Migrate to PyQt6 and run SAM/DINO inference in-process

[cofade]

@bnsreenu — proposing two coupled changes that I think materially improve the cross-version story for this project. The PR is a single set of commits because the second change unblocks the first; splitting them would require a temporary half-state that's worse than either endpoint.

Closes #67 (PyQt5 + Python 3.14 + Torch DLL conflict → PyQt6 migration).
Closes #68 (multi-dim slice support in DINO "Detect All Images").

Summary

• PyQt5 → PyQt6 across the codebase (~30 source modules + tests + CI). Bulk enum-namespacing codemod, QAction / QShortcut module relocation, QDesktopWidget → QGuiApplication.primaryScreen(), event.position() + QPointF end-to-end in image_label.py. Documented in ADR-014.
• Retired the subprocess workers. sam_worker.py, dino_worker.py, and tools/check_worker_isolation.py are deleted. sam_utils.py and dino_utils.py load Ultralytics / Transformers models in-process and cache them on SAMUtils / DINOUtils. Each inference runs on a short-lived QThread; the calling thread pumps its event loop via _run_sync so the public API stays synchronous-looking and the UI stays responsive. ADR-011 is marked Superseded by new ADR-013.

End result: ~1-2 s faster per SAM/DINO call on Windows (no subprocess spawn, no model reload), cleaner Linux runtime (Qt 6 native integration), Python 3.13/3.14 supported out of the box. Bumped 0.8.12 → 0.9.0 to signal the binding switch.

Why they're coupled

The subprocess workers existed to dodge WinError 1114 — a PyQt5 + Torch DLL load-order conflict on Windows + Python 3.14, documented in the old ADR-011. Migrating to PyQt6 eliminates that conflict (Qt6 reshuffled its DLL packaging), so the entire subprocess isolation layer becomes dead code. Doing both in one PR avoids paying the migration tax twice.

Phase 0 coexistence gate

tools/check_pyqt6_torch_coexistence.py imports PyQt6 → torch → torchvision → transformers → ultralytics in that order and constructs a QApplication (the platform plugin is what actually loaded the conflicting DLL in the original failure). Pass = subprocess removal is safe. Verified on Linux + Python 3.11 and Windows + Python 3.14.

What's also in the PR

The migration exposed a handful of latent issues. Rather than ship them as follow-ups against a freshly-broken PyQt6 build, they're addressed in this same PR — every fix has a one-line explanation in arc42 / ADR docs so future maintainers don't re-derive them. Highlights:

DINO + multi-dim images (closes #68):

• Detect All Images now flattens every loaded slice into the batch via _collect_dino_batch_work_items. The current implementation silently skipped multi-dim slices including the active one.
• _navigate_to_image_or_slice handles slice names in batch review (slices live in slice_list, not image_list).
• _refresh_dino_temp_for_current re-syncs image_label.temp_annotations from dino_batch_results on every image / slice switch. Without this, masks from the previously-reviewed image visually bleed onto every subsequent slice.
• Application-wide _DINOReviewEventFilter makes Enter / Escape work during review regardless of focus (otherwise QListWidget consumes Enter for itemActivated). See ADR-015.
• Auto-accept dropdown now honored by both Detect Current Image and Detect All Images — was batch-only before.
• DINO phrase editor: row 0 (class-name phrase) is now renamable; auto-prepend of original class-name in get_phrases_for removed so the rename actually takes effect.

Canvas (pan + zoom-to-cursor):

• Pan switched to event.globalPosition() — widget-local coords were absorbing half the cursor delta as the scrolled widget shifted under the cursor, producing effective half-speed pan.
• Cursor-anchored Ctrl+wheel zoom newly implemented. Post-zoom offset derived analytically from viewport().width() because self.width() is stale on zoom-out before layout settles.

Multi-dim TIFF:

• load_tiff reads tifffile.series[0].axes and pre-fills the dimension dialog with the right T/Z/C/H/W mapping. Previously the default_dimensions[-ndim:] slice silently degraded for ndim ≥ 5 — a 5D TZCYX TIFF produced 2560 one-row "slices" of wrong content on OK.

GPU memory:

• Tools → Unload AI Models (Free GPU Memory) now uses the full recipe: model.cpu() → gc.collect() → torch.cuda.empty_cache() + ipc_collect() + synchronize(). Without the explicit gc.collect(), Ultralytics' circular refs pinned the model on the device until the next GC tick — Task Manager / nvidia-smi showed zero drop. Dialog now also discloses that PyTorch keeps a per-process CUDA context that survives unload (requires app restart for full reclaim).

