spike(engine): prototype a read-only source-content capability classifier

[McanKul]

Goal

Prototype a bounded, read-only backend that identifies existing text and image occurrences and explains whether each occurrence could be edited safely.

This is a child deliverable of #11 and must not modify or save a PDF.

Scope

• Inspect the original source path and source page, not the temporary one-page pagePdf preview copy.
• Return PDF-space bounds, a stable opaque locator, an element kind, and an explicit capability result.
• Use reason codes for unsupported cases such as:
  • missing or ambiguous Unicode mapping
  • subset, Type 3, vertical, clipped, patterned, or nested text
  • shared, inline, masked, or nested images
  • encrypted, malformed, signed, or stale source documents
• Compare a bounded lopdf prototype with a PDFium sidecar prototype.
• Measure binary size, memory use, cross-platform build/signing cost, and native crash isolation.
• Keep PDF bytes and engine-native object handles out of React state.

Acceptance criteria

• The classifier is deterministic on the Spike: evaluate safe editing of existing PDF text and images #11 fixture corpus.
• Every result is supported or unsupported with a specific reason; there is no silent fallback.
• Opaque locators are tied to a source fingerprint and fail as stale if the source changes before save.
• Reused pages and reused XObjects are reported as separate occurrences without implying independent mutability.
• Parsing and decompression are bounded; malformed input returns a recoverable error.
• If PDFium is tested, it runs out of process so a native crash cannot terminate the OffPDF UI.
• The spike ends with a compatibility/performance matrix and a recommendation recorded on Spike: evaluate safe editing of existing PDF text and images #11.

Out of scope

• Writing a mutation back to the document.
• Paragraph reflow or a promise to edit every PDF.