track-open-worklisted

# Track Open Work — the tracker is the entrypoint, not the graph ## The boundary **A session does not write its own memory.** Episodic writes happen *after* a session ends, by `thalamus extract` over the retained transcript. `thalamus write` and `thalamus extract --force --write` survive as operator actions **from outside a session**, and `write-guard.sh` blocks both from inside one. Reading `thalamus write` as a general permission is the specific mistake this exists to stop. The rationale prices it: distillation writes the session regardless, so a live write is a *second* pass over the same session — claims are content-addressed on (kind, normalized description) so a re-phrased one mints a new node instead of converging, and **threads get fresh ids, so both stay open in `memory_open_threads`, the surface the next session reads first.** A synthetic Session written mid-flight corrupts that same entrypoint, which is why it was rejected as an alternative. The one write verb a session does hold is **closing** a thread, and only through approval: `thalamus thread propose` writes a ledger row and nothing to the graph, then the operator approves. Report the title, a 1–2 sentence description, **and** the proposal id — all three, every time. The operator approves remotely and cannot read the ledger to find out what they are approving. ## Where a thing goes | It is… | Where | |---|---| | Work a future session should pick up | **A GitHub issue** — file it | | Something this sessio