bug(kb): exception-driven _fallback_chunking also bypasses subclass metadata enrichment #14480

Summary

Same defect class as #14467, different trigger. SemanticChunkerBase._fallback_chunking() — the path chunk_text() falls back to on any exception in the primary chunking pipeline — builds each chunk's metadata directly in _create_fallback_chunk():

metadata={"fallback_chunking": True, **(metadata or {})}

This never calls _build_chunk_metadata() / _extra_chunk_metadata(), so a chunk produced via the exception fallback is missing chunk_index, total_chunks, embedding_model, chunking_method, and any subclass-contributed key such as optimization_version (see GPUSemanticChunker._extra_chunk_metadata()) — the same shape gap #14467 fixed for the zero/one-sentence path.

Where it was found

While fixing #14467 (single/zero-sentence chunks bypassing enrichment), the same audit covered every return path out of chunk_text(). _fallback_chunking() is the third one (autobot-backend/utils/semantic_chunker_base.py, _create_fallback_chunk and _fallback_chunking), and it has an identical gap. Kept out of that PR to keep the fix scoped to the reported symptom.

Consumer impact

Same conclusion as #14467: no current production consumer reads the subclass-contributed keys off chunk.metadata (knowledge_sync_incremental.py builds its own metadata dict from base_metadata and does not read chunk.metadata; advanced_rag_optimizer.py never calls chunk_text() directly). No stored data is affected today, but the inconsistency is a live seam for a future subclass or consumer.

What is needed

Route _create_fallback_chunk() through the same _enrich_chunks_with_metadata()/_build_chunk_metadata() call the other two paths now use, so all three return paths out of chunk_text() agree on one enrichment point.

Acceptance criteria

• A chunk produced via the exception-driven fallback path carries the same subclass metadata keys as one from the primary path
• Asserted via a test that exercises chunk_text() with a stubbed exception in the primary pipeline (not _create_fallback_chunk() directly)

Refs #14467