fix(computer): emit one full WHEEL_DELTA per libnut scroll tick

[quanru]

Summary

• Move the libnut-fallback per-detent loop into ComputerInputDriver.emitScrollDetents so scroll dispatch lives next to the rest of the input primitives.
• Emit exactly one platform-correct detent per libnut.scrollMouse call (120 on Windows = one WHEEL_DELTA, 1 elsewhere) and pace calls with LIBNUT_FALLBACK_TICK_DELAY_MS so blink doesn't coalesce them into a single tick.
• Add input-driver-scroll.test.ts covering per-call magnitude, pacing, horizontal pass-through, and destroy-mid-scroll rejection.

Why

libnut.scrollMouse(x, y) is not a portable pixel API:

Platform	y semantics
macOS	CG pixel scroll delta
Linux	XButton4/5 press-release pairs (1 = one wheel notch)
Windows	Forwarded directly to MOUSEEVENTF_WHEEL's mouseData, where the unit is WHEEL_DELTA (= 120) per detent

The old fallback computed ticks = ceil(distance / 100) and made a single scrollMouse(0, ticks) call. On Windows that sends mouseData < 120 for any normal request — well below one detent — and Chromium's WheelEventQueue silently accumulates and drops it, which is exactly what we saw in Lark / Feishu Electron clients reporting "scroll did nothing". macOS still works because the phased-scroll helper takes the path; Linux happens to work because 1 detent ≈ 1 notch already.

Edge scrolls (scrollToTop etc.) had the same problem in the libnut path: the unit vector was hard-coded as [0, 10] etc. — fine on Linux as 10 notches, near-useless on Windows as mouseData = 10.

After this change, both the directional fallback and edge fallback go through emitScrollDetents, which fires LIBNUT_FALLBACK_DETENT_AMOUNT per call — 120 on Windows, 1 everywhere else — and the destroy-mid-scroll rejection wired into the pending-input-delay set keeps shutdown clean.

Test plan

• pnpm run lint
• npx nx test computer — 53/53 passing, including 4 new tests in input-driver-scroll.test.ts
• npx nx build computer
• Windows manual verification in Lark / Feishu Electron client (single scroll request now visibly moves the page)

Related

• Pairs with fix(computer): self-heal phased-scroll +x bit lost in npm tarball #2574 (chmod self-heal). Both are scroll-reliability fixes in the libnut/phased-scroll path; this one is the Windows half.
• Supersedes branch fix/computer-scroll-win-fallback (commits 905124883, d020e1de4) — those predated the ComputerInputDriver refactor on main; this PR is the rewrite.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ec4694cbb1

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve enough detents for edge-scroll fallback

When phased scrolling is unavailable and the libnut fallback is used (Windows/Linux, or macOS with the libnut keyboard driver), scrollToTop/scrollToBottom now emits only SCROLL_REPEAT_COUNT one-detent ticks. With the new platform-correct detent amount, that is roughly 10 wheel notches (about 1,000px by this file's LIBNUT_FALLBACK_PIXELS_PER_DETENT), whereas edge scrolls are documented and implemented in the phased path as a 50,000px boundary scroll; common long pages will stop far short of the top/bottom. The old Linux fallback effectively sent 100 notches, and the Windows fix should scale the number of detents rather than reducing the edge action to a single-screen-ish scroll.

Useful? React with 👍 / 👎.

[cloudflare-workers-and-pages]

Deploying midscene with    Cloudflare Pages

Latest commit:	9a63ffb
Status:	✅  Deploy successful!
Preview URL:	https://9de9d4ac.midscene.pages.dev
Branch Preview URL:	https://fix-computer-libnut-detents.midscene.pages.dev

View logs