diff --git a/docs/plans/2026-09-26_lead-decisions.md b/docs/plans/2026-09-26_lead-decisions.md
index c62c69d09..2a32cf771 100644
--- a/docs/plans/2026-09-26_lead-decisions.md
+++ b/docs/plans/2026-09-26_lead-decisions.md
@@ -2838,3 +2838,22 @@ which stay with him. Each item names who decided it and what happened.
       owned. The harness suite was already in the gate.
     - The brief is re-pinned to the edited section, with the old section
       byte-identical at 51085ddd and HEAD before the edit.
+
+127. **Row 90 round 2 adds row 92's sandbox file (2026-10-12).** Decision
+    125's rebase changed two files. Dewey's gate on the rebase failed cli
+    (8) and seat (4): row 92 (57948a81) landed between the two bases, and
+    its bwrap sandbox binds only `RUNNER_FILES`. Row 90's `broker.mjs`
+    imports `packages/tasks/src/requirements.mjs`, which the runner then
+    loads through `client.mjs`, so every sandboxed role session would have
+    died at the next bus restart. No row 90 path changed, so the path
+    drift check couldn't see it.
+    - Round 2 is 21 paths. `packages/seat/src/sandbox.mjs` gains the one
+      `RUNNER_FILES` line. `requirements.json` stays out of the sandbox:
+      only `loadRequirements()` reads it, from the Broker constructor and
+      `verbs(ctx)`, and the runner calls neither. The seat test still
+      holds the list equal to the import walk.
+    - The full gate on a rebased candidate is what caught this, so it
+      stays required. A path drift check is not enough when a row in
+      between changed what a file's imports are allowed to load.
+    - Harness runs on the 21-path candidate in Filbert's review and in
+      Sage's landing gate, not in a separate Dewey slot.