fix(diagram): collapsed-to-collapsed edges carry no per-foreign-key style

[dimitri-yatsenko]

The problem

A collapsed node stands for a set of tables. An edge between two collapsed nodes therefore represents a bundle — every foreign key between the two sets of tables, drawn once. _apply_collapse built that edge like this:

if not new_graph.has_edge(new_src, new_dest):
    new_graph.add_edge(new_src, new_dest, **data)

First member wins, no aggregation. So multi (cardinality), primary (primary-vs-secondary) and aliased (renamed) — each a property of one foreign key — got attributed to a set of them, and the resulting style depended on graph traversal order.

Two schemas with identical structure, differing only in declaration order:

before:  order 'a' -> penwidth=2    style=solid
         order 'b' -> penwidth=0.75 style=dashed
after:   order 'a' -> penwidth=2    style=solid
         order 'b' -> penwidth=2    style=solid

It is stable within a process, which is why it went unnoticed: the same script always draws the same picture, but two environments can disagree, and a committed SVG is not reproducible.

The rule

An edge between two collapsed nodes claims none of the per-foreign-key properties: every bundle edge renders identically (solid, penwidth 2), in both the graphviz and mermaid paths.

This is about collapsed nodes, not schemas. A collapsed node may stand for any subset of tables — Diagram(schema).collapse() + Diagram(OneTable) leaves the unexpanded remainder as a group — and the rule applies to any edge between two such nodes.

An edge with only one collapsed end is unchanged. It still names a single table, so its cardinality and primary-vs-secondary styling stay meaningful and are preserved:

one end expanded, multi-valued FK -> penwidth=0.75 solid
one end expanded, secondary FK    -> penwidth=0.75 dashed

Why this shape

#1533 established that weight encodes cardinality and only cardinality. This follows from it: a bundle has no single cardinality to report, so it reports none. The alternative — aggregate over the bundle (thick only if every member is 1:1, dashed only if every member is secondary) — is well-defined and order-independent too, but it invents a second meaning for the same visual cues, one that holds only at the collapsed level. Uniform styling keeps one meaning per cue.

Tests

tests/integration/test_diagram_bundle_edge.py:

• test_bundle_edges_are_uniform — a bundle mixing a primary 1:1 foreign key with a secondary one renders uniform.
• test_bundle_style_is_order_independent — declaring the members in the opposite order changes nothing. This is the regression that would have caught the original bug.
• test_one_collapsed_end_preserves_edge_style — an edge into an expanded table keeps its own thin weight, so the bundle rule cannot leak.

All three pass on both backends, alongside the existing cardinality guard from #1533:

$ pytest tests/integration/test_diagram_bundle_edge.py tests/integration/test_diagram_edge_weight.py -q
8 passed in 15.09s

(pip install "testcontainers[mysql]" is needed for the session fixtures; without it these error at setup rather than fail.)

Downstream

Found while adding a generator for the diagrams in datajoint/datajoint-docs#265, where the prose claimed "a heavier edge carries a larger bundle" and the figure beneath it disagreed — lab → session and session → imaging bundle two foreign keys each and rendered 0.75 and 2. That prose is corrected separately; this changes what the renderer draws.

Once this ships, collapsed figures committed in datajoint-docs need regenerating — tracked on datajoint/datajoint-docs#246, which already covers the release-time re-verification step.

[MilagrosMarin]

Verified the implementation. _apply_collapse cleanly separates the two cases — both ends collapsed strips primary/multi/aliased and marks bundle, one end collapsed preserves the data untouched — and both the graphviz and mermaid paths honor it. test_bundle_style_is_order_independent genuinely reverses the declaration order rather than asserting on a single arrangement, so it's a real regression guard for the original bug.

Agreed on uniform-over-aggregate. Aggregation is well-defined, but it gives thick/dashed a second meaning that only holds at the collapsed level, and #1533's whole point was that weight means one thing. Reporting nothing is the honest read of a set that has no single cardinality.

One follow-on, not about this PR's files: bundle edges render penwidth 2 solid, which is exactly what the docs spec table assigns to "1:1 dependency" + "primary foreign key" — and that section currently says "line weight encodes cardinality, and only cardinality — it is binary," with no mention of bundles. This PR ships no docs, so the obligation lands on datajoint-docs#265's § Edge Styles. Worth a row there before this ships, so the normative table isn't contradicted by the renderer. Context disambiguates in practice (both endpoints are collapsed group boxes), so it's spec coverage rather than a rendering problem.

Approving.