Memory/runtime maintenance is healthy, but operational backlog remains material. The latest restore and model-heartbeat completed successfully (restore_result.ok=true, rc=0, acpx_rc=0, NO_CHANGE). The preferred model conversation log has no entries after 2026-09-05T10:13Z; this is not treated as proof of no current conversations because authoritative turns/backlog surfaces were unavailable.
• `memory-heartbeat.json`: 7,528 source entries, 6,487 visible; restore packet generated 2026-09-09T04:09Z; `semantic_changed=true`.
• `model-heartbeat.json`: latest run 2026-09-09T04:09Z, successful, no durable curation change.
• `model-conversation-log.jsonl`: 6,394 rows; latest visible direct user/assistant activity is 2026-09-05T10:13Z (Hamburg coffee search).
• `semantic-candidates.jsonl`: latest row remains `needs_more_evidence` and is transient cultural-event context; no promotion recommended.
• `actions.jsonl` / `notices.jsonl`: open backlog still includes the durable polling control-plane design, scheduled-worker model/config repair, stale decision-state, semantic-memory audit, and candidate review.
• `directed/latest.md` and `directed/index.json`: no newer directed report than the historical May 26 material; the index was refreshed but does not represent a new report.
1. **Healthy memory path, unresolved control-plane drift.** Restore and curator heartbeat are succeeding, but the scheduled-worker repair remains open because prior failures identified obsolete kimi-for-coding/k2p6 selection. Do not edit generated continuity output directly; repair the source/config and prove two clean runs.
2. **Polling architecture remains the highest-value engineering action.** Preserve the agreed isolated design: arbitrary per-watch schedules, durable queue, dedupe/debounce, retries/backoff, deadlines, wake budgets, shadow mode, and visible subagent lifecycle. ACP resume is only the transitional dispatch adapter.
3. **Candidate state is review material, not fact memory.** The last semantic candidate is a one-off event request and should remain unpromoted. The prior curator report says 13 open memory candidates, but this cycle could not independently recount helper status.
4. **State freshness is incomplete.** decision-state.json is still dated 2026-05-27 and directed synthesis has not produced a newer report. This is durable maintenance debt, not evidence of missing user conversations.
• The apparent quiet period in the preferred model log may reflect filtered/direct-ACP ingestion and unavailable authoritative turn status rather than absence of user activity. Verify with `turns.jsonl` and `persistent_chat_log_backlog_status` when exposed.
• Generated model-identity drift and scheduled-worker failures likely share a stale worker configuration/source path; inspect the worker invocation rather than repeatedly correcting generated summaries.
• Repair scheduled-worker model/config selection and verify two consecutive successful worker runs.
• Review the 13 open memory candidates and the single `needs_more_evidence` semantic candidate through curator helpers; do not promote the latter.
• Stage the isolated polling control plane in shadow mode before touching the live timer dispatcher.
• Refresh `decision-state.json` through its owning workflow and harden directed-index notice repopulation against resolved/obsolete reports.
• Native reads of `/home/matrix-lite/state/turns.jsonl` and `/home/matrix-lite/workspace/state/timers/timers.json` were denied.
• `observability.jsonl` is absent; timer/subagent/backlog status tools were not available in this cycle.
• No subagent was spawned; investigation used read-only file inspection only.
Updated recent-topics.md with this cycle's current operational state. No curated-memory proposal: there is no new durable user fact or preference to promote.