feat(providers/llm): real Anthropic Claude adapter with prompt caching

[wesleysimplicio]

Context

ClaudeProvider at lib/providers/llm.ts:42-46 is a mock. Per AGENTS.md routing, Claude is the default for orchestration, script, compliance, humanization (and fallback for caption / translation).

Goal

Replace the mock with a real adapter against the Anthropic API. Use the claude-api skill guidance: import @anthropic-ai/sdk, default to the latest model (claude-opus-4-7 for high-quality / claude-sonnet-4-6 for cost-sensitive), enable prompt caching on the system message and brand context.

Acceptance criteria

• npm i @anthropic-ai/sdk
• ClaudeProvider.generate calls the Messages API with the correct model per task (compliance / script → opus-4-7; caption / humanization → sonnet-4-6)
• Prompt caching enabled (cache_control: { type: 'ephemeral' }) on stable preamble blocks: brand voice, persona, channel rules
• Returns GenerationResult with real tokens, cost_usd, latency_ms
• Reads ANTHROPIC_API_KEY from process env (loaded by CLI)
• Retry on transient errors (5xx / network), respect 429 with backoff
• Timeout 60s for short tasks, 180s for script
• Skipped when DRY_RUN=true (mock variant used; see #provider-mocks)
• E2E spec with mocked HTTP layer asserts payload shape (system prompt + cache control + messages)

Files

• lib/providers/llm.ts:42-46 (ClaudeProvider)
• package.json (new dep)

Depends on

• #provider-mocks-isolation (cleaner ground)
• #env-loader (env loaded consistently — see Sprint 0 acceptance)

[wesleysimplicio]

Closed after PR #40 follow-up stabilization in commit f40b58d. Current main passes npm run typecheck and npm run test:e2e (119 Playwright specs green), and this tracker item is being retired per repository cleanup.