make bench-cloud-vs-local: timed head-to-head, local --burn vs cloud-ask

Both paths grind on the same data (LOCAL_DB defaults to web.db, the
file that was uploaded to the Spaces shard at SHARD_URL). --burn busts
the local QA cache so we measure a fresh inference both sides; the
cloud path has no caching layer so it's always fresh.

Reports answer + audit_mode + verifier_method + sources + wall time
for each path, then prints local-vs-cloud delta + ratio.

Sample run (Q='who developed virt-back?'):
  local: 3.8s  STRICT  (verbatim quote verified)
  cloud: 11.8s HYBRID  (entity-name verified)
  delta: +8.0s  (3.10x — the HTTP RANGE * ~99 requests tax)

Cloud audit_mode lands one rung lower because the cloud-ask prompt is
simpler than query.py's full pipeline, so the LLM paraphrases instead
of quoting. Verifier honestly demotes paraphrase to HYBRID. Both
answers are correct; the audit_mode difference is a prompt-shape
artifact, not a cloud-path correctness gap.

Makefile

@@ -1580,6 +1580,43 @@ cloud-ask: bootstrap ## bucket-direct end-to-end: FTS → LLM → verify [Q="...
 	    --max-context-chars $(or $(MAX_CONTEXT),24000) \
 	    --cache-mb $(or $(CACHE_MB),64)
 
+# ---------------------------------------------------------------------------
+# Head-to-head: local query (BURN=1) vs cloud-ask, same question, same data.
+# ---------------------------------------------------------------------------
+# Defaults align the two endpoints: LOCAL_DB is the same file that was
+# uploaded to Spaces under SHARD_URL, so the two paths grind on the
+# byte-identical corpus. --burn busts the local QA cache so we time
+# a fresh inference both sides.
+
+LOCAL_DB ?= $(HOME)/.arborist/crawl/web.db
+
+bench-cloud-vs-local: bootstrap ## head-to-head: local query --burn vs cloud-ask, same question, timed [Q="..."]
+	@test -n "$(Q)" || { echo 'usage: make bench-cloud-vs-local Q="your question" [LOCAL_DB=path] [SHARD_URL=https://...]'; exit 2; }
+	@echo "=== bench-cloud-vs-local ==="
+	@echo "Q:         $(Q)"
+	@echo "LOCAL_DB:  $(LOCAL_DB)"
+	@echo "SHARD_URL: $(SHARD_URL)"
+	@echo
+	@echo '--- LOCAL (--burn, fresh inference) ---'
+	@T0=$$(date +%s.%N); \
+	$(ARBORIST) --db $(LOCAL_DB) query --top-k 4 --burn --json '$(Q)' > /tmp/.bench-local.json 2>&1 || true; \
+	T1=$$(date +%s.%N); \
+	LOCAL_DT=$$(awk "BEGIN { print $$T1 - $$T0 }"); \
+	$(PY) -c "import json; d=json.load(open('/tmp/.bench-local.json')); print('answer:    ', d.get('answer_text','').strip()[:200]); print('audit:     ', d.get('audit_mode')); print('verifier:  ', d.get('verifier_method')); print('n_verified:', d.get('n_verified'), '/', d.get('n_quotes')); s=d.get('sources') or []; print('sources:   ', len(s), '→', s[0]['document_uri'] if s else '(none)')"; \
+	printf 'wall:       %.3fs\n' "$$LOCAL_DT"; \
+	echo; \
+	echo '--- CLOUD (bucket-direct, always fresh) ---'; \
+	T0=$$(date +%s.%N); \
+	$(ARBORIST) cloud ask '$(Q)' --shard-url "$(SHARD_URL)" --top-k 4 > /tmp/.bench-cloud.json 2>&1 || true; \
+	T1=$$(date +%s.%N); \
+	CLOUD_DT=$$(awk "BEGIN { print $$T1 - $$T0 }"); \
+	$(PY) -c "import json; d=json.load(open('/tmp/.bench-cloud.json')); print('answer:    ', d.get('answer_text','').strip()[:200]); print('audit:     ', d.get('audit_mode')); print('verifier:  ', d.get('verifier_method')); print('n_verified:', d.get('n_verified'), '/', d.get('n_quotes')); s=d.get('sources') or []; print('sources:   ', len(s), '→', s[0]['document_uri'] if s else '(none)'); st=d.get('stats') or {}; print('http_reqs: ', st.get('http_requests')); cs=st.get('cache') or {}; print('bytes:     ', cs.get('bytes_fetched'))"; \
+	printf 'wall:       %.3fs\n' "$$CLOUD_DT"; \
+	echo; \
+	echo "=== summary ==="; \
+	awk "BEGIN { printf \"  local: %.3fs\\n  cloud: %.3fs  (delta %+.3fs, ratio %.2fx)\\n\", $$LOCAL_DT, $$CLOUD_DT, $$CLOUD_DT-$$LOCAL_DT, $$CLOUD_DT/$$LOCAL_DT }"; \
+	rm -f /tmp/.bench-local.json /tmp/.bench-cloud.json
+
 CLOUD_DEMO_PORT ?= 18785
 
 cloud-demo: bootstrap ## bucket-direct end-to-end proof on tiny in-process corpus [CLOUD_DEMO_PORT=N]