PMAT-3797: CRUX serve verb — apr serve, llama-server and ollama through ONE OpenAI client, streaming + non-streaming

[noahgift]

CRUX slice 4: the serve verb

Closes #3797. Stacked on #3796 (→ #3789 → #3767), so this diff is crux-3739-chat...HEAD.

keep-open: #3739 stays open. TTFT and decode rate via apr test llm bench, max context, code, thinking ON, consumer rungs and T-1 wiring are later increments; all are named in every receipt's not_covered.

What it adds

• --verbs run,serve runs one serve cell per model under one hold of the GPU lock.
  • apr serve run and llama-server each load the model once, next to the running ollama daemon.
  • Every golden prompt goes through scripts/lib/crux_openai_client.py, non-streaming and streaming.
  • Then every server stops, and ollama unloads (keep_alive 0, plus the unload wait).
• crux_openai_client.py is the one OpenAI client: the same body for every server, and the row-contract JSON out. It reads no clock and computes no rate; TTFT and throughput are apr test llm bench's job (PERF-009).
• The judge: serve rows are read through the client JSON. The key gains an additive mode (nonstream | stream), and apr serve's cells carry backend_verified: false.
• Correspondence (per aprender-97, pv SHACL release-readiness shape: the tag is REFUSED unless every (Q4_K model in host inventory × {lambda,gx10} × {run,chat,serve,code} × consumer-context) cell has a Pass receipt — a missing cell is a violation, never a skip #3715): the serve run entry declares modes: [nonstream, stream], with per-engine modes, so pv never names a verb.

Measured on lambda

cpu lane, qwen2.5-coder-1.5b Q4_K_M, --verbs run,serve: PASS, 9 cells, 9 GREEN. apr serve, llama-server@d1d3c3396 and ollama 0.33.2 each answered every golden prompt in both modes, and nothing was left running.

• The first run found a teardown bug. Every request was answered, but the cell hung: a bare wait after killing apr serve blocked on llama-server. Teardown now kills every server first and waits on those PIDs only. The rerun finished unaided.
• apr serve renders a different prompt. It answers 2+2 from an 18-token prompt with "4", where llama-server and ollama send 36 tokens (the GGUF template's default system prompt) and answer "2 + 2 equals 4.". This is recorded, not judged here; the byte-equal tmpl rows against HF will judge it.

Verified

• bash scripts/check_crux_inference_judge.sh: 44 ok, 0 broke. Serve rows: stream and non-stream are distinct GREEN cells; apr serve answering nothing (connection refused) while llama-server answers is RED. Two serve mutants each break a row.
• bash scripts/guard_tree.sh --no-cargo: 77 checks, 0 failed on fa8623a5d. check_no_competing_harnesses.sh: count=0. The verb name serve run and my first rate-field names had tripped its text test, so the fields are now decode_rate with a separate rate_unit.

🤖 Generated with Claude Code

[noahgift]

Every commit is already in release/0.69.1-batch-2 (git cherry: 0 unfolded). It reaches main through the release batch PR #3835. Closing as folded, not rejected.