Poll for self-driven snow frames instead of a fixed sleep

[TheGreatAxios]

Summary

• The idle landing snow test used a fixed 3s sleep then a single flush. Under CI load the mount's setInterval can fire after that window, so flush finds nothing scheduled and the capture is still the mount frame — a flake that fails unrelated PRs.
• Replace the fixed sleep with a poll up to a 5s deadline: flush → check mark rows → 50ms yield. Nothing in the wait path calls paintLanding/renderMark/renderOnce, so a frozen timer still fails the assertion. Early exit drops average suite cost below the old fixed 3s wait.
• Path is under the post-flatten src/tui/landing.test.ts (rebased onto main after Flatten the OpenTUI shell into src/tui #435).

Fixes CL-5766.

Test plan

• bun test src/tui/landing.test.ts (25 pass; snow test ~166ms via early exit)
• bun test ./src ./tests ./evals (4346 pass, 0 fail)
• CI green on this PR

[linear-code]

CL-5766

[TheGreatAxios]

Self-review

Root cause: fixed 3s sleep + single flush races under CI load when the mount-scoped setInterval fires late. flush only drains already-scheduled renders; if the timer has not fired yet, the capture is still the mount frame and the assertion fails with two identical marks.

Fix: poll until markRows differ or a 5s deadline: flush → compare → 50ms yield. No paintLanding / renderMark / renderOnce in the wait path, so the load-bearing property is intact (frozen timer still fails). Early exit keeps average wall time below the old fixed 3s sleep (local snow test ~166ms).

Scope: single file src/tui/landing.test.ts on post-#435 main. No product code change.

Risks: worst-case wait is 5s instead of 3s if snow never moves (already fails the test). Under normal load this should be cheaper. Ablation property preserved by design — nothing pumps frames.