Inline multi-line ANSI progress redraws can accumulate duplicate transcript lines #388

Problem / observed symptom

During a Homebrew-driven install inside NMSh, the maintainer observed progress output accumulating/spamming transcript lines instead of redrawing in place. The exact brew command and byte trace have not yet been captured; this is a report requiring physical reproduction, not a confirmed Homebrew-specific regression.

The generic source-level gap is narrower and independently visible on release/v0.18.0:

• src/output/AnsiOutputParser.ts handles carriage return (\r) by resetting the column and can clear/replace the current line with CSI K/G/C/D. It does not implement upward/downward or absolute row movement (e.g. CSI A/B/H/f), yet it appends on every newline.
• src/output/Classifier.ts detects cursor positioning as hasCursorMovement, but moves ordinary output to LIVE after time has passed; cursor motion alone is deliberately not enough for PASSTHROUGH.
• Consequently, inline multi-line progress displays that redraw previous rows using CSI A/2K or H/K are liable to become multiple permanent transcript lines, even though single-line \r rewriting works.
• tests/outputBuffer.test.ts covers 10%\r20%\r100%\n, and tests/interactiveCli.test.ts verifies progress output does not incorrectly steal the terminal; neither proves multi-row redraws remain bounded and visually correct.

This needs an explicit rendering contract and controlled reproduction, not an assumption that all ANSI output should be printed literally.

Minimal diagnostic fixture

Compare the same command in plain Ghostty/Terminal.app versus NMSh, with normal and narrow width, and record the actual byte stream (safely, without private output):

node -e 'for (let i=0;i<=10;i++){process.stdout.write((i?"\x1b[1A\x1b[2K":"")+"Downloading "+i*10+"%\n")}'

A single progress line rewritten via CR is an existing passing control:

node -e 'process.stdout.write("10%\r20%\r100%\n")'

Also test two-or-more-line redraws, chunk boundaries split inside CSI, ANSI colors/hyperlinks, multiple progress components, window resize, completion, ctrl-C, persistent shell, detach/reattach, transcript copy/search/folding, NO_COLOR and Safe mode. Distinguish actual tool output from NMSh's own progress UI.

Design / acceptance questions

1. Decide how NMSh's bounded inline transcript representation supports multi-row in-place overwrites without growing a history entry per visual frame. Avoid turning arbitrary ANSI into a full terminal emulator; if unsupported constructs are encountered, choose an explicit safe degradation.
2. Preserve the distinction between raw PTY event data and the current presented screen. Do not silently delete useful logs or corrupt /copy, transcript search, saved sessions, and replay. Document what copy means for transient redraws and expose raw recorded output where supported.
3. Don't auto-enter PASSTHROUGH solely on progress bars, CSI A, cursor hiding, or high volume: interactive terminal ownership needs independent evidence and the composer should stay usable.
4. Add focused parser/classifier/real-PTY regressions reproducing the reported repeated-line failure; include representative brew install/brew update when possible, plus another independent renderer (e.g. package installer progress).
5. Benchmark frequent updates and long hostile escape sequences; bound memory, CPU, rendering frequency, retained history and row count.
6. Test normal/narrow Ghostty and Terminal.app visually; record the exact before/after. Distinguish ordinary progress from genuine interactive TUIs and preserve existing mode/passthrough behavior.

Coordination

Follow-up to ongoing compatibility coverage #14 and structured presentation #4; related inline-UI handover distinction #211. No duplicate architecture or per-program parsing heuristics. Shared frontend files (src/app/TerminalApp.ts) are owned by the active #359 stabilization track: do not edit those files or disrupt Codex's work until #359 gives an integration checkpoint.

The Reddit discussion was shared as context (https://www.reddit.com/r/commandline/s/WCZeIyUbCh), but its short URL could not be resolved during this review. Do not infer the post's content, comments, or linked repository from the share token. Add the canonical post and any upstream evidence once accessible.