v0.9.0

Closes the forecast loop: forecasts bind to dispatches through verified receipts, observed tokens correct token forecasts, and a backtest scores duration forecasts.

Changed

• Point the claude, codex, production, and sonnet aliases at Claude 5 and GPT-6 models (reliability policy defaults).
• Require PyYAML 6.0.1, since 6.0 fails to build from source (minimum dependencies).

Added

• Add local-policy reliability limits for Claude 5 and GPT-6 models (reliability policy defaults).
• Forecast subscription meter points from a caller-supplied meter table, as an experimental opt-in (subscription points).
• Correct token forecasts per segment from caller-supplied observed tokens (measured correction).
• Report expected cache-read tokens as a third token slot (cache-read share).
• Inspect audit-corpus coverage and score verified native duration forecasts without changing calibration (backtest).
• Identify matching forecast inputs with a fingerprint that new nullable fields leave unchanged (forecast keys).
• Bind forecasts to dispatches with immutable, verified receipts (binding receipts).