Claude-native telemetry: token & cost accounting

Problem

Token usage is the primary efficiency signal, but the raw stream splits tokens across input / cache-creation / cache-read / output and the local costUSD field is unpopulated (observed 0). Without a correct accounting layer, cost and efficiency claims are unreliable.

Capability

Turn normalized turns into a correct token ledger and computed cost, aggregated over the sprint window.

Behaviour

• Per-turn ledger: billable input = input + cache-creation; cache-read tracked separately (~0.1×, session overhead, not per-tool); output; cache TTL split; server tool use.
• Compute cost from tokens × an offline pricing table keyed by model; never trust local costUSD; treat missing price as "unknown," not 0.
• Model registry/alias resolution; record unknown models.
• Aggregate by sprint window (git refs), then session/day/model/project.

Acceptance criteria

• Billable tokens for a fixture match a hand-computed value.
• Cache-read reported separately and excluded from per-tool attribution.
• Unknown-model turn counts tokens and is flagged "cost unknown," never priced 0.
• Sprint-window totals equal the sum of in-window deduped turns.

Dimensions: Delivery Predictability (cost trend, outliers), Collaboration Efficiency (cache-read / output ratios). Tier 2. Depends on #36. PRD: docs/prd/features/02-token-cost-accounting.md.

Start by reading docs/prd/features/02-token-cost-accounting.md and issue #36, then trace where normalized turns and sprint-window git refs enter the system. Done means deduped in-window totals satisfy the listed ledger, pricing, cache-read, unknown-model, and aggregation acceptance criteria.