fix(anthropic): preserve rendered-context provenance

Problem

PXPipe can render static Anthropic system/configuration context into image blocks attached to the first user message because Anthropic system blocks cannot contain images.

The rendered image banner currently tells Claude to follow those user-role images as operating instructions. In real Claude Code usage this can be interpreted as apparent prompt injection / reasoning-extraction framing and cause a false refusal, even though the content originated from the legitimate request.

Expected behavior

PXPipe should preserve truthful provenance when relocating trusted configuration across modality/role boundaries:

• rendered image framing should be descriptive, not role-elevating;
• the actual system field should state that PXPipe produced those images from the request's configuration;
• Claude Code's preserved identity must remain first;
• tool schemas, request content, cache semantics, and unrelated providers must remain unchanged.

Observed validation

After applying the fix locally:

• focused Anthropic provenance/safety/cache regressions pass;
• production dependency audit passes;
• TypeScript typecheck passes;
• full test suite passes;
• build and command-entrypoint smoke checks pass;
• a freshly rebuilt/restarted PXPipe daemon works correctly with Claude without the previous false refusal.