Add equal-grid snapping for split panes

[brsbl]

Why

Users usually want a newly created split to return to equal-sized panes, but free resizing makes that canonical grid hard to recover precisely. This layer adds one shared equal-grid interaction to tab, page/thread, and thread/right-panel dividers without removing arbitrary sizing.

User-visible behavior

• Two siblings snap to halves, three to thirds, four to quarters, and larger sibling groups use the same equal-grid rule.
• Nested row and column split nodes own their grids independently; there is no diagonal snapping.
• Approaching within 12 px—or crossing a target during a fast drag—captures the equal boundary.
• Once captured, the divider holds through a 30 px release band. A second consecutive out-of-band sample confirms release after coarse pointer movement.
• A temporary full-span guide marks the active equal boundary.
• The behavior applies to right-panel tab splits, workspace page/thread splits, and the center divider between a thread and its right panel.
• Existing 15–85% bounds and arbitrary post-release sizes remain available.

Reliability and performance

• Grid and divider geometry are read once when a drag begins; pointer movement uses cached geometry.
• Center, tab, page, and thread dividers share the same free → captured → released state machine and resistance thresholds.
• The center-divider bridge previews adjacent flex sizes locally and commits once through the panel API on release, matching the internal split path.
• Release, cancellation, lost pointer capture, a zero-button sample, window blur, replacement drag, and unmount all clear the guide and resize state.

Screenshot evidence

All captures use Chrome for Testing 151.0.7922.71 at the same 1440×900 CSS viewport (DPR 2), the same one-file Split tab UX verification fixture, and one coarse pointer jump that crosses the equal boundary by 80 px. The before revision is exact parent head 1dbfc83bd587b98968d6b0a839da9f81751ce3b9; the after revision is exact PR head a128a9917ffe49588cb2c84d573f01113834e5a6.

Thread / right-panel divider

Before, the fast crossing follows the sampled pointer past center and leaves an arbitrary-width right panel with no guide.

After, the same crossing holds at 50/50 and renders a full-height guide through the owning split.

Vertical tab split

Before, the stacked Diff/Info divider lands 80 px beyond the equal-height boundary with no guide.

After, the same crossing holds both panes at equal heights and renders the full-width horizontal guide.

Validation

• Capture manifests record the exact parent/head revisions above, identical fixture and viewport, zero rendered error states, no guide and an 80 px overshoot before this layer, and exact 50/50 geometry with the correct axis-spanning guide after it.
• Regression coverage in this layer targets equal boundaries for multiple sibling counts and both axes, fast-crossing capture, shared 12/30 px hysteresis, confirmed coarse-sample release, center-divider single-writer behavior, persistence, and cleanup on every drag terminator.
• Tests, typechecks, lint, and other CI-equivalent checks run on the pull request through GitHub Actions; none were run locally for this screenshot audit.

BB-Thread-ID: thr_q8degf2y66

AGENT GENERATED: by GPT-5.6-Sol

[brsbl]

Closing this feature layer because split-resize snapping has been cut from the stack. PRs #2486 and #2487 remain unchanged.

[brsbl]

Reopened after correcting the interaction model: this now uses one shared midpoint grid per split surface with fast-crossing capture and magnetic release hysteresis.

[brsbl]

@slopcop Please review the full native PR stack #2486 → #2487 → #2488. The screenshot evidence has been audited and refreshed against each layer’s exact merge-base or parent-head and PR-head revisions, using matched fixtures and viewports throughout.

[bb-slop-cop]

🚨 SLOP COP 🚨 · review

I am SlopCop. I will review this pull request against the configured guidelines.

I will check security, code quality, performance, architecture, duplication, tests, and user behavior.

I will post one final report after the parallel checks finish.

[bb-slop-cop]

🚨 slopcop/review — The panel adapter applies the wrong size limits.

The shared resolver limits the leading panel to 15%–85%. These panels require 30%–76%, plus a collapsed state.

The adapter blocks the panel library's move events. Thus, the panels cannot reach their collapse thresholds.

A browser check showed a 15% preview. The panel then jumped to 24% after release.

Pass the panel limits and collapse rules into this path. Add tests for both limits and both collapse directions.

[bb-slop-cop]

🚨 slopcop/review — A pointer cancellation saves a temporary size.

pointercancel calls the commit path. The custom split dividers restore the original size after a cancellation.

Route this event to cancelDrag. Add a test that restores both flex values and preserves the panel state.

[bb-slop-cop]

🚨 SLOP COP 🚨 · review

Plain English summary

This change makes a divider pause at equal pane sizes. It adds the same guide and resistance to inner splits and right panels.

Findings

1. High — The right-panel adapter uses the generic 15%–85% split limits.

   The panels require 30%–76%, plus a collapsed state. The adapter also blocks the panel library's move events.

   Thus, a user cannot collapse either panel with a drag. Extreme previews also jump to a valid size after release.

2. Medium — A pointercancel event commits the temporary preview.

   A browser or device can cancel a pointer action. This event must restore the original size, as the custom dividers do.

Other review results

I found no security problem. I also found no new performance problem in the pointer-move path.

Three resize paths repeat pair state and release logic. A shared drag-session helper could keep their behavior consistent.

Verification

• The Turbo type check for @bb/app passed.
• The five focused test files passed all 99 tests.
• The full app suite passed 3,477 tests. Nine failures occurred in unchanged test files.
• A Chrome check reproduced both findings on the real right-panel handle.
• The final GPT-5.6 review gate confirmed both findings and returned REQUEST CHANGES.
• The author rebased the branch during this review. The two affected feature files remained identical.

Please fix both regressions before merge. I posted this as a comment-only review, as required.