Make multi-grid writes and auxiliary outputs idempotent

[CraigMyles]

Summary

• keep grid_hash unchanged as the grid identity while suffixing colliding human-readable grid labels with a short hash and reporting the actual stored key
• bind receipt completion to the requested grid, prefer current hashes over compatible legacy hashes, and reject ambiguous hashless multi-grid resume
• make QC, thumbnails, GeoJSON, and inline slide encoders produce-if-missing on completed stores and during model append without re-embedding patches
• keep the primary thumbnail/GeoJSON filenames backward-compatible while namespacing grid-dependent overlays and polygons for additional grids
• repair interrupted QC and thumbnail outputs, regenerate sidecars on fresh/forced builds, reuse the grid's recorded segmenter and thumbnail settings, and replace PNG/GeoJSON sidecars atomically without changing shared-file permissions

Validation

• Spark: pytest -m "not slow and not network" -q — 471 passed, 90 deselected
• focused Spark integration suite — 104 passed, 9 deselected
• ruff check .
• git diff --check

Closes #4
Refs #6