm pool. Weights cached in a Modal volume, so only the first-ever cold start pays the ~16 GB download.
• Smoke harness: smoke_grounder.py drives flow's real OpenAICompatibleGrounder (not a re-implementation) against two committed flow fixtures (benchmark/dense_surface/record_seed1.png, replay_native_arial_seed1.png) and records hit/miss/abstain, latency, and per-request cost into RESULTS.md.
• Operator wiring: deployment.snippet.yaml — the flow runtime: block pointing the grounder at this endpoint (api_key_env reference, PHI allowlist entry).
• CI guard: tests/test_qwen_endpoint.py (stdlib-only, runs in docs CI) pins the invariants: secret name, <=120 s scaledown, no warm pool, exact model + vLLM pins, no long secret-like literals.

Not deployed yet — intentionally

Preflight found Modal auth present but no qwen-endpoint-token secret. Creating it is founder-only (one command, RUNBOOK.md section 0; registered as NEEDS_YOU B28). Once it exists, RUNBOOK sections 1-2 deploy and verify: 401 check, cold/warm latency, scale-to-zero check, and the 5-request grounder smoke — total GPU budget $5, expected ~$0.30. RESULTS.md holds the pre-registered protocol and blank tables.

The full grounding accuracy probe runs against Together in a sibling effort; this endpoint's smoke is wire-proof, not an accuracy claim.

Test plan

• uv run --extra dev pytest tests/ -q — 149 passed (incl. 8 new)
• uv run python scripts/validate_docs.py — clean
• Post-secret: RUNBOOK sections 1-2 (deploy, 401, latency, scale-to-zero, smoke) — blocked on NEEDS_YOU B28

🤖 Generated with Claude Code