Clone the fedorov/grandqc idc-dicom-fixes branch, which carries the
DICOM-directory, device-detection, PyTorch >=2.6 load, small-slide
overlay, and fail-loud fixes as proper commits. Replace the in-notebook
patch_grandqc_for_dicom() cell with a verification cell that asserts the
fixes are present in the cloned scripts, and update Part 2 prose plus the
resolution notes accordingly.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The demo selection picked the smallest TCGA-BRCA slides, which are
frozen-section (TS/BS) slides absent from GrandQC's Zenodo reference
archive (DX diagnostic slides only) — so Part 6 validation matched
nothing. Switch to five verified DX barcodes (variety by magnification,
not tissue type, since all DX slides are primary tumor).

Part 6 now downloads just those 5 reference masks (~700 KB, re-hosted)
instead of the full 2 GB BRCA.tar, and fixes the mask-name matching:
reference masks are {barcode}.{UUID}.svs_mask.png while local masks are
keyed by the bare ContainerIdentifier, so a barcode->filename map is
built. Drop the cohort-wide statistics that required the full archive.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

New section 6.4 overlays per-class artifact region boundaries for the
reference (Zenodo) and local DICOM runs on a shared grid: solid outline
for the reference, dashed for the local run, in the matching artifact
colour. This exposes boundary-level localisation differences that the
aggregate per-class fractions in 6.3 can hide. The reference mask is
resized onto the local grid (nearest-neighbour) since the two can differ
slightly in pixel dimensions. Existing cohort-scaling note renumbered to 6.5.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

If the download cell is re-run after a rename, both the UID directory
and the renamed barcode directory exist, and os.rename(src, dst) onto a
non-empty dst raised OSError. Handle the both-present case by dropping
the redundant UID copy and keeping the barcode directory, so the cell is
safe to re-run in any order.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

With five demo slides the single-row figures made each panel tiny.
Switch the tissue-detection overlays, artifact overlays, and the
artifact-boundary comparison to a 2-column grid with larger per-panel
sizes, hiding any unused trailing panel.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The overlaid solid/dashed contours were hard to read. Instead show the
reference (Zenodo) and local DICOM artifact maps side by side, one slide
per row, each colourised with the same GrandQC palette and a shared
class legend. No resizing needed since the panels are not overlaid.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Upstream GrandQC tiles slides with floor division and fills the sub-patch
right/bottom edge remainder with a background buffer, so tissue/artifacts in
that margin are never inferred. This adds Part 7 to
grandqc_slide_quality_with_idc.ipynb implementing the alternative
boundary-anchored strategy: a ceil grid plus edge tiles read from a
back-shifted (extent - p_s) origin and realigned to the grid.

Part 7 changes ONLY the edge handling — OpenSlide DICOM reading, the
GrandQC_MPP15 model, MPP 1.5, the Step 1 tissue masks, the tissue gate, and
the argmax label convention (including stock's interior +1 read offset) are
all kept identical to Part 4, so interior tiles are inferred identically and
the two runs differ only along the slide edges. 7.2 compares stock vs
anchored masks and quantifies the edge tissue recovered. Parts 1-6 are
unchanged.

Also adds grandqc_edge_tiling_notes.md documenting both strategies with code
and a worked example.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>