emrg: the evolution readings walk a log directory whose writer was removed, so the recent-improvements list stops at 2026-08-23 #2187

The daemon's two self-evolution readings walk <logs>/evolution-*.json:

• EmrgServer._evolution_count() (the pong frame's evolution_count, i.e. the GUI growth counter) walks it for its "disk" half;
• the evolution_summary frame walks it for the recent list.

No writer fills that directory any more. e401246e (2026-08-19, rants 14:16/14:18/14:20) removed _write_evolution_log along with the archival it belonged to, and the same commit left the file behind — its own diff says "evolution-.json was never consumed"*. The durable store that replaced it is <logs>/task-runs/<task>.jsonl, written by TaskHandler._append_task_run and restored at init by _load_task_runs.

Measured on this host, 2026-10-13 (cyc20261013-080459):

So:

1. The recent list could not show a cycle written after 2026-08-23 — seven weeks of evolutions missing from a feature whose whole point is showing them (emrg: WorkBuddy P3 fixes — evolution_summary + result-panel export fix (rant 21:35) #502, WorkBuddy P3).
2. The count's disk half answers 27, a number that reaches the reader whenever it is the larger of the two the reading takes the max of — a host whose tasks are all disabled, or a daemon whose scheduler is not up (_scheduler is None falls back to the legacy self.evolutions, which is never appended). On this host the in-memory figure currently hides it; on a host with a smaller live store the growth card would be inflated to the dead directory's count.

Fix

One derivation of the durable store, in emrg/server/scheduler.py, used by both daemon readings and by the handler's own restore:

• task_run_files(dir) — the store's files, by name;
• read_task_run_records(path) — one file's records, oldest first, each naming its task; a missing/torn/not-a-regular-file subject answered by the reader itself (the kind asked before the open, as elsewhere in this module).

_evolution_count and the evolution_summary frame read that; _load_task_runs is routed through it instead of parsing the same lines a second time.

Acceptance

• A record written to the live store is visible to both readings — and the dead directory's files are not counted (both directions).
• The recent list is newest-first across tasks and carries the store's own fields.
• A FIFO or a directory planted at <logs>/task-runs/<task>.jsonl does not stop either reading, and the skip still names the kind.
• Full pytest suite green.