v3.0

Fix: dual-spec BOFF seat cluster matching (warp_dialog.py)

When a seat like "Commander Engineering-Temporal Operative" had all recognised abilities
from the spec profession (Temporal), _find_cluster('Engineering', 'Temporal') failed because
c_primary == 'Temporal' ≠ 'Engineering'. Added fallback: if the base-profession search fails
for a seat with a spec, also try _find_cluster(spec_prof, None). Covers all dual-spec seat
combinations (Tactical-Temporal, Science-Command, etc.) symmetrically.

Fix: X-position gap detection for BOFF slot index assignment (warp_dialog.py)

Replaced sequential rank_slot = 0, 1, 2, … filling with position-based index mapping.
New _slot_indices_from_x(cluster_items, rank) helper:

• Computes minimum X-gap between detected ability icons as the baseline "1 slot" step
• A gap of ~2× step → slot index jumps by 2 (1 missing slot between)
• Single ability defaults to slot 0
• Slot indices are logged alongside ability names for diagnostics (slots=[0, 2, 3])
  Previously an empty slot 1 between two detected abilities caused abilities to land on slot 1 and
  slot 2 instead of slot 0 and slot 2.

Feature: empty / inactive BOFF slot position detection (layout_detector.py, warp_importer.py, warp_dialog.py)

_fill_boff_gaps(bboxes_abs, img, icon_est, x_min, max_slots=4) — new method on LayoutDetector:

• Called after _find_icon_bboxes_in_strip in both OCR (Strategy A) and color (Strategy B) paths of _detect_boffs
• Scans expected grid positions at step intervals from the leftmost active icon
• For each position without an active bbox: extracts a crop and calls _classify_cell() → 'empty' or 'inactive'
• Returns 5-tuples (x, y, w, h, state) — active icons unchanged, virtual positions added
  warp_importer._process_image():
• Detects 5-element bboxes before the normal bx, by, bw, bh = bbox unpack
• For state in ('empty', 'inactive'): creates RecognisedItem(name='__empty__'/'__inactive__', conf=1.0) directly — no icon matcher call
• Virtual items flow into BOFF clusters alongside active abilities
  warp_dialog.py Phase 3:
• VIRTUAL_ITEM_NAMES check before boff-cache lookup — silent continue, no "not in cache" log noise
• Virtual items contribute their X-position to _slot_indices_from_x for precise gap anchoring even when only 1 active ability is present in a seat

Training pipeline for __empty__ / __inactive__ crops

sync.py has no filter on virtual names — confirmed __empty__/__inactive__ crops are uploaded to
HuggingFace alongside normal equipment crops. admin_train.py includes them as regular EfficientNet
classes. When the model matures, icon_matcher will output these labels directly, making gap detection
independent of _fill_boff_gaps heuristics.