Strict bastion prompts (QAX/奇安信) treat one SSH channel write as a single keystroke and silently drop every multi-character chunk, so an IME commit of two or more hanzi and a Chinese paste both reach the bastion's search prompt as nothing at all. Single characters work, which is why typing them one at a time is the only path that works today.
Split plain multi-character text at the send boundary instead of adding a setting: the write side of handleTerminalInputData now iterates one chunk per Unicode code point, and all bookkeeping (command buffer, script recorder, autocomplete, sudo autofill, broadcast, scroll) still sees the original payload exactly once. IME commits split whenever the committed text has more than one code point; raw (non-bracketed) onData splits up to 32 characters and longer pastes keep the single write. Anything containing an escape sequence - bracketed paste wrappers, CSI-u composition, forwarded Kitty key sequences, cursor reports - is never split, and broadcast peers still receive the original string once.
Each chunk carries the same writeToSession options as the original single write, so the sensitive marker and automation flags are kept per character. Bridge writes stay one channel write per IPC call, which reproduces what Xshell sends on the same bastion (#3077).
Co-Authored-By: Claude <noreply@anthropic.com>
When the punctuation keyup was dropped and a previously held key was released
next, the recovery flushed the deferred key (recording a Kitty press) but then
released only the unrelated key, leaving the deferred press forwarded until
focus loss.
Replace the sentinel-rewrite predicate with resolveDeferredKeyupRelease, which
enumerates all three cases: sentinel releases take over the deferred key's
release identity, the deferred key's own release pairs the press directly, and
an unrelated release keeps its identity while the deferred press gets a
synthesized release.
A modified keydown (Ctrl+C, Alt+...) is a command, not the continuation of the
lost punctuation keystroke. Leaving the deferral armed across it meant the next
ordinary keydown flushed the pending character in front of the user's input,
right after the interrupt or shortcut ran.
Discard the stale deferral when the next keydown carries Ctrl/Alt/Meta; an
active composition still resolves it via insertText.
The keydown recovery for a swallowed punctuation keyup flushes the deferred
key, which records a Kitty press locally and for broadcast peers. Since that
branch only runs when the release will never arrive, the press stayed
forwarded and the TUI saw the key held until focus loss or session reset.
Extract the forwarded-press release pairing into releaseForwardedKittyPress and
call it from the recovery with a release synthesized from the deferred physical
key, so the press/release pair completes in the same keystroke.
A real keyup for another held key could land while the punctuation deferral
was still armed (the deferred key's own release was swallowed). That release
was rewritten to the deferred physical key, so in Kitty report-event modes the
held key's forwarded press never got its release and the TUI saw it stuck down
until focus loss or session reset.
Restrict the rewrite to IME sentinel releases (Process/229/Unidentified/
Dead/Compose, mirroring the non-text DOM keys the Kitty encoder already
refuses to send as text). Any real release still ends the stale deferral, so
the #3103 recovery is unchanged.
Plain wheel in screen/vim/codex still forwards arrow keys (xterm's
alternate-scroll), and shift+wheel previewed xterm's normal buffer, which
holds no scrollback once an app owns the alternate buffer - so the preview
clamped to a single screen and looked dead (#2516).
When the alternate buffer is active and the normal buffer has no rows above
the viewport, the preview now renders the session's captured output instead:
a bounded plain-text transcript kept next to the connection log capture,
after the programmatic command rewriter (masked commands stay masked) and
before the replay sanitizer, which drops alternate-screen output on purpose.
Rows are wrapped to the terminal columns with wide-character awareness and
reuse the existing wrap-flag selection joining.
Mouse-reporting apps are untouched: their wheel events still go to the app.
Co-Authored-By: Claude <noreply@anthropic.com>
The CJK IME punctuation deferral (#2833) armed on a punctuation keydown and
only resolved on an insertText remap, an exact-key keyup, or a session reset.
Windows IMEs report Process/keyCode 229 as the keyup of a key they consumed,
or drop the keyup, so the deferred key stayed armed: the deferral then blocked
every keypress, and a second "/" search in vi (or any later keystroke routed
through keypress) never reached the PTY (#3103).
