mobile-debug-mcp 0.32.0 → 0.32.1

package/docs/benchmarks/ui-tree-performance/comparison.md

@@ -0,0 +1,41 @@
+# UI tree observation performance
+
+Fixture SHA-256: be322549a4ba154c0b1c85876559157bb21a911673906d270a7ec26dc2fcfa89. Android 16, API 36. Three warm-ups and ten measured attempts per flow and mode. Failed attempts are visible in the JSON reports and excluded from percentiles.
+
+| Mode | Flow | p50 ms | p95 ms | Failures |
+|---|---|---:|---:|---:|
+| baseline | static | 52402.254875 | 52973.044208 | 0 |
+| baseline | network_100ms | 5607.070875 | 5851.747375 | 0 |
+| baseline | network_1500ms | 6996.555542 | 7165.898416 | 0 |
+| candidate | static | 43823.508666 | 44747.601625 | 0 |
+| candidate | network_100ms | 4761.979625 | 4967.290833 | 0 |
+| candidate | network_1500ms | 6176.188083 | 6469.3265 | 0 |
+
+| Acceptance gate | Result |
+|---|---|
+| static_p50_improvement | pass |
+| static_p95_regression | pass |
+| network_100ms_p95_regression | pass |
+| network_1500ms_p95_regression | pass |
+| zero_measured_failures | pass |
+
+Raw per-attempt records include acquisition and consumer events, response-send timestamps, and assertion-read timestamps. Device serials and UI contents are not recorded.
+
+The candidate improves static p50 by 16.4% (52.40 s to 43.82 s). Both modes
+made 200 physical tree reads across their ten measured static attempts. The
+baseline recorded 200 approximately 400 ms first-attempt waits; the candidate
+recorded none. Separate assertion samples and same-step target consumers remain
+visible in the acquisition records. The candidate also lowers network p95 at
+both response delays. The fixture recorded no `VERIFICATION_UNAVAILABLE` in
+either mode, so this result does not establish its rate in other apps.
+
+`initial-baseline.json` preserves the standalone instrumentation-only capture
+made before the acquisition change. The paired `baseline.json`,
+`candidate.json`, and `comparison.json` use one APK SHA-256 and the same
+emulator settings. An earlier paired attempt is retained as
+`comparison-attempt-1.json`: one measured baseline 100 ms network read failed
+after a 181.97 s physical acquisition, and one candidate 1,500 ms warm-up
+failed before an observation event. That attempt failed the zero-failure gate.
+The unchanged repeat above had no failures. The cause of the first attempt's
+stalls was not established, so the repeat supports the latency result but does
+not prove the stall cannot recur. iOS latency was not measured.