PIGS-864: Add render_pdf_page tool for inline page image rendering

[rpalekar]

https://gonitro.atlassian.net/browse/PIGS-864

Summary

• New MCP tool render_pdf_page(inputPath, pageIndex, dpi, clipBox?) returning the rendered page as an inline base64 PNG content block — no disk writes.
• Built by composing existing endpoints (splitPdf → convertFile → in-memory unzip + crop + resize via fflate/jimp) to avoid cross-repo work. See the Jira comment for the two approaches considered.
• clipBox is in 72-dpi PDF points (top-left origin) so detection boxes from extract_pii / search_text_in_pdf / redact_pdf round-trip directly into close-up inspection.
• DPI range 50–300 (default 150); the platform always rasterises at 300 DPI, so we downscale client-side and can't go sharper than 300.
• Unlocks the PwC Demo 3 use case (visual verification of policy-driven redactions).

Trade-offs

• Two backend round-trips per render (split + convert). A future follow-up will replace the pipeline with a single platform render-page method and keep the public tool surface unchanged.
• Always renders at 300 DPI server-side then scales down — wasted backend work for low-DPI full-page renders.

Test plan

• task n:check — format, types, lint all clean
• task n:test — 168 tests pass (1 skipped), rendering.ts at 100% statement coverage
• Manual smoke test in Claude Desktop: full page, low-DPI thumbnail, clipped region, and the PII-verification end-to-end flow all return an inline image that the model can describe visually.

🤖 Generated with Claude Code