feat(bench): serving benchmark — the baseline instrument for the migration

A scripted 8-turn conversation through the real pipeline (tool call,
memory recall, librarian route, small talk), run against any base URL so
the same instrument measures Ollama today and llama-server after the
cutover. Non-stream mode records wall time and usage; --stream records
time-to-first-token; a VRAM poller and an optional interleaved second
tenant capture the contention the serving plan calls out.

Planted facts carry a BENCH-MARKER prefix so anything the biographer
learns from a bench run stays recognizable in the memory stores.

First smoke against production: 9.7s mean per simple turn, and TTFT
equals wall time — the client sees nothing until the whole
steward-orchestrate-synthesize pipeline has finished. Time-to-first-word
for desklock is currently the full pipeline latency, which makes
streamed synthesis a first-class goal of the serving work.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Makefile

@@ -1,4 +1,4 @@
-.PHONY: help setup run test test-unit test-integration test-contracts lint typecheck clean
+.PHONY: help setup run test test-unit test-integration test-contracts bench-serving lint typecheck clean
 
 VENV := .venv
 PYTHON := $(VENV)/bin/python
@@ -50,6 +50,9 @@ test-integration: ## Run integration tests (needs Claude/Ollama)
 test-contracts: ## Wire-level contract tests against live service boundaries
 	$(PYTEST) tests/contracts -v --no-cov
 
+bench-serving: ## Serving benchmark vs a live backend (BENCH_ARGS="--base-url ... --stream")
+	$(PYTHON) scripts/bench_serving.py $(BENCH_ARGS)
+
 lint: ## Run ruff linter and formatter check
 	$(RUFF) check src tests
 	$(RUFF) format --check src tests