Report per-agent and per-engine token usage from a sealed run

[apresmoi]

What

A sealed run now reports what it cost. observe derives per-agent and per-engine token usage from the exported Daimon ledger at raw/daimon/usage.jsonl, surfaced as usage[] and usage_summary on the observe report.

No lifecycle-contract change was needed. Simfile already reads exported raw artifacts directly — this follows exactly how memoryBanks.ts derives per-bank counts from raw/mneme/<bank>/. The pinned 43-row capability contract is untouched, so this needs no coordinated Spawnfile release.

Boundaries hold: no Spawnfile import, own zod schema with .strict() so an unknown key is rejected rather than stripped, file reads only, nothing selects or wakes cognition.

The lower-bound guarantee is structural

usage_summary.lower_bound is z.literal(true), not z.boolean(). A report asserting these counts are exact is not representable in simfile.observe.v1 — there is a test that a lower_bound: false report fails to parse.

This matters because the qualification is load-bearing. An all-zero usage block means unknown, not free; AGY reports no cost field at all; codex is unmetered by design. unknown_turns carries the producer's complete: false signal per agent and in aggregate. A report that quietly presented these as exact would recreate the exact failure this feature exists to prevent — a dead subscription that looked like zero cost.

Absent is not zero

collectUsage returns undefined when neither ledger generation exists, and the report omits usage entirely — following worldGrants.ts, which already establishes that convention and documents it. A present-but-empty ledger aggregates to [], an observed zero. The two are distinguishable.

Both rotation generations are read, oldest first. Engine attribution comes from an agent's earliest record, so an agent that switches engine mid-run reports the one it started on.

Verification

1687 tests passing. tsc --noEmit clean on both tsconfig.json and tsconfig.web.json.

Four mutations, each confirmed against the named subtest it kills rather than the failure count: reversing generation order, treating an absent ledger as empty, relaxing lower_bound to a boolean, and dropping the incomplete-turn signal.

Related

Depends on the export side in noopolis/spawnfile#7, which carries both ledger generations into exported run artifacts.