feat(ai): enforced budgets + the callModel funnel — metering that can refuse #5626

Summary

PR A of the enterprise metering build (from the 2026-09-02 audit: 74 call sites / 16 metered / 5 quota-checked / zero dollar-cap refusals; per-org spend unrepresentable).

• Migration 0136: ai_cost_events.org_id (+index) and org_ai_budgets (monthly_cents, per_category_cents JSONB → per-feature allocation, the "big corps allocate so much AI funding" model).
• recordAiCost carries orgId, derived from repositoryId via a 5-min-cached lookup when not passed.
• assertAiSpend() — the first pre-call refusal path: platform daily ceiling → org monthly cap → per-category allocation; throws AiBudgetExceededError (scope + spent + cap). Fails open on ledger faults; cap 0 = no cap, never "spend nothing".
• callModel() — gate → call → noteAiFailure → recordAiCost in one wrapper; the only sanctioned way for new AI code to reach the model.
• ai-metering-coverage.test.ts — pins the 54 legacy direct-caller files as a shrink-only list; a new direct caller fails CI with instructions to use callModel().
• Agent daily caps enforced for real: hosted loop pre-checks getAgentUsage and refuses before spending; chargeAgent's previously-discarded boolean is honoured, with overshoot recorded via recordAgentOverdraft instead of silently vanishing from the books.

Test plan

• New: ai-metering-coverage (2), ai-budget-enforcement (4) — pass
• Neighbors: agent-multiplayer, ai-budget, ai-cost-tracker, hosted-claude-loop, hosted-loop-secrets, hosted-loop-spend-floor — 85 pass, 0 fail
• tsc --noEmit clean, bun build clean
• bun run db:migrate on deploy applies 0136