The deferral now ends on any real key release (composing releases excluded, so
an active composition still resolves via insertText) and a stale deferral is
flushed when the next unmodified, non-composing keydown for a different key
arrives. The keypress block is scoped to the deferred key instead of every
keypress, so a stale deferral can no longer swallow unrelated input. A mangled
release pairs its kitty release from the deferred physical key instead of the
Process/229 event.
Two changes based on Codex bot review feedback on PR #3156:
1. Preserve ZWNJ (U+200C) and ZWJ (U+200D) in terminal input.
These carry meaning in Persian orthography and emoji sequences;
PTYs can process the raw bytes, so stripping them could corrupt
filenames, args, and passwords that legitimately contain them.
2. Sanitize composition text before Kitty encoding.
When Kitty report-all + associated-text mode is negotiated,
encodeKittyCompositionText converts U+200B into an escape sequence
(e.g. ESC[0;;8203u) BEFORE reaching the handleTerminalInputData
guard. Sanitize the raw text in both composition paths
(commitImeTextInput remap path and onData handler) so zero-width
artifacts are stripped before encoding, not after.
Update pattern-matching tests to reflect the new variable names
(sanitizedText / sanitizedData) in the Kitty composition paths.
CJK IMEs (notably Microsoft Pinyin / Sogou on Windows) occasionally emit
zero-width Unicode characters (U+200B ZWSP, U+FEFF BOM, etc.) when switching
composition modes. xterm.js sends these to the PTY via onData, and with the
15-graphemes Unicode version they render at width 0 — so the command line
looks normal but contains hidden characters that cause the executed command
to fail.
Add sanitizeTerminalInput() and call it at the top of
handleTerminalInputData so every input path (keyboard, IME composition,
paste, Kitty keyboard, Shift+Enter) strips zero-width formatting chars
before they reach the PTY. If the input was zero-width-only, return early
to skip all downstream processing (command buffer, broadcast, sudo autofill).
* feat(terminal): pass Esc+. through for readline yank-last-arg
Autocomplete was swallowing Escape, and macOS Option+. typed ≥ instead of Meta-., so the standard shell shortcut never reached bash/zsh.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(terminal): allow Shift when Esc+_ yanks the last argument
Underscore is Shift+Minus, so rejecting all shiftKey follow-ups dropped M-_ after autocomplete dismissed.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(terminal): keep Esc+_ armed across the Shift keydown
Typing underscore emits Shift before Minus. Clearing the Meta prefix on that modifier event dropped M-_ after autocomplete dismiss.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(terminal): do not treat Esc+. as typed input
Yank-last-arg rewrites the shell line. Walking the ESC+. bytes appended a literal period to the command buffer and left autocomplete's typed buffer marked reliable.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
User-invoked Select All / Select Word now pulse the gesture so they
still copy. touchcancel releases the same way as touchend so an
interrupted touch cannot leave SearchAddon able to overwrite the
clipboard.
Closes#3019
Codex and other agents emit OSC 9 when a turn completes or approval
is needed, but Netcatty ignored those sequences. Register OSC 9, 777
notify, and kitty OSC 99 handlers, map them to Electron notifications,
and reuse tray focus so a click jumps back to the emitting session.
Ignore ConEmu OSC 9;n progress sequences so they do not become toasts.
#3026
Dragging on the overlay can blur xterm, so native copy never hit the
container listener and the raw wrapped rows went to the clipboard.
Listen on document while the preview is open. Also keep a selected
hard newline at a row boundary instead of dropping it.
#3026
Select All built an element-boundary range, so wrap joining never ran
and copied extra newlines. Context-menu paste selection also left the
overlay up, unlike the keyboard path. Select the text node, accept
element-boundary offsets, and dismiss the preview before pasting.
#3026
The overlay had to keep pointer events so history could be copied, but
then every later click stayed on the preview. A click without a drag
now closes it and restores terminal focus. Copy also joins overlay
rows that xterm marked as soft-wrapped, so a long command does not
pick up extra newlines.
#3026
Urgent interrupt still asked only xterm for a selection. With history
preview text selected, Ctrl+C therefore sent SIGINT into vim before the
copy path ran. Use the same copyable-selection flag for interrupt,
shortcut passthrough, and Kitty broadcast.
