Measure genuine CPU query startup first input and warm latency

docs/QUERY-CONTEXT.md

@@ -93,3 +93,32 @@ and six paired measurements also passed; see
 effectively unchanged in those samples. Native macOS and actual GPU context
 allocation/parity, pressure behavior, long code queries, representative query
 latency and updated installed desktop packaging remain release checks.
+
+## Current Windows CPU worker timing, 2026-10-13
+
+The normal current worker (`c01cb3ec13cfca722ee768e8350e1399a8b9a4d14a24a35adc3473a5cfc82a73`)
+was measured with CPU explicitly selected and one inference thread. Three paired
+indexing/query processes alternate order; each retains a separate first request
+and twenty distinct warmed inputs with native debug timings. Real 768-dimensional
+vectors were checked. Output preservation now refuses existing reports.
+
+| Query-role repeat | Process startup ms | First inference/IPC ms | Warm p95 ms (20 samples) |
+| --- | ---: | ---: | ---: |
+| 0 | 729.27 | 100.02 | 71.30 |
+| 1 | 662.83 | 106.18 | 78.38 |
+| 2 | 704.21 | 100.84 | 71.75 |
+
+One-time text-model size/SHA verification took 192.45 ms and is separate from
+worker startup. Indexing-role startup ranged 624.49–779.14 ms, first query
+101.64–110.53 ms, warm p95 73.03–80.47 ms. These are cached-file native-worker
+measurements, not cold disk, Rust query-cache, ANN, hybrid retrieval, actual UI
+first input or paint. The installed desktop/query worker remained present;
+background activity was not controlled, so this is not an idle reference-host
+acceptance measurement. All test inference used CPU; no GPU provider was loaded.
+
+Full ignored report: `.searchi/qa/query-context-warm-current.json`, SHA-256
+`ea7dd0b72eeb1dbd79f06fce92e0475dfdc321d7781e5d06d58fe12755faf0fd`.
+The query-role memory observation after first inference was 239,878,144 bytes
+private commitment and 274,141,184 bytes resident working set for repeat zero;
+these are different metrics and must not be summed. Native desktop first-input
+and complete search latency remain separate open checks.

tools/embeddinggemma2/query_context_benchmark.py

@@ -8,6 +8,7 @@ import argparse
 import ctypes
 import hashlib
 import json
+import math
 import os
 from pathlib import Path
 import platform
@@ -53,6 +54,8 @@ def main():
     parser.add_argument("--output", type=Path, required=True)
     parser.add_argument("--repeats", type=int, default=3)
     args = parser.parse_args()
+    if args.output.exists():
+        parser.error("output must be new; preserve previous evidence")
     if not 1 <= args.repeats <= 20:
         parser.error("repeats must be 1..20")
     manifest = json.loads((ROOT / "tools/embeddinggemma2/model-manifest.json").read_text())
@@ -77,6 +80,15 @@ def main():
                 response, first_ms = session.worker({"op": "query", "text": "cancel the supplier partnership", "debug_performance": True})
                 vectors(response, 1)
                 sample.update(first_query_ms=first_ms, worker=response["performance"], after_query_memory=memory(session.process.pid))
+                warm = []
+                # Distinct inputs exercise native query inference without a Rust
+                # query cache. Keep first input separate from these warmed calls.
+                for index in range(20):
+                    response, duration = session.worker({"op": "query", "text": f"supplier cancellation notice example {index}", "debug_performance": True})
+                    vectors(response, 1)
+                    warm.append({"round_trip_ms": duration, "worker": response["performance"]})
+                ordered = sorted(item["round_trip_ms"] for item in warm)
+                sample["warm_queries"] = {"samples": warm, "p95_ms": ordered[math.ceil(len(ordered) * .95) - 1]}
                 if role == "query":
                     session.send({"op": "documents", "texts": ["test"], "title": "none"})
                     rejected = session.next()
@@ -88,7 +100,8 @@ def main():
                 session.close()
     report = {"platform": platform.platform(), "cpu_count": os.cpu_count(), "verification_ms": verification_ms, "cache_note": "Cached model bytes; startup excludes Rust verification and desktop preparation. No GPU executed. Repeated fresh processes, not cold disk.", "worker_sha256": hashlib.sha256(args.worker.read_bytes()).hexdigest(), "samples": samples}
     args.output.parent.mkdir(parents=True, exist_ok=True)
-    args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    with args.output.open("x", encoding="utf-8") as output:
+        output.write(json.dumps(report, indent=2) + "\n")
     for role in ["indexing", "query"]:
         rows = [s for s in samples if s["role"] == role]
         print(role, "startup median ms", statistics.median(s["startup_ms"] for s in rows), "ready memory", rows[0]["ready_memory"], "after query", rows[0]["after_query_memory"])