hir-123: swap fetch() to @anthropic-ai/sdk in triageReplyLLM

[jaredzwick]

Follow-up from CTO review on HIR-122 for #14. Founder pre-approved deferring the SDK swap out of the original PR.

Summary

• Replace fetch() against https://api.anthropic.com/v1/messages with new Anthropic({ apiKey }).messages.create({...}) from @anthropic-ai/sdk.
• Same input/output contract for triageReplyLLM: still returns TriageResult | null, still validates intent against the enum, still clamps confidence, still wrapped in the same try/catch so it never throws.
• Same model (ANTHROPIC_MODEL env / claude-haiku-4-5-20251001 default), same TRIAGE_PROMPT, same heuristic fallback path — none of those move.
• Tests now mock the @anthropic-ai/sdk module directly via vi.mock + vi.hoisted instead of overriding globalThis.fetch. All 14 replyTriage cases keep their original assertions.

Stacked on #14 — base branch is hir-120/warm-reply-triage. Will rebase to main after #14 merges.

Out of scope (per HIR-123)

• Prompt changes, model changes, fallback heuristic changes.
• Streaming / tool-use / response_format JSON. We still parse the text block ourselves through parseLLMJson.

Test plan

• pnpm test:int — 109 passed, 1 skipped. The unrelated api.int.spec.ts Payload-secret env failure exists on the base branch too.
• npx tsc --noEmit — clean.
• CTO review.