[feedback] Model: a document's headline gets mapped, its body's recommendations die; subagents don't know where a finding lives

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 lead turned one research report's recommendations into a decision and could not show the same for the other reports of the day. A read-only subagent sweep then found 3,257 recommendation-shaped items (recommendations, options, 'should' sentences, owed items) across four days of reports, ~575 with no tracked home. Judgement subagents produced 14-23 measured findings per pass but placed them inconsistently until the report format required a 'recommendations -> tracked home' table.

Reproduction. Give the model a 40-page report with a headline verdict and 30 inline recommendations; ask it to 'apply the report'; count how many of the 30 reach a tracked item.

Ask. Keep an explicit obligations list per document read and map every item before calling the document handled; make 'where does this live' part of a finding; support a report format whose seal line counts obligations mapped vs unmapped.