feat(cli): full-width dim separator above and below user bubble

[CnsMaple]

🤖 Generated by AI (CnsMaple + Reasonix)

Problem

When mouse-scrolling back to find an older turn in the chat transcript, the
user's submitted prompt is hard to spot. The assistant's markdown reply ends
without a strong visual edge into the user bubble, so the eye has to hunt
through dense, dimmed text to find the › … line. This is a small but real
ergonomic pain on long sessions.

Fix

A full-width dim ─ rule above and below each committed user bubble. The
rule frames the bubble and acts as a scroll anchor — when you scroll up
through an older turn, the two rules bracket the user's input cleanly.

The rule is committed as a tiny placeholder token (NUL + U+E000 + NUL)
and expanded to a real renderUserSeparator(width) call at wrap time, so
it always spans the current viewport width. Terminal resizes, replays at
a different width, and any other rewrap pick up the new width with no
re-commit.

What changed

• internal/cli/transcript.go
  • New userSeparatorToken constant, renderUserSeparatorToken() factory,
    and expandSeparatorTokens(lines, width) expander. The expander walks
    transcript entries and replaces every token with a freshly rendered
    renderUserSeparator(width), so resize is implicit.
  • renderUserSeparator(width int) string is kept as a public helper for
    direct use; it falls back to ASCII - when colorEnabled is false and
    to "" when width <= 0.
• internal/cli/chat_tui.go
  • wrapTranscript is now preceded by expandSeparatorTokens(cm.transcript, contentW) on every transcript change — one insertion at the call site.
  • renderUserBubbleBlock now returns the token form (no width
    parameter) and is reused by both the submit path (chat_tui.go:3180)
    and the replay path (replaySectionsFor).
  • renderUserBubble is untouched; the regression guard
    TestUserBubbleIsLightweightTranscriptLine still passes.

Tests

• TestRenderUserSeparatorIsFullWidthAndDimmed — covers color/ASCII paths
  and width <= 0.
• TestRenderUserBubbleBlockEmitsThreeLines — middle line is unchanged;
  separators are the token.
• TestUserBubbleCommitPathIncludesSeparators — three transcript entries
  after a submit, top/bottom are tokens.
• TestUserBubbleBlockScrollsAsThreeLines — full pipeline
  expandSeparatorTokens → wrapTranscript; top/bottom are exactly
  content-width cells.
• TestSeparatorTokenAdaptsToResize — same token-form transcript rendered
  at width 60 and width 100 produces rules of the matching widths and the
  wider one contains strictly more ─ runes (anti-cache guard).

Notes

• The rule is a real transcript entry, not a viewport background, so it
  scrolls with the content and survives resumed sessions through the replay
  path. Historical sessions are not backfilled — only turns submitted after
  this change pick up the rule, matching the "rendering change" semantics.
• No new state, no new key bindings, no UI surface change outside the
  transcript. The user bubble itself is unchanged.
• AI-assisted implementation, reviewed manually by the author.

[CnsMaple]

Here is a side-by-side ASCII comparison of the transcript area before and
after this change. Terminal width is 60 columns in both renders. The
[dim] marker denotes ANSI dim SGR — in a real terminal the rule below
is rendered as a continuous dim ─ line, not as visible [dim] text.

Before

│ The assistant's last reply ends here. Lorem ipsum dolor sit
│ amet, consectetur adipiscing elit. Sed do eiusmod tempor
│ incididunt ut labore et dolore magna aliqua. Ut enim ad minim
│ veniam, quis nostrud exercitation ullamco laboris.
  › hi
  Hi! How can I help you today?
│ Sure, here is the next reply. The text flows on for a while
│ and the eye has to hunt for the `› hi` line in the middle
│ of the previous assistant block above.

After

│ The assistant's last reply ends here. Lorem ipsum dolor sit
│ amet, consectetur adipiscing elit. Sed do eiusmod tempor
│ incididunt ut labore et dolore magna aliqua. Ut enim ad minim
│ veniam, quis nostrud exercitation ullamco laboris.
[dim]────────────────────────────────────────────────────────────
  › hi
[dim]────────────────────────────────────────────────────────────
  Hi! How can I help you today?
│ Sure, here is the next reply. The text flows on for a while
│ and the eye now lands on the dim `─` rule instantly —
│ the user turn is bracketed by two visual anchors and is
│ trivial to spot while scrolling.

Why the rule adapts to resize

The ─ characters above are committed as a 3-byte placeholder
(\x00\uE000\x00) and replaced at wrap time by
renderUserSeparator(contentW). So the same user turn from an
earlier session renders as a 60-cell rule on a 60-col terminal and
as a 100-cell rule on a 100-col terminal — no re-commit, no stale
hard-coded width:

  60-col terminal                 100-col terminal
  ────────────────────────        ──────────────────────────────────────────────────────────────────────
   › hi                            › hi
  ────────────────────────        ──────────────────────────────────────────────────────────────────────

The placeholder is invisible in the raw transcript (a NUL-framed
Private Use Area code point), so it costs ~3 bytes per user turn and
adds no per-frame overhead for non-user-bubble entries — the expander
is a single strings.Contains short-circuit per line.

[SivanCola]

已处理这两个 review 点并推到 PR head：4f744d95。

• chat_tui_test.go：删除了 TestUserBubbleCommitPathIncludesSeparators 里的空 if ln == "" {} 分支，避免 SA9003: empty branch。
• chat_tui.go：native scrollback/Termux 路径现在在 tea.Println 前通过 nativeScrollbackOutput 展开 separator token，避免把 NUL/PUA placeholder 打到终端；补了 TestNativeScrollbackOutputExpandsUserSeparatorTokens 覆盖该路径。

验证：

• git diff --check
• GOTOOLCHAIN=auto go test ./internal/cli

[SivanCola]

approve