Ledger of what the router sends to providers, kept in memory and journaled to disk.

It listens to [:mead, :router, :route] and keeps:

• a counter row per session, dropped by forget/2 when the session is deleted;
• a counter row per {route, provider};
• a ring of the last 100 turns;
• a count of the events it could not read.

The process owns the two public ETS tables, so they live as long as it does. The :telemetry handler writes to them in the process that emitted the event, which keeps the ledger off the critical path of every conversation. Because :telemetry detaches a handler that raises, the handler never raises: an event it cannot read is counted as dropped.

With a storage directory (config :mead, :storage, directory: ..., or the :directory option) the handler also casts every event to the ledger, which appends it to ledger.journal (Mead.Ledger.JournalFile) in batches of up to 100 events or 50 milliseconds, with one sync per batch, and again when it stops. A turn never waits for the disk. At boot the ledger replays the journal into the tables before it attaches the handler. A crash of the whole VM loses the events still waiting for the next batch. Without a directory the ledger is memory only and receives no casts.

A counter row is {key, turns, failures, input_tokens, output_tokens, duration_ms}.