#3026
Shift+scroll already overlays the normal-buffer history on the
alternate screen, but the overlay ignored pointer events and any
mousedown or copy chord dismissed it. Selection now stays on the
preview so the usual copy shortcuts, context menu, and copy-on-select
can take that text without leaving vi.
Related to #2980
SearchAddon can leak incremental-match decorations, and WebGL bakes
those yellow backgrounds into the glyph atlas. After Esc the handles
are gone but stained cells remain. Track search decorations on
registerDecoration, paint fill on the HTML overlay only, and resweep
after the search-bar close refit.
Key-chord fan-out now delivers remapped LF through handleTerminalInputData
on each peer. handlingKittyBroadcast already blocks re-broadcast, but that
made canBroadcastInput false and armed password assistance on the target.
Keep the re-fan-out guard and only arm autofill for locally originated
submits.
skipBroadcast only stops the raw-string fan-out so peers can re-resolve
the chord from their own buffer. Sudo autofill still has to treat that
key as a broadcast, or a main-buffer sudo/su submit arms password
assistance on the source session.
* fix(terminal): avoid keyword highlight flicker
* fix: address Codex review on PR #2951
* fix: address Codex review on PR #2951
* perf(terminal): defer highlight catch-up under load
* perf(terminal): bound deferred highlight work
* fix(terminal): bound highlight catch-up backlog
* fix(terminal): mirror history wipes in highlighter
* fix(ci): load throughput baseline from PR base
* fix(terminal): preserve highlight rebuild ordering
* fix(terminal): drain highlights before exit capture
* fix(terminal): preserve line state across clear
* perf(terminal): slice history serialization
* fix(terminal): preserve parser state on recolor
* fix(terminal): preserve protocol state on recolor
* fix(terminal): retain live state during recolor
* fix(terminal): close recolor state gaps
* fix(terminal): retain saved colors across recolor
* fix(terminal): serialize recolor side effects
* fix(terminal): color keywords in place without history rebuild
The previous approach rewrote the stream and replayed a second
terminal, which forced private xterm state snapshots and still
risked flicker-adjacent reset work. Writes now stay pristine and
only dirty cells get a new foreground. Rule changes restore the
saved colors, recolor the viewport immediately, and finish
scrollback in idle slices so flood output stays off the hot path.
Related to #2784
Related to #2879
* fix(terminal): keep in-place keyword coloring correct under load
Enter no longer restyles the previous row, and write/catch-up
positions now follow xterm markers so saturated scrollback cannot
skip a just-written match. Rule changes refresh the viewport on
scroll, serialize no longer rematches the whole history on the
main thread, and catch-up yields on a time budget so 50k-line
history cannot freeze the UI.
Related to #2784
Related to #2879
* perf(terminal): keep flood and scroll off the highlight hot path
Pinned log tails were rematching the whole viewport on every
output-driven scroll, and every write allocated a marker. Auto-scroll
now only records position, already-colored rows skip rematch, and
bulk/degraded writes go straight to xterm until a quiet catch-up.
Related to #2784
Related to #2879
* perf(terminal): merge flood writes on a 16ms burst window
Electerm keeps log tails smooth by coalescing attach writes for one
frame after a flush. Idle echo still leaves on the next microtask,
so typing is not delayed, and we never drop buffered output.
Related to #2784
Related to #2879
* fix(terminal): keep keyword originals across append, clear, and alt screen
Same-line appends no longer throw away saved colors just because the
line fingerprint changed. Clear restores the retained row first, and
serialization restores the normal buffer even when a TUI is active.
Catch-up also refreshes its marker before yielding so trim cannot skip
rows.
Related to #2784
Related to #2879
* fix(ci): stop resetting highlight catch-up on every flood write
The throughput gate was comparing cell-color wrapping to raw xterm
and failing at ~18% on DOM CI. Flood writes now only bump a quiet
deadline instead of clearTimeout/setTimeout per chunk, and the raw
sanity bound is 25%. The product check remains 10% versus main.
