fix(vt): disable alternate screen scrollback

[armenr]

• I have read CONTRIBUTING.md.
• I have created a discussion that was approved by a maintainer (for new features). Not applicable: this is a bug fix.

Fixes #947

Summary

Disable the alternate screen's scrollback buffer when constructing an Emulator and extend the existing alternate-screen regression to assert that the hidden buffer is absent after scrolling.

NewScreen enables a 10,000-line history by default, and NewEmulator uses it for both screens. The emulator's public scrollback methods expose and configure only the primary screen, so alternate-screen output currently clones rows into a buffer callers cannot inspect, bound, or clear.

Before and after

Before this change, the focused regression fails on current main:

$ go test -count=1 -run='TestScrollback/alt_screen_does_not_have_scrollback' ./...
--- FAIL: TestScrollback (0.00s)
    --- FAIL: TestScrollback/alt_screen_does_not_have_scrollback (0.00s)
        scrollback_test.go:118: expected nil alt-screen scrollback, got max=10000 len=6
FAIL

After this change, the same regression and the complete vt suite pass.

Why this fix

Screen.SetScrollback(nil) is already the documented way to disable history. Applying it once to scrs[1] keeps primary history and visible terminal behavior unchanged while preventing unreachable alternate-screen retention. It adds no API and does not alter parsing, resizing, or screen switching.

In a separate SwitchBoard integration measurement on an earlier x/vt revision with the same retention path, a 2 MiB alternate-screen flood retained 104,234,528 bytes of incremental live heap for one untouched emulator versus 6,025,248 bytes with this constructor change. Those workload-specific numbers are supporting evidence, not an upstream benchmark or memory guarantee.

Validation

Run on Linux x86_64 with Go 1.26.5-X:nodwarf5, from vt/:

• gofmt -l . — clean
• git diff --check — clean
• go mod verify — pass
• focused regression on unmodified main — fails as expected with hidden max=10000 len=6
• focused regression after the fix, 100 repetitions — pass
• go build -v ./... — pass
• go test -count=1 -race -covermode=atomic -coverprofile=coverage.txt ./... — pass, 41.2% statement coverage
• go vet ./... — pass
• Charm's current golangci.yml with golangci-lint v2.12.2, restricted to changes since upstream/main — 0 issues
• compile-only go test -c checks — pass for Linux/arm64, Darwin/amd64, Darwin/arm64, and Windows/amd64

The unrestricted shared lint command currently reports two unrelated findings already present on main (vt/emulator.go's unused nolint:gosec directive and vt/csi.go's QF1012). Upstream's current-main vt workflow run 31709584952 fails on those same two findings; this patch adds no lint findings.

Compatibility

• No public API changes.
• Primary scrollback behavior and its configurable size are unchanged.
• Visible primary/alternate screen behavior is unchanged by the existing test suite.
• Callers that somehow relied on the private, inaccessible alternate buffer cannot observe it through Emulator; if alternate history is desired later, it should be exposed and bounded explicitly.