Real-vault retrieval receipt

scripts/benchmark_retrieval_latency.py produces a content-free latency receipt. It records query-set hash, verified index identity, execution-shape timings, sanitized provider binding, peak RSS, optional NVIDIA device memory, and execution errors. It never writes query text, note content, snippets, paths, or the doctor report's vault path to the receipt.

Peak RSS is recorded per execution shape. NVIDIA telemetry is sampled only at benchmark run start/end so nvidia-smi latency cannot contaminate MCP or query timings; it is explicitly device-wide and may include unrelated processes.

The normal benchmark remains diagnostic. For the roadmap's real-vault dense evidence gate, require both an immutable generation and an actual provider session:

python scripts/benchmark_retrieval_latency.py \
  --vault /path/to/isolated-real-vault \
  --fixture /path/to/frozen-query-fixture \
  --modes semantic hybrid reranked \
  --rounds 5 \
  --cold-rounds 3 \
  --query-limit 10 \
  --probe-provider \
  --require-provider-binding \
  --require-immutable-generation \
  --output /tmp/power-real-vault-retrieval.json

The command exits non-zero if the generation is not verified immutable, the provider probe does not create a session with an active provider, or any execution shape records an error. gpu_memory_used_bytes is an optional device-wide nvidia-smi reading; null means that the host does not expose that telemetry and is not silently treated as zero.

This receipt is evidence infrastructure, not acceptance by itself. The real-vault gate still requires an exact source snapshot, enough repeated samples for p50/p95, cold/warm/process/MCP controls, resource readings, and content_free=true. Synthetic fixtures and CPU feasibility runs remain diagnostic only; they do not justify an ANN or reranker quality claim.