…iewport

insert_history_lines emitted ESC[1;1r when area.top() == 1 — an invalid
DECSTBM region (bottom margin must exceed top) that terminals ignore,
falling back to a full-screen scroll region. Every inserted history line
then scrolled the live viewport itself, and ratatui's diff buffer never
repaired the damage because it believed the viewport was intact. Large
completed-turn outputs made this state both more reachable (tall
viewports) and more destructive (one full-screen scroll per line).

Extend the existing top()==0 guard (#358) to top() <= 1: warn, keep the
call cursor-position-neutral, persist any viewport move made by the
scroll-down branch, and return false so the caller retains pending lines
for retry. The deferral is self-healing: the next insert scrolls the
viewport down further, opening a valid region.

Adds four VT100-backed regression tests: static top()==1 corruption and
return contract, the scroll-down-into-y==1 path, and retry convergence.
🤖 Generated with [Nori](https://noriagentic.com)

Co-Authored-By: Nori <contact@tilework.tech>

…iewport (#519)

## Summary
🤖 Generated with [Nori](https://noriagentic.com/)

- `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 #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
- [x] TDD: three repro tests written first and confirmed RED against the
unfixed code (viewport rows visibly overwritten by inserted lines),
GREEN after the guard
- [x] `cargo test -p nori-tui` — 1349 passed, 0 failed; no pending insta
snapshots
- [x] `just fmt` / `just fix -p nori-tui` clean
- [x] `cargo build --bin nori && cargo test -p tui-pty-e2e` — all
runnable suites pass (requires `mock_acp_agent` binary built)
- [x] 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

Co-authored-by: Nori <contact@tilework.tech>