fix: canonicalize the frozen-row boundary across all comparison sites

[6pac-ai]

⚠️ Merge order: this PR assumes #1266 and #1267 are merged first. It has no textual conflicts with either, but it was authored and validated against a master that includes them — please land those two before this one.

The bug (Q3 in discussion #1247)

The frozen-row boundary was compared differently at six sites, contradicting each other and the render split (rows >= actualFrozenRow render in the bottom canvas):

Site	Drift	Observable
appendRowHtml frozen class	row <= frozenRow (a count compare)	bottom mode: the first rows classed .frozen, the actually-pinned rows never; top mode: off-by-one
cleanupRows / cleanUpCells	<= actualFrozenRow	first scrollable row never evicted/cleaned in top mode
_getContainerElement (getCanvasNode/getViewportNode)	>= actualFrozenRow + 1 in top mode	returns the TOP pane for a row whose DOM lives in the BOTTOM canvas — overlays/menus positioned via the public API mis-target exactly the boundary row
scrollRowIntoView	actualFrozenRow - 1 boundaries	bottom mode: refuses to scroll the LAST scrollable row (navigateToPos, one screen away, already compares correctly)

The fix

Two predicates matching the render split — isBottomBandRow (row >= actualFrozenRow) and isFrozenRowIdx (frozenBottom ? row >= actualFrozenRow : row < actualFrozenRow) — with every site routed through them. Net −5 lines.

Downstream check (executed)

slickgrid-universal maintains its own fork of this grid (verified by clone + grep — it does not consume this file at runtime), so this cannot break it. Its fork carries the same isBottomSide off-by-one and may want to port the fix.

Test

cypress/e2e/quirk-frozen-row-boundary.cy.ts — self-hosting (harness via cy.intercept; no example-page dependency), one check per drifted site plus a frozen-row-stays-cached control. Pre-fix it fails on every drifted site with the predicted signatures (frozen class [0,1,2,3] in both modes; getCanvasNode returning the top canvas while the row's DOM is in the bottom; scrollRowIntoView(996) leaving viewport.top=0; boundary row never evicted); post-fix all pass. Frozen-family suites including the merged #1255/#1258 quirk specs ran 27/27.

⚠️ Temporary example page

examples/example-quirk-frozen-row-boundary.html (tinted .frozen rows + pane-lookup/scroll readouts) is intended to be deleted before merge — the cypress test is fully independent of it.

(Quirks-triage remaining-items wave: see discussion #1247; siblings #1266, #1267.)

[ghiscoding]

I replicated this PR as well but it seems to cause a constant error in Cypress test example-auto-scroll-when-dragging.cy.ts which would probably need to be reviewed before merging

I asked copilot to summarize why it might be failing because I have the same example and test, and it replied with:

Likely CI flake, not grid logic bug.

Short version you can post:
The failure is probably Cypress actionability during drag start: the target cell is briefly covered in headless CI, so mousedown/mousemove gets blocked. The drag helper should be hardened by forcing mousedown + first mousemove, and ending with mouseup on body (then grid) to avoid stale drag state when pointer leaves the container. This fixed the same frozen-grid test on my side without changing assertions.

Side note, I actually had to modify my similar Cypress test as well (example17.cy.ts in my case), so you might want to ask Claude to replicate my Cypress changes in this PR as well to make the tests pass and avoid flakiness.
See my PR for reference: ghiscoding/slickgrid-universal#2706

[6pac-ai]

Thanks for the pointer @ghiscoding — reproduced here too, and it was indeed failing constantly on this PR's CI (example-auto-scroll-when-dragging.cy.ts, frozen-grid drag test, ensureElIsNotCovered).

Root cause checks out as test actionability rather than grid logic: the covered target is a live grid cell, and the covering element is the example page's own descriptive text (<div>better to multiple of MIN</div>) — with scrollBehavior: false, the drag's target coordinates now land where the page text overlaps the grid, because the canonicalized boundary changes where the auto-scroll legitimately lands. The dedicated frozen-boundary spec (which asserts canvas routing and classification directly) passes in the same runs, so rows aren't being misrouted.

Ported your #2706 hardening to cypress/support/drag.ts in fce4bcb — force on the mousedown and first mousemove in dragStart (force was previously missing entirely here) plus the body mouseup in dragEnd, assertions unchanged. The failing test now passes 5/5 consecutive local runs and the full suite is green. (The example17 median-sampling change wasn't needed on this side — our equivalent MAX-interval test is already skipped.)