[Task 12] Pricing sync core

[lubobill1990]

Task 12 — Pricing sync core

Depends on: 02
Unblocks: 13

Goal

Two fetchers + one LLM caller + validators. No DB writes yet (task 13 owns
the version-write logic).

Scope

New file src/lib/pricing-sources.ts:

export async function fetchAzureRetailPrices(): Promise<AzureRow[]>
export async function fetchAnthropicPricingHtml(): Promise<string | null>

• Azure: https://prices.azure.com/api/retail/prices?$filter=serviceName eq 'Cognitive Services' and serviceFamily eq 'AI + Machine Learning'
  with paging (NextPageLink). Return all rows.
• Anthropic: GET https://www.anthropic.com/pricing, locate the pricing
  section, return the HTML slice (or null if not found / claude not in
  state.models).

New file src/lib/pricing-sync.ts (validators + LLM caller part only):

export async function buildSyncRequest(): Promise<SyncRequest>
export async function callSyncLlm(req: SyncRequest, model: string): Promise<ParsedPricing>
export function pickSyncModel(state, cliFlag): string
export function priceChanged(oldRow, newRow): boolean       // 0.5%
export function sanityFails(oldRow, newRow): boolean        // 10x

pickSyncModel priority:

1. CLI flag value if it is in state.models.
2. First match in whitelist ['gpt-5', 'gpt-4.1', 'gpt-4o', 'claude-sonnet-4', 'claude-3-7-sonnet'].
3. state.models.data[0].id.

LLM call always sets x-internal-pricing-sync: 1 and
response_format: { type: 'json_object' }.

System prompt: see design doc 03. The user message carries
{ knownModels, azureRows, anthropicHtml }.

Definition of Done

• Unit tests for priceChanged (epsilon edges, NULL vs value, zero).
• Unit tests for sanityFails (10x boundary, NULL handling).
• Unit test for pickSyncModel — flag set / flag invalid / fallback path.
• Manual run of fetchAzureRetailPrices() returns >100 rows and
  populates expected fields.

---

• Source: docs/tasks/12-pricing-sync-core.md
• Design: docs/design/
• Phase: pricing
• Depends on tasks: 2