Related to #2784
Related to #2879
* fix(terminal): rematch recycled rows and cap wrapped catch-up
Quiet catch-up now treats a recycled BufferLine as stale when its
saved highlight colors are gone, so repeated keywords on a full
scrollback still color. Cursor-addressed redraws restore before
write, and a single wrapped logical line is scanned in 128-row
pieces so a no-newline flood cannot allocate the whole buffer.
Related to #2784
Related to #2879
* fix(terminal): accept plugin #RRGGBBAA highlight colors
Plugin decorations allow 8-digit hex. Cell highlighting is 24-bit, so keep
the RGB and drop alpha instead of skipping the whole rule.
* fix(terminal): color cells for partial grapheme matches
A match that ends inside a combining cell maps both endpoints to the same
xterm column. Color that cell instead of building an empty range.
* fix(terminal): rematch flood rows after empty-line stamps
A hot-path \r\n write stamped the following empty cursor row as current.
The next flood chunk reused that BufferLine with ERROR text, and quiet
catch-up skipped it. Fingerprint the stamped content so filled or
recycled identical rows are recolored.
* perf(terminal): avoid flood-path registerMarker after trim
Once a catch-up range exists, later bypassed writes only keep the earliest
bound. A disposed marker collapses to 0 without allocating a new one on
every saturated-scrollback chunk.
* fix(terminal): recolor the row retained by clear
xterm clear keeps the cursor row after we restore originals. Recolor that
surviving line so enabled rules still apply until the next write.
* perf(terminal): stamp no-match rows without cell snapshots
Catch-up only needs a generation and fingerprint when a line has no
keyword. Allocate typed-array originals when a cell is actually colored.
* perf(terminal): treat pinned-bottom recycle as output scroll
Once scrollback is full, baseY and viewportY stay put while rows recycle.
That was treated as a user scroll, so every flood write rematched the
visible area. Keep pinned-to-bottom updates off the rematch path.
* perf(terminal): keep quiet catch-up to a single viewport paint
Catch-up already refreshes visible slices. A trailing recolorVisible()
caused a second settle paint and tripped the atomic-repaint gate.
---------
Co-authored-by: netcatty-bot <308658023+netcatty-bot@users.noreply.github.com>
* Cursor: Apply local changes for cloud agent
* fix(terminal): stop first SSH fail under StrictMode remount
Defer backend boot past the double-invoke, cancel superseded vault init,
gate start on vault hydration, and harden hotkey/attach cleanup so the
first connect no longer races a stale close or empty keychain.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf: finish StrictMode mega-subscriber plan (accent isolation + docs)
Stabilize published terminal themes across custom-accent drag, apply accent
at chrome/log leaves and CSS injection only, land Phase 0 design/plan docs,
and lock Host-bag / Bugbot-clean contracts for the isolation checklist.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* docs: add StrictMode mega-subscriber design and plan artifacts
Force-add Phase 0 spec/plan under docs/superpowers (gitignored locally)
so the delivery branch carries the approved design and plan copy.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: clear remaining StrictMode doubles and notes editor jank
Gate every AI send and reject concurrent runTurn, latch tray connect flush
and settings persistMounted remounts, and stop notes content drafts from
rebuilding MDXEditor/shell on each keystroke.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf: gate hidden notes panel subscribe and skip host-link annotate on keystrokes
Hidden terminal notes mounts keep a live snapshot read but stop listening to
publishes. Host-link DOM annotation follows hosts/mode (preview MutationObserver)
instead of every markdown value change.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: latch SSH transport idle TTL notify under StrictMode
Avoid duplicate main-process notifySettingsChanged on effect double-invoke
when the idle TTL value has not changed.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* perf: freeze inactive vault notes against hosts churn and prefetch MDXEditor
Gate the retained VaultNotesSection notesStore subscription while hidden,
ignore hosts identity in its memo compare when inactive, and warm the lazy
InlineMarkdownEditor chunk when notes become active.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: clear remaining AI StrictMode double-fire hazards
Latch Ask-AI terminal selection on requestId before draft mutation, hoist
Codex App Server IPC bridge to App (ref-counted), and make Catty
requestApproval join an existing waiter for the same toolCallId.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: remount-safe AI send latch and skip stale MCP approval IPC
Replace the component sendInFlightRef with a module key latch so StrictMode
panel remount cannot reopen a mid-flight send, and no-op resolveApproval when
the pending entry is already drained.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* test: allow gated useNotesStore in VaultNotesSection memo contract
The retained notes section now passes enabled:isActive so hidden mounts do
not subscribe; update the VaultView memo source assertion accordingly.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: keep note drafts across mode toggle and external note publishes
Flush ref-only content before edit/preview remount, and skip prop→editor
setMarkdown while a local draft diverges from the last synced value.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* chore: strip StrictMode sermon comments after Bugbot/Comment Sicko pass
Keep effect-order and public-export contracts; drop remount/perf narration
that the latches and tests already encode.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* fix: remount-safe Ask-AI terminal selection latch
Move consumed requestIds to a module Set so StrictMode remount cannot
append the same selection chip twice before parent state clears.
