add i6/i7: i3/i4 with agent-view docs (64-token snippets, no docs_id tags; notes untagged, 128)

clean_note: drop forward-planning sentences from notes (keep relevance verdicts + gap statements); affects i4/i5

redesigned-format training data i0-i5 (docs/query_input_redesign.md): Main Question / Current Subquery / Previous Interactions, [docs_id:x] tags, <empty> placeholders, 128/128 truncation, full history

initial commit