Add the digest layer end to end. What ships: - new `talent/digest.md` as a tier 3 cogitate talent with `schedule: "none"` - polymorphic `sol call identity digest` CLI where default mode triggers, waits, and prints a short success line, while `--write --value` is the persistence sink - `_maybe_submit_startup_digest()` in the supervisor, fired once per boot after `_task_queue.set_ready()` - `digest.md` seeded with `"not yet generated\n"` via `ensure_identity_directory()` This keeps D1 on the polymorphic verb path, mirroring the `awareness_tender` pattern so trigger and write sink share one verb and no separate hook file is needed. For D2, startup uses `_task_queue.submit(["sol", "call", "identity", "digest"])` so the fire-and-forget path gets subprocess logs, env inheritance, and the same codepath manual callers use. Startup submission is skipped under `--no-cortex` and `--remote`, and still runs under `--no-daily` and `--no-schedule`. Not in scope here: chat talents, chat stream, UI, or auto-cadence work. Those stay deferred to later lodes. The talent prompt now tells the talent to compute a 7-days-ago `YYYYMMDD` at runtime before calling entity search, since that CLI only accepts absolute dates and a baked-in value would go stale. `make ci` is green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>