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
* chore: drop remount latch comment
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Local terminal paste with an image-only clipboard now forwards raw Ctrl+V into the PTY so nested TUIs (e.g. Claude Code) can read the OS clipboard. Text and file-path paste, and remote sessions, are unchanged.
Fixes#2758
macOS copy defaults to Cmd+C. With no terminal selection, the shortcut
handler used to consume the chord as an empty clipboard write, so nested
TUIs that expect Kitty Super+C (e.g. Herdr) never received the key.
Pass plain Meta+C through the same no-selection path as Ctrl+C, while
keeping other unbound copy chords as a safe no-op (#1461).
Closes#2728
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: 陈大猫 <binaricat@users.noreply.github.com>
Bump @xterm/* betas so shared-atlas clear (#6055), no-mipmap upload
(#5987), and capacity eviction (#6043) ship in addon-webgl. Replace the
local isolation/backport postinstall mutator with a verify-only gate.
Default new-terminal scrollback to 3000 (closer to VS Code's 1000).
Load @xterm/addon-image in the shared xterm runtime so remote programs can
draw real raster images (kitten icat, chafa --format kitty, image previews
in file managers and TUIs) instead of ANSI approximations.
- domain/terminalInlineImages.ts owns the limits and turns TerminalSettings
into addon options; disabled settings resolve to null so the addon is not
loaded at all and the terminal keeps advertising no graphics capability.
- The addon is loaded right after term.open, before the WebGL renderer, so it
patches IRenderService.setRenderer first and its canvas layers survive every
DOM <-> WebGL swap (deferred WebGL, soft-hide suspend, context-loss rebuild).
It is disposed before term.dispose so bitmaps never outlive their terminal.
- Hibernate restores a tab from a text-only snapshot, which cannot carry image
data. XTermRuntime.hasInlineImages reads the live image cache, and Terminal
degrades such a session to soft-hide: hibernateRuntime takes the soft-hide
branch, canFinishHibernate re-checks after every async step, and an eviction
request leaves the session soft-hidden. The block lifts by itself once the
cache is emptied by a terminal reset or FIFO eviction.
- Settings > Terminal > Inline Images exposes the master switch, per-protocol
switches (Kitty / SIXEL / iTerm IIP) and the memory limits, translated for
en, zh-CN, zh-TW and ru. Protocol switches cloud-sync; the memory limits stay
local like the other per-device renderer tuning.
Co-authored-by: Wine Fox <fox@ling.plus>
* fix(terminal): batch hidden log output
* fix(terminal): preserve ordered hidden output
* fix(terminal): bound hidden output overhead
* fix(terminal): respect custom tab stops in timestamps
* fix(terminal): preserve terminal parser state
* fix(terminal): preserve hidden output boundaries
* fix(terminal): retain batching with prompt formatting
* fix(terminal): preserve prompt boundaries through transforms
* fix(terminal): handle unbatched prompt output
* fix(terminal): preserve batched prompt and time boundaries
* fix(terminal): retain background timestamp boundaries
* fix(terminal): preserve visible timestamp boundaries
* fix(terminal): separate shell output from tui frames
* fix(terminal): separate output after tui exit
* fix(terminal): preserve prompts across output slices
* fix(terminal): flush output before context reads
* fix(terminal): time-slice timestamp marker cleanup
* fix(terminal): preserve prompt columns after bare LF
* fix(terminal): preserve background write yielding
* fix(terminal): retain yields across timestamp boundaries
* perf(terminal): scan output boundaries once
* fix(terminal): track prompt cursor movement
* fix(terminal): resume prompt scan after resets
* fix(terminal): preserve stateful prompt ordering
* fix(terminal): drain close output before capture
* fix(terminal): measure control-only prompt prefixes
* fix(terminal): preserve prompt control boundaries
* fix(terminal): preserve prompt mode toggles
* fix(terminal): bound hidden batches and unfocus policy
Always drain write-coalescer size caps even when scheduled frames are gated,
keep unfocused-but-visible panes on throttled batching instead of per-chunk
force flush, compute prompt starts on post-transform display data, and model
full-width wrap so prompt-on-newline does not insert a blank line.
