Unlock full scrollback depth in railway ca session panes

[codyde]

What

railway ca session panes could only scroll back one screenful, no matter how much history the emulator held. This removes that limit: scrollback now reaches the full retained history (4000 lines).

Why it was limited

The one-page clamp in Session::scroll_by was a deliberate workaround, documented in place: vt100 0.15.2 composes the scrolled view as scrollback[len-offset..] ++ rows[..rows_len-offset], and an offset deeper than the screen height underflows the second subtraction — a panic in debug, and in release a silent wrap that renders the live screen ("scrolling does nothing"). resize() carried a matching clamp so a shrinking pane couldn't strand an offset past the new height.

The workaround's comment said upgrading vt100 was blocked on ratatui's exact pin of unicode-width = "=0.2.0". Both halves of that blocker have since resolved upstream:

• vt100 0.16 fixes exactly this bug (visible_rows now bounds with take(rows_len) + saturating_sub).
• ratatui 0.30 moved to ratatui-core, which accepts unicode-width >=0.2.0, <=0.2.2, so it co-resolves with vt100 0.16.

Changes

• Bump ratatui 0.29 → 0.30, vt100 0.15 → 0.16.
• vt100 API move: Parser::set_size / set_scrollback are now on Screen (via screen_mut()), and Cell::contents() returns &str.
• Remove the one-screen ceiling in scroll_by — the emulator clamps to the history that exists, and the settled offset is read back as before.
• Remove the resize clamp; the held offset now survives reshapes (safe on 0.16) and is re-read from the emulator after set_size.

No ratatui API in this codebase changed in 0.30: the upgrade produced zero ratatui compile errors across all five TUIs (cloud agent, develop, metrics, scale, volume browser), and none of the 0.30 breaking-changes items (block::Title, Flex::SpaceAround, layout::Alignment, custom Backend impls, braille symbol tables) are used here.

Tests

Six new tests pressure the removed limit from every direction, in both debug (where the old emulator panicked) and release (where it silently no-opped):

• scrolling_reaches_the_whole_history — 20 screens of output; scroll_by(isize::MAX) lands on the first line ever printed, isize::MIN returns to live.
• scrolling_walks_past_one_screenful — successive wheel notches through the real scroll() entry point keep going past the pane height (the exact thing the clamp prevented).
• a_deep_scroll_survives_resize — a deep offset rides through shrinking and growing panes without underflow or a stuck view.
• scrollback_survives_churn — live output, scrolls, and reshapes interleaved; the held offset always matches the emulator and the view always renders.
• scrollback_clamps_at_capacity — 4200+ lines overflow the 4000-line retention; full depth clamps to exactly 4000 and shows real (oldest retained) history.
• a_deeply_scrolled_pane_draws_old_history — full-frame render via render_session → screen_lines → Screen::cell with the view several screens deep: old lines drawn, tail out of view, "scrolled back" badge shown.

Full suite: 954 passed. The 3 auth_sim failures are pre-existing on master in this environment (fail identically on an unmodified checkout) and unrelated.