diff --git a/docs/agent-handoff.md b/docs/agent-handoff.md
index cf9d3d5..2b3e3c8 100644
--- a/docs/agent-handoff.md
+++ b/docs/agent-handoff.md
@@ -2079,16 +2079,20 @@ before trusting them:
   `m6_8_supervisor_reaps_all_children_across_cycles`) are timing-based;
   `editor::composition_overhead_under_ten_percent` is a render-ratio
   microbenchmark that fails ~1/3 even isolated single-threaded (already
-  `cfg!(macos)`-disabled). *(One local measurement, 2026-08-04, 16-core
-  Linux, tree-primitive branch: it went red inside the full `--features
-  crdt` lib run at `dispatch overhead 30.7%`, then 5/5 green run alone
-  at ratios of -0.6% to +0.2%, and the next full-suite run was also
-  green. Recorded as a measurement, NOT a cause — 5 isolated greens
-  establish that the ratio is far from the threshold when alone, not
-  that in-suite contention is what pushed it over. It is also not a
-  `ci-red-signatures.md` row: that registry judges red **CI** runs, and
-  this was local. The branch could not reach it — its diff versus `main`
-  touched no `src/`, no crate, and no manifest.)* Vterm Stage 3's merge CI saw one macOS timeout in
+  `cfg!(macos)`-disabled). *(Local measurements, 2026-08-05, 16-core
+  Linux, tree-primitive branch, which could not reach it — its diff
+  versus `main` touches no `src/`, no crate, no manifest. **Two reds,
+  both inside the full `--features crdt` lib run**, at `dispatch
+  overhead` 30.7% and 10.3%; two full-suite runs at the same tips were
+  green; and **10/10 green run alone**, ratios spanning -2.3% to
+  +1.5%. So the in-suite/isolated split has now been seen twice, and
+  the second red cleared the 10% budget by 0.3 points — the threshold
+  is marginal, not comfortably clear. Recorded as measurements, NOT a
+  cause: ten isolated greens establish that the ratio sits nowhere near
+  the threshold when alone, not that in-suite contention is what pushed
+  it over, and two reds against two greens in-suite is intermittence
+  rather than a mechanism. Not a `ci-red-signatures.md` row either —
+  that registry judges red **CI** runs, and these were local.)* Vterm Stage 3's merge CI saw one macOS timeout in
   `real_tui_terminal_smoke_restores_host_after_output_input_resize_scroll_copy_and_bell`;
   the complete failed-job rerun passed. The required-GPU gate also failed once
   in `headless_diag_face_recolors_band_counter_despite_unchanged_text`, then