Mechanical PyQt6 cleanup:

• 15× surviving .exec_() → .exec() calls across QMenu / QDialog / QApplication / QMessageBox call sites in 6 files (would crash on right-click in the class list, on Project Search, on the DICOM converter, etc.).
• Missing import traceback in dino_merge_dialog.py (latent NameError).
• F2 (Snake game) moved from keyPressEvent to QShortcut(ApplicationShortcut) because QTableWidget consumed F2 for in-cell-edit before it bubbled.

Dark mode:

• Dark mode on by default at startup.
• Removed hardcoded background: #e0e0e0 from the DINO threshold table header and background: #f5f5f5 from the DINO status label — both punched bright rectangles into the dark sidebar.
• Added QRadioButton / QCheckBox / QHeaderView / QTableWidget / QSpinBox / QDoubleSpinBox / QComboBox / QGroupBox rules to soft_dark_stylesheet.py. Dataset splitter radio buttons (and a few other widgets) were rendering at the OS default which on Windows dark mode meant near-invisible selected-state indicators.
• Annotated-slice highlight changed from light blue (173, 216, 230) to muted steel-blue (58, 95, 140) with light text.

YOLO export:

• Exact-key match first in image_paths.get(image_name); substring fallback only after no exact match. "bee.jpg" in "honeybee.jpg" was True under the old substring-only lookup.

Threading + safety net

• _run_sync runs the inference callable on a QThread, captures any exception on the worker instance, and re-raises on the calling thread.
• _inference_in_flight module flag + InferenceBusyError exception serialise concurrent calls. A QMutex was considered and rejected — same-thread re-acquisition of a non-recursive mutex deadlocks, recursive defeats serialisation.
• apply_sam_prediction carries its own _sam_inference_in_flight flag at the call site, because the SAM debounce timer can fire while an earlier inference is pumping inside _run_sync. Defence in depth.
• _qimage_to_numpy always returns a fresh copy. The earlier alias-the-QImage-buffer pattern was a latent UAF in the fallback path that threading would make more dangerous.

Quality gate

Five senior-reviewer passes on the diff, each followed by a fix commit. Each pass surfaced real issues (silent KeyError in cross-project state carryover, row-0 rename illusion, missing update_slice_list_colors after auto-accept, orphan batch-review entries, etc.). Final verdict: "Mergeable with the P1 fixes applied."

Test plan

• All 65 pytest tests/ pass on Linux + Python 3.11 under QT_QPA_PLATFORM=offscreen
• Full app constructs and renders headlessly
• Phase 0 coexistence smoke test passes (Linux + Py3.11; Windows + Py3.14 verified locally)
• Manual UI smoke on Windows: SAM click-segment, SAM box, DINO single + batch (regular images + multi-dim slices), multi-dim TIFF loading + dimension dialog, project save/load roundtrip, dark-mode toggle, Tools → Unload AI Models, YOLO + COCO + Pascal VOC export, all Tool menu dialogs (DICOM converter, image patcher, image augmenter, dataset splitter, slice registration, stack interpolator, annotation statistics, COCO combiner, Merge COCO for Training), right-click context menus, F2 Snake easter egg, project search, project details
• CI green across Ubuntu / Windows / macOS × Python 3.10-3.13 — please confirm once CI runs

Architecture docs touched

• ADR-001 (Use PyQt5) marked Superseded → ADR-014
• ADR-011 (subprocess isolation) marked Superseded → ADR-013
• ADR-013 (in-process inference + QThread) new
• ADR-014 (PyQt5 → PyQt6 migration) new
• ADR-015 (application-wide event filter for DINO review shortcuts) new
• docs/05_building_block_view.md: SAMUtils block rewritten, DINO subprocess box dropped
• docs/06_runtime_view.md: SAM and DINO inference sequence updated for in-process path
• docs/08_crosscutting_concepts.md: new sections for Pan + Zoom Reference Frames, Dark Mode No Hardcoded Colors Rule, Releasing Model GPU Memory, DINO Temp Annotations lifecycle (single field, switch sync, batch over slices, navigation, event filter, auto-accept), Multi-dim TIFF Axis Defaults, Export Format Filename Matching
• docs/12_glossary.md: "Subprocess Worker" entry marked historical
• README.md, docs/01_introduction_and_goals.md, docs/02_architecture_constraints.md: PyQt5 → PyQt6
• PYTHON314_SETUP.md deleted (described migration as future work)
• TESTING.md: stale Py3.14/DLL section excised, Py3.13 added

Stats

15 files changed in the manual-testing fix pack on top of the original migration commits, with ~1000 lines net added — most of it diagnostic logging + the four new arc42 sections + the application-wide event filter. The subprocess machinery removal in the parent commits is the net deletion (~300 lines).

---

Happy to discuss any of the calls before merge — especially the dark-mode default, the dino_batch_mode semantics applying to both detect paths, and whether you'd rather keep the diagnostic [DINO] / [SAM] print logs un-gated (current state, per the manual-test session) or move them behind an env var. None of those are load-bearing for the binding migration itself.