fix(Tokenizer): avoid blank trailing input row - #6057

Addressed in dc1e1d0. The empty compact input now keeps the measured end-lane reserve as logical outer spacing, so its flex base remains compact while its hit rectangle stops before Clear all. I added real-browser geometry guards for both LTR and RTL that require the input to stay on the token row, retain positive width, and not overlap the clear button. Both new guards pass in CI: https://github.com/facebook/astryx/actions/runs/33931922038/job/101212993760

Could you take another look?

You were right—the previous guard raced the ResizeObserver and could measure before the end-lane reserve was applied. Addressed in 3705ac0:

• the empty input is now a minimal, non-growing flex item; the wrapper remains the larger click-to-focus surface
• the guard waits for the measured lane before asserting
• LTR and RTL each require one-row alignment, positive input width, no Clear all overlap, a successful center-point browser hit test, and focus after click

Both hardened guards pass in Chromium: https://github.com/facebook/astryx/actions/runs/33933639943/job/101217896127

I also replaced the PR evidence with apples-to-apples captures using the same three filters, 600 CSS-pixel fixture width, and 2x output as the original report. Could you take another look?

Agreed—the 4px in-flow target still had an impossible boundary case: at some widths it had to overlap the clear lane or wrap. I replaced that approach in a0f64a2.

While empty, the combobox is now an absolutely positioned first-row background hit surface bounded one pixel before the measured end lane. It consumes no flex width, and pills plus end controls are layered above it. Typing restores the normal in-flow input.

The guard now waits for ResizeObserver before measuring, then verifies in both LTR and RTL: one-row alignment, positive input width, no Clear all overlap, correct hit-testing at the final pill and clear-button centers, an exposed point that hit-tests to the input, and focus after clicking that point. Both variants pass in Chromium: https://github.com/facebook/astryx/actions/runs/33935679227/job/101223840998

I also replaced the PR explanation and confirmed the before/after captures are apples-to-apples: identical filters, 600 CSS-pixel fixture width, 2x output, and the after capture was taken only after the served iframe hash matched a0f64a2. Please take another look.

Addressed on 00ebfea84.

I first tightened the guard and reproduced the reported failure in Linux Chromium with pointer: coarse: the empty combobox overlapped the effective Clear all hit target by exactly 1.00px in both LTR and RTL (failing run).

The empty input now reserves the clear button's full 2px coarse-pointer outset plus the existing 1px separation. The durable browser guard runs both directions with touch enabled, reads the ::after hit-area geometry, verifies all pills stay on row one, verifies the input stays on that row, checks non-overlap and hit-testing, and clicks an exposed input point to confirm focus. The final clean-head run passes.

I also replaced the mismatched screenshots in the PR description with an A/B captured in the same Chromium run using identical data, theme, viewport, scale, and fonts.