[feedback] Model: a repeated instruction is treated as a reminder, not as a missing mechanism

Context. Solo-operator project using Claude Code (desktop app, Windows 11) with a lead session on Claude Fable 5.1 orchestrating Opus 5.5 / Sonnet subagents (in-process, worktree-isolated), a JSON task store (~1,100 records), PreToolUse/SubagentStop hooks and a harvest-train commit flow. Everything below was measured on 2026-09-28 in one session of ~20 dispatches, two blind judge pairs and nine commits. Numbers are from tool output, not recollection; no secrets or private paths included.

Finding. The operator restated one standing rule ('every lane chases multiple sub-goal closes') four times in a day. Each time the lead complied for the item at hand; only at the fourth did it generalise the rule into a mechanism (a mandatory brief field checked by a hook). Same shape as an earlier drift where a conditional safety rule was restated >100 times.

Reproduction. Give a standing rule; restate it after one violation; observe whether the second restatement yields a local fix or a mechanism (a template field, a lint, a hook).

Ask. The second restatement of an instruction should trigger 'what mechanism makes this unnecessary' by default: propose the wall, say so, then comply.