fix(tui): guard degenerate scroll region when one row remains above viewport

[ritammehta]

Summary

🤖 Generated with Nori

• insert_history_lines emitted an invalid ESC[1;1r DECSTBM when exactly one row remained above the inline viewport (area.top() == 1). Terminals ignore invalid DECSTBM and fall back to a full-screen scroll region, so each inserted history line scrolled the live viewport itself — persistent display corruption, since ratatui's diff buffer believed the viewport was intact and never repainted it. Large completed-turn outputs make this state both more reachable (tall viewports) and more destructive (one full-screen scroll per line). This is the unguarded sibling of the top()==0 fix from fix(tui): Guard against degenerate scroll region when viewport fills screen #358.
• The guard now covers area.top() <= 1: warn, stay cursor-position-neutral, persist any viewport move made by the preceding scroll-down branch, and return false so Tui::draw retains pending lines. The deferral is self-healing — the next insert scrolls the viewport down further and succeeds (covered by a retry-convergence test).
• Four VT100-backed regression tests (the vt100 parser implements the same degenerate-DECSTBM full-screen fallback as real terminals, verified against its source) plus nori-rs/tui/docs.md updates.

Follow-ups found during investigation (not in this diff)

• tui.rs scroll_region_up(0..area.top(), …) on viewport expansion emits the same invalid ESC[1;1r when area.top() == 1; masked by the subsequent terminal.clear(), but can push stale viewport rows into scrollback.
• The MAX_PENDING_LINES (1000) trim in Tui::draw runs on every draw before insertion is attempted, so a single >1000-line batch loses its oldest lines even when insertion succeeds — contrary to its comment ("while blocked").
• Architectural: VS Code terminal, Windows Terminal, tmux, and Zellij discard region-scrolled lines instead of moving them to scrollback (same open upstream family: openai/codex #27644, #24849, #15380); upstream mitigates via resize-triggered transcript reflow (#18575), which this fork predates.

Test Plan

• TDD: three repro tests written first and confirmed RED against the unfixed code (viewport rows visibly overwritten by inserted lines), GREEN after the guard
• cargo test -p nori-tui — 1349 passed, 0 failed; no pending insta snapshots
• just fmt / just fix -p nori-tui clean
• cargo build --bin nori && cargo test -p tui-pty-e2e — all runnable suites pass (requires mock_acp_agent binary built)
• Closed the loop: drove ./target/debug/nori --agent elizacp in an isolated tmux session — full turn completes, reply inserted into history above the viewport, no corruption

Share Nori with your team: https://www.npmjs.com/package/nori-skillsets