document the leaf-deletion blind spot in check:sri and point to the settled-tree leaf-count signal

internal/sri-consistency-check.mjs

@@ -24,6 +24,18 @@
  *   safety advice created a different hazard. The fix is not to ban single-bundle builds —
  *   it is to make sure an inconsistent tree can never SHIP.
  *
+ * ★ WHAT THIS CHECK CANNOT SEE — a DELETED leaf (so never run it alone before a publish).
+ *   This script walks the leaves that EXIST and asserts each one's baked hash matches its
+ *   bundle. It asserts NOTHING about how many leaves there should be. A leaf that no longer
+ *   exists is simply not checked — the gate passes because there is nothing left to disagree.
+ *   So a tree gutted by a per-bundle `BUNDLE=x vite build` (which empties OTHER bundles' leaf
+ *   index.html) returns a confident PASS here while being unpublishable.
+ *   The leaf COUNT is settled-tree-gate.mjs's signal 4 (leaves >= MIN_LEAVES). The two checks
+ *   are complementary and the publish runbook runs BOTH:
+ *       bun run check:settled && bun run check:sri
+ *   Named by the grid-palette agent, 2026-09-20, after a five-sample poll watched leaves climb
+ *   0 -> 58 mid-build; a single sample of a moving tree cannot tell "destroyed" from "rebuilding".
+ *
  *   NOT the same thing as the 5-minute SRI CACHE trap documented in the publish runbook.
  *   That one is TRANSIENT (republishing the same bundle inside the 300s max-age serves a
  *   stale .js against a fresh integrity) and resolves itself. THIS one is PERMANENT —