feat(cli): update @anthropic-ai/sdk to 0.92.0 with prompt caching

[0xAxiom]

What

Updates the Anthropic SDK from 0.32.1 to 0.92.0 (60 releases) and adds prompt caching to cut costs on repeated pipeline runs.

Changes in CLI/src/core/anthropic.ts

Prompt caching — the biggest functional improvement. AppFactory runs multi-stage pipelines where large system prompts are sent to Claude on every stage call. By marking system prompts with cache_control: { type: "ephemeral" }, the Anthropic API caches the tokenised system prompt for 5 minutes and applies a ~90% discount on cached input tokens.

// Before
system: systemPrompt

// After — system prompt is cached between repeated calls
system: [{ type: 'text', text: systemPrompt, cache_control: { type: 'ephemeral' } }]

Updated default model — claude-sonnet-4-20250514 → claude-sonnet-4-6 (latest Sonnet generation).

Typed error handling — replaced fragile string-matching retry logic (msg.includes('rate')) with proper SDK error classes:

// Before
if (errorMessage.includes('rate') || errorMessage.includes('429')) { ... }

// After
if (err instanceof Anthropic.RateLimitError) return true;
if (err instanceof Anthropic.InternalServerError) return true;

Idiomatic streaming — streamClaude now uses .on('text', ...) + finalMessage() instead of manually iterating raw SSE events.

Why

• @anthropic-ai/sdk 0.32.1 is ~60 minor versions behind the current 0.92.0, missing improvements to types, error handling, and streaming helpers
• Prompt caching directly reduces per-run API costs; pipeline runs with large shared system prompts are a perfect fit for it
• String-based retry detection is fragile — using the SDK's typed error hierarchy is more reliable and maintainable

Tested

• TypeScript compiles clean (tsc --noEmit)
• All 252 unit + integration tests pass (npm run test)
• No behaviour changes in stub/test mode (existing test coverage validates all paths)