feat(signals): report research agent reads steering notes and fleet memory

What The signals report research agent — the sandbox-backed multi-turn stage that produces a report's findings, priority, actionability, title, summary, and suggested reviewers — should read the team's steering notes and the scout fleet's shared scratchpad before judging, and record which notes it applied. Today it reads no user preference at all. Meanwhile the inbox already converts user feedback into steering notes via three derived origins — report_dismissal, report_discussion, report_feedback (see products/signals/backend/scout_harness/tools/notes.py plus dismissal_notes.py, discussion_notes.py, feedback_notes.py) — that only scheduled scout runs ever read. Reading them at the research stage closes the steering loop users feel most: "I dismissed a report with a reason, and the next report respected it." It works for every signal source at once, because every source funnels through report generation. Why this is (almost) prompt-only • The scout me…