feat: native Anthropic provider adapter (C2)

[MANVENDRA-github]

Adds a real adapter for Anthropic's Messages API (which is NOT OpenAI-compatible),
so the advertised 4-provider support actually holds. Clients keep sending the
OpenAI shape; Sentinel translates in and out. (Closes the provider gap from the
v0.1.x audit.)

What changed

• New providers/anthropic.ts: OpenAI chat request → Anthropic Messages body
  (system message hoisted to top-level system, required max_tokens defaulted,
  top_p/stop mapped), and Anthropic response + SSE stream → the OpenAI shape.
  Auth via x-api-key + anthropic-version; redirect: 'error' preserved; the
  upstream response is parsed through a Zod schema (unexpected shape ⇒ 502).
• Config: provider type widened to 'openai-compatible' | 'anthropic' and
  threaded through ResolvedProvider; the registry builds the right adapter by
  type. Example config + .env.example carry an Anthropic entry.
• Docs: README / CLAUDE / CHANGELOG now state native Anthropic support;
  SECURITY_REVIEW_LOG SR-007 covers the new outbound auth path, SSRF, and
  response-shape parsing.

Tests

• Adapter request/response/stream translation, x-api-key header + no-key case,
  non-OK + bad-shape errors, default max_tokens (+ stop_reason mapping), array
  content / multi-system join / top_p+stop, and stream event filtering.
• Registry builds the Anthropic adapter for type: 'anthropic'; config parses
  it and rejects unknown types.

How to verify

• pnpm verify → green (230 tests, coverage ≥90%)
• pnpm build → green

Existing OpenAI-compatible providers are unchanged.