feat: add native Anthropic provider

[weeta-code]

Summary

Adds Anthropic as a first-class provider alongside OpenRouter, OpenAI, and Z.ai. Users can bring an sk-ant-... key directly and talk to the Messages API from the browser — no OpenRouter account, no local proxy.

How it works

• New "anthropic" entry in AI_PROVIDER_PRESETS; the sidebar settings UI picks it up automatically (no UI code changes needed — the existing map over AI_PROVIDER_PRESETS handles it, and the grounding toggle is already scoped to openrouter/openai only).
• getProviderHeaders branches to emit x-api-key + anthropic-version: 2023-06-01 + anthropic-dangerous-direct-browser-access: true instead of Authorization: Bearer.
• ai-enrich.ts and ai-ghost.ts branch on provider === "anthropic" at the HTTP-call boundary:
  • POST to {baseUrl}/messages instead of /chat/completions
  • System prompt moved to top-level system field (Anthropic's Messages API shape)
  • Structured output via forced tool_use — a synthetic tool (enrichment_result / emergent_thesis) is defined with input_schema and forced via tool_choice: {type: "tool", name: "...", disable_parallel_tool_use: true}. Schema fidelity is equivalent to OpenAI's strict json_schema mode; no regex fallback needed.
  • Response parsing reads content[].type === "tool_use" and uses .input directly.
• CSP connect-src in next.config.mjs gains https://api.anthropic.com (without this, Firefox silently shows NetworkError when attempting to fetch resource; Chrome says "Refused to connect... violates CSP").
• README gets an "Anthropic (direct)" subsection mirroring the existing provider entries.

Design notes

Why no thinking field is sent, even on 4.7: forced tool_choice: {type: "tool"} is explicitly incompatible with extended/adaptive thinking — the API rejects the combination with a 400. Thinking defaults to off on all exposed models (Opus 4.5/4.6/4.7, Sonnet 4.5/4.6, Haiku 4.5), so omitting the field is correct, forward-compatible, and gives deterministic low-latency enrichment that matches the schema every time. Task-fit also supports this: classification, connection inference, and short annotations don't meaningfully benefit from deep reasoning, and enrichment is the user-facing latency path.

Ghost synthesis could plausibly benefit from adaptive thinking, but that would require dropping forced tool_choice to auto and handling two response shapes (tool_use + text fallback). That's a cleaner follow-up PR than bundling with provider addition.

Web grounding is intentionally not wired up on the Anthropic path. Anthropic has a server-side web-search tool but it's a separate feature; supportsGrounding: false on all Claude models cleanly hides the grounding UI (same pattern as the Z.ai provider).

Files changed

File	Purpose
lib/ai-settings.ts	Add "anthropic" to provider union, preset, ANTHROPIC_MODELS, header branch
lib/ai-enrich.ts	Messages API + forced tool_use branch
lib/ai-ghost.ts	Same pattern for synthesis
next.config.mjs	CSP connect-src allowlist
README.md	"Anthropic (direct)" subsection

Testing

• npx tsc --noEmit — clean
• npm run build — passes (2.3s compile, all pages generated)
• Live test against claude-sonnet-4-6 and claude-opus-4-7: enrichment produces valid classifications, annotations, and influencedByIndices; ghost synthesis fires correctly after accumulating notes across categories
• Existing OpenRouter, OpenAI, and Z.ai code paths unaffected (branches occur before divergent logic)