telemetry: per-turn fold records (know if cache/memory work landed) #560

Parent

#548 (agent session architecture)

Problem

We will not know if #549 / #557 / #558 / #552 worked. The usage chip is a
turn sum (#547) plus, later, occupancy (#556). There is no per-turn
record of what we sent and whether we stayed put.

Orrery writes a routing_record even when the policy is “stay.” We do not
need their router. We need the table.

Goal

One structured row per completed (or aborted) turn, host-local or a small
session meta field — not a new product:

• model id, effort if any
• occupancy / last-step in (status bar: show current context occupancy in addition to summed turn usage #556 when it exists)
• cache-read tokens if the provider sent them
• prefix-ish size (system + tools + notes), if we can measure without a
  client tokenizer — otherwise omit, do not invent
• did we compact, did the operator switch model, tool-error / stall bits
  if cheap

Success: after a week of use we can answer “did #558 raise cache-read
ratio?” and “did Next mid-session blow the prefix?” without reading Blob.

Constraints

• Not a learned policy. Not multi-provider fallback. Not Factory eval.
• Do not block A1–A7 on this; ship a sink once A1/A7 exist so the
  schema does not churn.
• No PII in an export path unless explicitly flagged (Orrery replay-set
  lesson).
• Not status bar: show current context occupancy in addition to summed turn usage #556 (how we paint occupancy). This is the log.

Suggested next

create-plan after #558’s assembly order is real enough to name the fields.