Fill screenshot transparent areas with the session background

Screenshots rendered the terminal content with transparent margins and
uncovered areas. Composite the session's real background behind the
content so the result is opaque and matches what's on screen: a
background image with its scaling mode and default-background blend, or
the solid background color when there's no image. Window transparency is
deliberately ignored since a saved screenshot has nothing behind it.

The background is laid out as though the pane were the size of the
screenshot (each multi-part piece gets its own), and is rendered one
slice at a time so memory stays bounded the way the streaming encoder
already is. Slices are produced on the main thread; compositing stays
off-main via CGContext. Applied at every output path: the standard
save/copy render, the streaming/multi-part encoder, and the live
preview.

The preview also shows a checkerboard in the area outside the image so
the boundary of what will be saved is clear when the content is smaller
than the preview.

ChatViewController: fix oversized gaps when the gutter panel first re…

…veals an existing chat

When the right-gutter chat panel first reveals an existing chat,
layoutPanels() attaches the panel (synchronously triggering load() ->
reloadData() via viewDidMoveToWindow) before positionPanels() sets the
panel frame and before any layout pass. Every row height was therefore
measured against a table width of 0, so maxBubbleWidth collapsed to its
floor, each bubble wrapped to a sliver and computed absurdly tall, and
the bubbles rendered with large gaps. The width-change self-heal in
performLayoutNow only re-measures on a width change, and once
lastTableViewWidth latches the final width it never re-fires, so the bad
heights stuck.

Force the re-measure the existing re-scroll comment always intended: on
the next runloop, performLayoutNow() lays the table out at the panel's
final width, then an explicit reloadData() re-queries heightOfRow against
it, then scrollToBottom(). Scoped to the inline-panel reveal path; window
mode is unaffected.

Add advanced setting to open tmux windows and tabs in the background

Adds the tmuxWindowsOpenInBackground advanced setting (default off). When
enabled, new or restored tmux windows and tabs are revealed without taking
keyboard focus, so attaching with tmux -CC or having tmux create a window no
longer steals focus from the session you are working in.

The "open in background" intent is threaded as an explicit openInBackground:
parameter from loadTmuxLayout: through openTabWithTmuxLayout: down to
insertTab:atIndex:openInBackground: (with appendTab: and
addTabAtAutomaticallyDeterminedLocation: variants). Existing callers go
through thin wrappers that pass NO, so non-tmux tab insertion is unchanged.

When opening in background, insertTab: leaves the current tab selected,
reveals the window with orderWindow:NSWindowBelow: instead of
makeKeyAndOrderFront:, and skips setCurrentTerminal: so clients do not observe
a focus or window steal.

Based on the approach in PR #684 by Yifan Gao (gaoyifan), reworked to carry
the flag as an explicit parameter rather than transient window state.

3.7.0beta5

Tweak release notes

Clarify cursor Shadow tooltip and fix typo

AI: extend Anthropic prompt caching to conversation history

Prompt caching previously marked only the system block and the last
tool definition, so the conversation history (the bulk of an
orchestration request, dominated by accumulated tool results) was
re-sent uncached on every turn and billed at full input rate.

Add a rolling ephemeral cache breakpoint on the last content block of
the last message. Anthropic reads the longest matching prefix and
excludes cache_control from the cache key, so the block marked this
turn becomes a cache read on the next turn once a newer turn is
appended after it. String content is promoted to a one-element text
block so it has somewhere to carry the marker; array-form messages
(tool results) are marked in place and stay byte-stable across turns.

To support this, message content blocks gain an optional cache_control
field (omitted from the wire when nil, so unmarked blocks serialize
unchanged), and AnthropicCacheControl is lifted to file scope so both
the system/tools breakpoints and the message breakpoint share it.

The live prompt-caching test asserted call 2 re-created zero cache;
that held only when system+tools were the sole cached segment. With
the rolling marker, call 2 legitimately writes a small entry for its
changed last message while reading the shared prefix, so the
assertion now requires the read to dominate the tail write.

Regenerate the Anthropic live-test cassettes: every Anthropic request
body changed, so the prior recordings no longer match and are removed
in favor of freshly recorded ones (the safety benchmarks included).
Non-Anthropic cassettes are untouched.

Mark faint text, cursor, and images in get_screen_contents output

The orchestrator's get_screen_contents tool flattened the terminal grid
to plain text, discarding every cell's rendering attributes. Inline shell
suggestions and ghost completions (zsh-autosuggestions, fish autosuggest,
Claude Code's suggested reply) are rendered faint, so they came through
byte-for-byte identical to text the user actually typed, which routinely
fooled an agent into mistaking a suggestion for an entered command.

Capture the per-cell faint bit (via a new ScreenCharIsFaint accessor,
since Swift can't read C bitfields) and the per-character grid coords,
then weave markup into the extracted text:
  - faint runs wrapped in dim markers
  - a cursor marker spliced at the cursor position
  - images collapsed to an image placeholder
All three use U+27E8/U+27E9 angle brackets, which effectively never occur
in real terminal output. The tool description documents the tokens.

extractScreenText/cursorMarkupIndex now take a VT100Screen instead of a
PTYSession so the logic is unit-testable; added tests covering each
markup end-to-end and composed.

Update iTerm2-shell-integration: backport OSC 133 aid= to bash source

Close the source/bash.1 drift so "make bash" no longer drops the
OSC 133 aid= feature.

AI: remove dead claude-sonnet-4-0 model and its fixtures

The Anthropic model claude-sonnet-4-0 returns HTTP 404 for current API
keys; it is no longer a reachable alias. Remove it from AIMetadata so the
model picker and the per-model live harness sweeps stop offering it.

Also delete the now-orphaned captured fixtures that only existed for this
model: its two SafetyRefusalFixtures and the 16 AICassettes whose request
body targets claude-sonnet-4-0. The live harness derives its model list
from AIMetadata, so there are no per-model test methods to remove; the
AIMetadataFixtureCoverageTest still passes because no remaining model
lacks a refusal fixture.