Make comparison controls and room actions readable on phones

[HomenShum]

A person running the comparison on a phone could lose part of the model/count controls or the completion badge, and a room participant could lose the Invite button beyond the header. This change wraps existing control groups and room actions and keeps the complete numeric fields and status visible in the tested views.

Only layout classes and the comparison control group's accessible name change in the two existing owners. Playback, room handlers, state reducers, shared primitives, provider configuration and dependencies remain unchanged. The new HANDOFF provides keyless setup and the exact limits of the retained evidence.

Actual before and after

• Phone completion before / after.
• Enlarged text before / after.
• Room header before / focused Invite after.
• Evidence index includes normal desktop comparisons, all required viewport/font records and the failed scroll approaches.

Verification

• Local npm run doctor, npm run proof and npm run build pass. Doctor covers server/client/Convex types and 56 citations; proof covers 55 tests in 9 files plus the separate target-12 CLI smoke.
• One actual final browser run reaches 100/100 at 125.736 seconds and re-enables Run at 126.829 seconds within the unchanged 300-second budget. Its native request reports deterministic execution with no executed model.
• 337 scoped checks and 94 captures cover seven viewport pairs, enlarged computed text at 390/1440, reduced motion, native pointer and both keyboard directions, full control boxes, numeric spinners, and repeated Invite/State interaction. Two local contexts retain the expected missing-key error behavior after correction and a separate Start.
• Independent review verifies exact two-owner layout substitutions and actual pixels. The handoff preserves selected raw records and required image sets, explicit omissions, source hashes and a standard-library verifier. Six log copies use .log.txt so normal Git publication retains them without changing ignore policy.
• Current shared CI and conformance passed on first attempts. CI ran server/client types, 56 citations and 55 tests / 9 files, plus a target-10 CLI smoke that commits nine steps; it is not completion proof. GitGuardian actually scanned one commit with no secrets found. Shared CI omits build, this browser matrix and Convex typechecking; those remain separately bound local evidence.

Limits and rollback

The first after-pass had an insufficient interior-only control check; later forward/reverse Tab failures are retained. Ordinary wrapping removes the horizontal focus-scroll dependency without adding focus handlers.

Eleven final captures still contain off-viewport descendants in other detail regions. Selected-option ellipsis, detailed panes, short-height reading, full visual/accessibility grades, native devices/zoom, audible speech/microphone, actual provider/agent execution, hosted Convex and durable restart remain unverified. The baseline has three high development-tool package entries; this layout PR changes no dependencies. The local server still binds an unspecified host.

This is a reversible layout/documentation change with no schema migration. Revert the commit to roll it back. No production interaction or deployment claim is made.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
local-collab-mvp	Ready	Preview	Sep 5, 2026 11:06pm UTC