fix(web): floating preview resize ends cleanly without scale jumps#23002

Problem

Fixes interrupted floating-preview resize gestures reported in #22988 (GitCafe mirror #22998). A lost release could leave a resize active, so later buttonless hover enlarged or moved the preview and left it stranded over the conversation. Resizing also mixed a captured frame with the live source aspect ratio, allowing a source update to change its size during the gesture.

Change

Capture fractional geometry from the preview element at pointerdown and freeze that frame's ratio for the gesture. Each handle owns its pointer capture. Release, cancellation, capture loss, blur, and missing primary-button state end the gesture; idle moves do nothing. A new pointerdown discards stale capture state if a handle disappeared. Finish clamps position through the existing resolver while preserving the user's chosen width and interaction type.

Scope and approval

Fixes #22988

Triaged bug via the automated issue flow; GitCafe mirror #22998.

This is gesture-only: no chat/composer inset, details-card folding, anchor, or placement-policy changes. #16831 supplied the interrupted-gesture findings; its removal of intentional inset/folding behavior was why it was closed unmerged, and those layout changes are excluded here.

Verification

Evidence re-captured as narrated before/after recordings per EVIDENCE.md. Compared before 680152e27b with after db94329f89. Captured all nine rows of the fixture matrix: evidence index · independent proof audit: SUFFICIENT.

Every recording pair has matching scripted steps and decoded-video timestamps above it. Normal gestures use native drags above the minimum size. Interruption/transition rows acquire native pointer capture before deterministic event injection. Physical outside-window release is explicitly excluded.

Blur now releases capture and holds at 355×222; before continues to 460×288 despite blur. Missing release now holds at 280×175; before continues to 318×199.

Adverse transition finding (needs maintainer review): R5 after ends with its bottom edge at 656 instead of 612, a 44 px drift absent before. R5/R6 rendered aspect ratios also change during source resizing. These recordings document failed expectations, not passing transition behaviour. Normal samples use whole-pixel geometry and do not independently distinguish the fractional-geometry fix or transient self-correcting jumps.

Out of scope: other handle/termination/theme combinations and drag/pill/actions keep shared-controller sampling within eight pairs. Physical outside-window release, touch and native Electron/device/PDF/mobile require separate input or environments. Providers, contracts, backend and relay/tunnel paths are unchanged.

Environment: Linux x86_64, Chromium 154, DPR 2, loaded system UI sans-serif; identical isolated seeds and local static fixture. CSS viewports 1440×900 and 390×844, light/dark. Per-capture UTC clocks appear in the index. MP4 rasters are 2880×1800 and 780×1688; static PNGs show the full desktop surface downsampled to 1280×800.

Static before:

S1-before.png

Static after:

S1-after.png

Static delta: 10,760/1,024,000 nonzero pixels; 962 above 16/255; maximum channel delta 71. All changes lie inside streamed content. S1 includes unmasked zero-threshold ×8 and >16/255 diff images. The auditor reproduced these metrics exactly.

What else changed: the PR extracts the pointer controller, wires blur/lost-capture termination into the shared shell, adds controller tests and makes an empty-obstacle constant private. No styles, transports, contracts, providers or backend changes. Capture-only differences include stream rasterization, annotation timing and local IDs.

Capture: Codex gpt-6.1-sol; proof audit: Claude Opus 5.5.

Model/harness: GPT-6.1-Sol through Codex in T3 Code; independent GPT-6.1-Sol/Codex and Claude Opus 5.5 reviews and evidence audits.