Investigate memory-file fragmentation from per-session compaction appends (2026-06-07.md has 5+ duplicate date headers)

Per LEARNINGS.md 2026-06-08 reflection. memory/2026-06-07.md is 82KB / 659 lines because at least 5 Slack-channel/thread sessions on Sunday triggered memory-flush compactions that each wrote a fresh `## 2026-06-07 — Memory` + `## Nightly self-reflection (04:45 UTC)` header followed by placeholder text like `[content preserved from earlier]` or `[... existing content preserved ...]` instead of appending below the last `---` separator. Result: the file is structurally noisy. A future-me reading it cold cannot find the nightly summary on first scroll because the file looks like 5 different daily memories spliced together. Stefan reading it on his phone would not find his standup. Observed durable content blocks (which are individually fine, just badly stitched): 1. Nightly self-reflection (the canonical 04:45 UTC block) 2. Afternoon session: ~/shared-knowledge/NO-AUTO-MERGE-RULE.md distribution gap investigation (13:00-13:29 UTC) 3. Multi-agent ops coordination observed (13:15-13:36 UTC) 4. Multi-agent #agent-ops observation (13:06-13:27 UTC) — overlaps in time with #3 but produced as a separate section 5. #agent-ops mono-merge-policy thread — afternoon group discussion (13:01-13:15 UTC) 6. #agent-ops PR-dedup thrash → CIR-501 guardrails (13:50-14:18 UTC) 7. Victor sessions_list correction + CIR-499 closed (15:13-15:14 UTC) 8. Pre-compaction observer note: shared-knowledge split-brain thread (15:27-15:35 UTC) Each gets its own `## 2026-06-07 — Memory` top-level header. That is the bug. DO THIS: 1. Find the compaction tool / prompt that writes to memory/YYYY-MM-DD.md. Likely in shared cron-prompt config or in a per-channel-session memory-flush instruction. Search ~/.openclaw/ and ~/shared-knowledge/ for memory/ write patterns. 2. Diagnose whether the tool (a) doesn't know the daily file already exists, (b) knows but re-emits the header unconditionally, or (c) is correct in its template but the per-channel/per-thread sessions running it have no shared lock so two of them race. 3. Choose fix at the right layer (May 7 blast-radius): - If the tool template is wrong: propose patch / land it directly if in my workspace. - If the tool is correct but lacks an idempotent append primitive: propose to Victor. - If the cron / channel-session config wires it incorrectly: propose to Victor. 4. As a secondary action, write a one-liner reconcile pass for 2026-06-07.md so this Sunday's memory has one canonical structure (merge the 5 duplicate headers into one date header + chronologically-ordered sections). Do NOT delete content; only restructure. P3, 14-day cap. Not a fire — the content is all there, it's just badly stitched. But it's an interesting failure mode (per-session memory-flush racing on a shared file with no append-idempotency) that's worth understanding before peer-agent storms become routine and every Sunday produces an 82KB tangled memory file. SEE: LEARNINGS.md 2026-06-08 entry (this entry, the cost-regression context that surfaced it); memory/2026-06-07.md (the artifact); ~/shared-knowledge/ for shared cron/memory templates.