feat(llm): cost telemetry + per-user daily token budget [STORY-012]

[khoks]

Summary

• Versioned cost calculator — MODEL_PRICING table for Opus/Sonnet/Haiku stamped with PRICING_VERSION = "2026-04-26"; costFor() returns cost_usd=0 + known_model=false for unknown models so analytics can flag operator-stale tables without breaking the runtime path.
• Per-user daily token budget — UsageStore interface + InMemoryUsageStore keyed by (user_id, UTC date); DailyTokenBudget with assertWithinBudget() (throws TokenBudgetExceededError at limit) + decideModel() (downgrade ladder Opus → Sonnet → Haiku, kicks in at threshold = 0.8).
• BudgetGatedLLMProvider decorator — wraps any LLMProvider: pre-call assert + decideModel (may downgrade), post-call record. Embed passes through. limit=0 means unlimited (self-hosted default); explicit req.model always wins.
• Telemetry schema extension — LLMTelemetryEventSchema gains cost_usd, pricing_version, optional session_id / cached_tokens / tool_used. AnthropicProvider now stamps cost via costFor(); toolCall populates tool_used from the first invocation. CompleteRequestSchema gains optional session_id.
• DB persistence split — agent_calls Drizzle migration + DB-backed sink + API 429 mapping moved to STORY-060 to keep STORY-012 at its S estimate. Interfaces (UsageStore, LLMTelemetrySink) are stable; STORY-060 just adds Drizzle impls behind them.

Acceptance criteria

• All telemetry fields present on LLMTelemetryEvent (provider, model, role, user_id, session_id, task, input/output_tokens, cached_tokens, cost_usd, pricing_version, tool_used, latency_ms, ok, decided_at, prompt_version) — agent_calls table is in STORY-060.
• Daily budget enforced server-side (BudgetGatedLLMProvider) — applied at the provider layer so any caller goes through it.
• Graceful model downgrade kicks in at threshold (default 80%) via the MODEL_TIERS ladder.
• At 100%, TokenBudgetExceededError carries a human-friendly message; HTTP 429 mapping in STORY-060.
• Cost calculation uses a versioned price table — PRICING_VERSION = "2026-04-26".

Test plan

• pnpm --filter @learnpro/llm test — 72 passed / 1 skipped (integration test, needs ANTHROPIC_API_KEY).
• 38 new tests across pricing.test.ts (6), budget.test.ts (20), budget-gated-provider.test.ts (12).
• pnpm typecheck — green across all 12 packages.
• pnpm lint — green across all 8 lintable packages.
• pnpm test — green across the monorepo.
• Manual smoke (deferred to STORY-060): with LEARNPRO_DAILY_TOKEN_LIMIT=100 + a real Anthropic key, hit the playground twice and observe the friendly 429 on call chore(meta): lift no-code rule, fix grace-days, add PR workflow #2.

🤖 Generated with Claude Code