End-to-end engine pipeline tests per modality

[martsokha]

Scope

Engine-level end-to-end coverage: run a real fixture through the full
DocumentPipeline::run_full body — extraction → detection → dedup →
redaction → validation → export — and assert the final audit record
and redacted bytes. This is the engine's regression net for the
per-modality pipeline shape.

The toolkit-only round-trip (decode → analyze → anonymize → encode,
without audit/export/validation) is tracked separately in the elide repo
(nvisycom/elide), since that layer lives there now.

What this issue becomes

Per-modality tests/pipeline_*.rs integration tests in nvisy-engine:

• pipeline_text.rs — TXT/JSON/HTML/PDF-text fixture → run pipeline →
  assert detected entities, applied redactions, redacted bytes, and the
  emitted audit record
• pipeline_tabular.rs — CSV/XLSX fixture, same shape
• pipeline_image.rs — PNG/PDF-image fixture with the noop OCR backend →
  assert the pipeline runs through to export; expand once a real OCR
  backend lands
• pipeline_audio.rs — WAV/MP3 fixture, noop STT path

Each test covers both the dry-run and full-apply paths. Lean on noop
backends to stay hermetic.

Scope when implemented

• crates/nvisy-engine/tests/pipeline_text.rs (+ tabular/image/audio siblings)
• Per-test fixtures under crates/nvisy-engine/tests/fixtures/
• Use the existing EngineInput workflow shape