* perf(terminal): skip off-path timestamps and share alt-screen probes
When host line timestamps are disabled, write plain bytes without the
segmenter/marker hot path. Reuse one alt-screen CSI probe per coalesced
chunk for frame split and schedule mode, and reuse a single full-batch
visual-row measure on the single-segment batched timestamp path.
* fix(terminal): read line-timestamp enable from live hostRef
Session data handlers close over boot-time ctx.host, so toolbar/vault
toggles of showLineTimestamps were ignored. Prefer hostRef.current for
the marker hot-path gate and keep hostRef synced on every Terminal render.
* perf(terminal): per-second timestamp ledger while gutter is off
When line timestamps are hidden, keep a lightweight one-stamp-per-second
ledger without xterm markers. Opening the gutter materializes recent
ledger entries into markers for paint; while visible, keep per-line
markers for accuracy.
* perf(terminal): separate timestamp record from gutter paint
Always record a one-stamp-per-second ledger on the write path with no
xterm markers. Gutter rendering maps the ledger to visible rows, so open
and closed modes share the same cheap record path.
* fix(terminal): fill gutter gaps with previous second stamp
When painting line timestamps from the per-second ledger, blank rows
between stamps inherit the preceding stamp's label so multi-line output
in the same second is fully labeled.
* fix(terminal): stop empty gutter paint and false scrollback rebase
Only rebase ledger anchors when the buffer is actually full and baseY
grows from top trim — not when baseY rises while the buffer is still
growing (which deleted MOTD/login stamps on scroll). Cap gutter paint at
the last content line so empty viewport rows past the cursor stay blank.
* fix(terminal): soft-wrap stamp placement and sparse reflow anchors
Estimate ledger lines with visual-row measurement so narrow terminals
do not pin the next second on a wrap continuation. After each write,
attach at most one xterm marker per stamped second so reflow and
scrollback trim keep anchors honest without per-line marker storms.
* fix(terminal): ignore alt-screen rows when placing timestamp anchors
Stamp line estimates froze incorrectly through vim/less alt-screen
output, so sparse reflow markers landed far below the restored normal
buffer. Skip visual-row advancement while DEC alt-screen is active,
key one-stamp-per-second from the batch arrival date, and only attach
anchors still present in the ledger after capacity trim.
* fix(terminal): seed stamp lines from normal buffer when already on alt
When a write begins on the alternate screen, active cursor is the TUI
row — freezing that through 1049 leave mis-pinned sparse anchors and
dropped them past normal bufferLength. Seed and restore stamp estimates
from buffer.normal, and cover multi-write already-on-alt leave+prompt.
* fix(terminal): keep same-chunk pre-enter line on alt leave
Leave CSI must not always re-read pre-write buffer.normal: when enter
happened in the same walk, the frozen estimate already includes normal
advances before the TUI. Only restore buffer.normal when the write
started already on the alternate screen.
* fix(terminal): pure alt stamp transitions that never rewind lines
Stamp estimates stay in normal-buffer coordinates: seed once (active or
buffer.normal when already on alt), freeze while alt is active, and on
leave only clear altActive — never re-read the live term mid-walk.
Export applyAltScreenAction with a table-driven unit test covering
spurious leave, plus write-path cases for already-on-alt, same-chunk
pre-enter advance, and leave-without-enter.