Memory dream, second pass: the notes themselves

The first dream rewrote the index and left every note on disk. It queued 35 facts the store does not hold and found three notes that contradict each other. This plan is that queue, written down so it does not evaporate with the session that found it.

Where the queue came from

Three analyzers read all 99 notes and the 56 run reports on 2026-11-09. Their proposals are in the dream run report, artifacts/runs/2026-11-09-memory-dream.md, and the full organizer accounting is in the run directory's organizer-report.md. The index that came out of that pass is committed in agent-config, and bun run dream check-index proves it still covers every note.

The three contradictions, worst first

The context window. session-context-measured.md says 835k in its description while its body says 1M and that 835k was one session. Every session budgets against this number, so a wrong one costs the most.
The answer channel. answer-channel-drops-answers.md and answer-channel-p2-p3.md describe a channel that loses answers, while p5-answer-on-finish.md says the Record button was deleted and finishing the card writes the answer. No proposal shows an answer reaching plans/decisions/answers.jsonl, so one of the two is still unverified.
DRAFT status. folding-a-note.md records the 2026-08-14 call that DRAFT gates nothing; the 2026-08-20 report shows status now hides a note from eight surfaces. Anyone reading the note alone will publish or unpublish on a rule that was reversed six days later.

Tasks

[x] Settle the context window in session-context-measured.md. Done 2026-11-09. The body already said 1M and named 835k as the largest of twelve sessions; the description was the wrong half. Corrected, with a dated line saying which half was wrong.
[x] Fix or refute the answer channel. Done 2026-11-09, and refuted. plans/decisions/answers.jsonl carries answers and acks through 2026-11-05, and PANTRY owns the route that writes it. The two old notes now say they describe the era before P5, and point at p5-answer-on-finish for the current write path. The defect they recorded was the Record button, not a dropping log.
[x] Correct the DRAFT paragraph in folding-a-note.md. Done 2026-11-09. src/content.ts now hides an explicit DRAFT, so the 2026-08-14 sentence was reversed on 2026-08-20; the older line stays as history and is marked reversed.
[ ] Work the 35 queued facts in the organizer report, in the order it lists them. Each one gets a note only if it still holds, and a note is one fact, per SESSION-LOOP section 4.
[ ] Fold the six duplicate clusters the analyzers named, deleting only where the survivor already carries every sentence.
[ ] Re-run bun run dream check-index after the folds, because a fold is where an index breaks.
[ ] Run one more dream over the result, with the analyzer prompt unchanged, and compare what it finds against this list.

What this is not

Not a rewrite of the store. The 99 notes are mostly right, and the first pass proved the index could be fixed without touching them. This plan only moves the ones a reader would otherwise be misled by.