by-links-layout-publishes-overlapping-positions

Summary

Makes the record-graph's by-links arrangement come to rest and makes the overlap gate measure it. The by-links layout settles islands with a spring layout that has no collision pass and packs two rim rows onto a circle whose circumference is smaller than the sum of the record arc-widths, so it publishes overlapping positions by construction; and the overlap gate counts only the by-date coil, so the overlap ships green. This spec adds one post-settle collision-resolution pass covering islands, pairs and rim under a single packing rule, and widens the gate to measure the by-links layout too.

Scope

In:

• internal/core/site/layout.go — the byLinks builder (lines 366–469): a post-settle collision pass after island/pair/rim placement, and the rim-row packing so the circumference matches the summed arc-widths.
• The overlap counter, today computed only inside coil() (layout.go:332–339): measure a.Links as well, in one coordinate space.
• internal/core/site/build.go (the Overlaps result field, lines 299–301,
  1. and the human-facing print in internal/surface/cli/site.go:197.
• Tests in layout_test.go and build_test.go.

Out:

• The renderer half is already fixed (iss-2608231243286557 — the collision resolver no longer demands 1.5 units of padding); the JS chart emitted from graphpage.go is unchanged.
• The by-date coil (coil(), layout.go:264–360) already settles and is not touched beyond sharing the widened overlap counter.

Approach

The two causes. byLinks (layout.go:366–469) settles islands via springLayout (called line 418, Fruchterman-Reingold at 511–587) then decompress/scale/place (419–422) with no node-radius collision pass — cause one. The two rim rows (445–468, radii rimRow0=0.905/rimRow1=0.97) sum each record's arc-width w = 2*a.Radius[i] + coilGap into total (456) and map each onto a circle at radius rr, but the summed widths exceed the circumference at that radius — overpacked by construction, cause two. Prior local patches (relaxing linked records alone; a rim-row growth rule that pushed the outer radius from 0.97 to 1.83) were measured and reverted because they changed the picture without settling it.

One packing rule. After island, pair and rim placement (after line 468), run a single post-settle collision-resolution pass over all by-links nodes together — the coil already demonstrates the forbidden-interval idea (layout.go:296–317). The pass iteratively separates any pair whose centre distance is less than the sum of their published radii (plus the same epsilon the coil uses), nudging both along their separation vector, bounded by an iteration cap, until no pair overlaps. Because islands, pairs and rim are resolved in the same pass and the same coordinate space, no region is packed against a promise a later region breaks. The rim rows are seeded so their circumference is at least the summed arc-widths (the row radius is derived from total, not fixed), so the pass starts from a near-feasible configuration rather than an over-packed one and converges without inflating the outer radius the way the reverted growth rule did.

Widen the gate, one coordinate space. The overlap count lives only in coil() (layout.go:332–339). Extract the counting loop into a helper countOverlaps(points []Point, radii []float64) int and call it for both a.Coil and a.Links, summing into a.Overlaps (Arrangements struct, layout.go:99–137). The known trap is that node positions are unit-disk-normalised while radii are in reference pixels (the mismatch that produced spurious overlaps in iss-2608231322321751); the helper normalises both into one space before comparing, so the count is real for by-links as it already is for by-date. build.go (299–301, 465) and the CLI print (site.go:197) surface the combined count as they do today.

How it satisfies each acceptance criterion

• By-links overlaps after the spring settle are separated; the published layout reports zero overlaps — the post-settle pass over all nodes. Test: over the record corpus, build the by-links layout and assert countOverlaps(a.Links, a.Radius) == 0 (a new TestByLinksNeverOverlaps mirroring TestCoilNeverOverlaps, layout_test.go:40).
• The overlap gate measures the by-links layout, not only the by-date coil — a.Overlaps now sums both arrangements. Test: TestBuildLayoutDoesNotOverlap (build_test.go:322) is extended to assert the count reflects a.Links.
• A by-links layout that still contains overlaps flags red — because the gate now counts a.Links, a residual overlap makes a.Overlaps != 0, which the CLI prints as a non-zero/red result (site.go:197). Test: inject an overlapping by-links fixture and assert a.Overlaps > 0.
• The by-date arrangement continues to pass — the coil's forbidden-interval packing is unchanged; the shared counter measures it exactly as before. Test: TestCoilNeverOverlaps stays green.

Decisions

One packing rule over all node classes, chosen over per-region patches. The two measured local fixes each moved the picture without settling it; the intent resolved that a correct fix is a design of the arrangement, not a patch. The single post-settle pass in one coordinate space is that design. The rim-row radius becomes derived from the summed arc-widths rather than fixed, so the pass converges without the outer-radius inflation the reverted growth rule caused.