v0.9.2 — Budget ↔ actual reconcile

Meridian v0.9.2

Budget ↔ actual token-meter reconcile on the 0.9.x track (v1.0 still deferred).

Highlights

• Pre-flight budgets still reserve on estimate
• After successful response with backend usage, token meters adjust to actual cost (prompt * prefill_weight + completion * decode_weight)
• Request counters never adjusted; failed/upstream-error requests keep the estimate
• Usage scrape when cost or budgets need tokens
• Metric: meridian_budget_reconciles_total{direction}

Images

• ghcr.io/imv-in/meridian:0.9.2 / :latest (release workflow)

Next (0.9.3)

Load/overhead numbers, richer e2e, ops polish for multi-tenant fleets.