[dbose]

Problem

Ticking a subject-area filter (e.g. one domain of 12 — 6 of 79 entities) left the viewport zoomed to the whole model: the filtered cards showed as tiny specks spread across empty space, and with onlyRenderVisibleElements they could be culled entirely (0 nodes rendered).

Root cause

Every entity stays in the node array when filtering (hidden ones flagged, never removed), each keeping its full-model position. The subset was laid out compactly (dagre over the visible nodes), but the post-layout fitView() framed all 79 — so the filtered cards landed in a corner of that huge bounding box.

Fix

When filtering, frame only the visible subset: compute the bounding rect from the visible nodes' positions + known sizes and fitBounds() to it — the same position-based approach the search-framing effect already uses, because fitView({nodes}) can't measure an off-screen/culled node. Unfiltered behaviour is unchanged (fit all, only when no saved layout); clearing the filter re-fits the whole model.

Also: dagre acyclicer: 'greedy' so FK cycles (self-refs, mutual refs — common in reverse-engineered/erwin models) don't degrade the layered "sources higher, derived lower" ordering.

Verification

Reproduced on a 79-entity / 12-domain model (matching the reported erwin_import case):

• Before: filter to a 6-entity domain → viewport fit to the 79-node extent, 0 nodes rendered (culled).
• After: all 6 cards render legibly, framed at scale ~0.56; "Show all" re-fits the full model.

tsc clean; canvas Vitest 55 passed; ruff clean; server suite green.

Scope note

This is the critical usability repair (frame the subset). The fuller "base tables higher, facts/dims/aggregation lower" hierarchical-rank layout (declared-layer → topological-rank → FK-direction fallback, with a TB direction option) is a separate follow-up.