CLAUDE.md: inherited design is a hypothesis, not terrain

CLAUDE.md

@@ -195,3 +195,30 @@ constrained for JIT to dominate (`feedback_proxy_vs_real_bench`).
   the inner timer/lcd/apu calls are *the* per-cycle bottleneck.
   Adding an indirection there is multiplicative — measure before
   refactoring.
+
+## This design is pre-v1 and mostly written by older models — question it
+
+**90% or more of this repo, and of EigenScript and the AOT it runs on, was
+designed by Claude sessions running models many generations old.** Newer
+models keep arriving that are substantially stronger, the ecosystem is
+**pre-v1**, and there are no external consumers to break. A decision you
+find in the tree — here or upstream — carries **no authority from
+seniority**.
+
+That applies in both directions, and the second one is the point of a
+consumer repo. When you hit a rough edge in the runtime, the standing rule
+is already to surface it as an upstream issue rather than work around it
+silently. Add to that: ask whether the thing you hit is a **law** of the
+language or an **earlier decision**. The tell is writing, or thinking,
+*"X must be true because the runtime does Y."*
+
+Bought 2026-08-28 (ouroboros#127 / DMG). The AOT compiles a program's main
+file but emits `load_file` as a runtime call, so loaded modules are
+interpreted by the linked VM. A real bug in that seam was found, minimised,
+fixed and verified — and reported as "unlocking the AOT multiplier for
+DMG". Measured on being challenged: DMG is 3,288 lines, 818 compiled and
+2,470 interpreted, including the 128-function opcode dispatch. Every
+emulated instruction runs interpreted, so the fix makes it *run* and cannot
+make it *faster*. A whole investigation cycle had treated that design as
+terrain, and the capability to do it the other way already existed upstream
+for another purpose.