fix(terminal): replay restores the DEC modes the capped buffer evicted (#843)

[Juliusolsson05]

Part 1 of #843 (the mouse wheel is dead in an OpenCode terminal pane after it remounts). Part 2, routing Jump to Latest through the TUI server, follows in its own PR.

Why

attachAgentPty / attachTerminal replay the last 512 KiB of PTY output to a newly mounted xterm.

• A full-screen TUI writes its modes once at startup. OpenCode 1.18.31, recorded for this PR, writes ?1049h ?2004h ?1000h ?1002h ?1003h ?1006h.
• At 60 fps of repaint, the cap evicts that preamble within minutes.
• The remounted xterm then sat on the normal buffer with mouse tracking off, so the wheel reached nobody. The pane looked right, because a full-frame repaint renders the same on either buffer.

Change

• CappedTextBuffer gets an onEvict hook. It reports evicted text in stream order, including an oversized chunk's discarded head.
• DecModeTracker folds the evicted bytes into the terminal state as of the replay's first byte. It follows xterm.js semantics, checked against the shipped @xterm/xterm bundle:
  • mouse protocols 9/1000/1002/1003 are exclusive, and resetting any of them turns tracking off;
  • encodings 1006/1016 are exclusive, a reset returns to default, and 1005/1015 are ignored as in xterm.js;
  • 47/1047/1049 select the alternate screen;
  • ESC c resets everything;
  • it also tracks the flags 1, 25, 1004 and 2004;
  • a sequence cut by an eviction boundary is carried over, as a flat copy so it can't retain a large evicted chunk (fix(#288): make sub-agent body truncation actually free memory (V8 slice trap + toolUseResult miss) #321).
• TerminalReplayBuffer.replay() = the mode prefix + the tail. Both attach paths use it. read() stays byte-exact for sessions.terminalRead paging.
• Not replayed: ?2026 synchronized output (per-frame; a prefix could freeze painting) and terminal queries (they would send stale answers to the program).

Why the state at the replay start, not the current state: the first fix prefixed the current modes and was reverted. It moved retained normal-screen output onto the alternate buffer, where the retained ?1049h wiped it (docs/superpowers/research/2026-09-08-post-merge-regression-audit.md).

Tests (real xterm headless, real OpenCode bytes)

• testing/fixtures/terminal-replay-modes/: the real OpenCode 1.18.31 startup, recorded under node-pty and sanitized. The README gives provenance.
• terminalReplayModes.test.ts compares the replayed terminal with one fed the live stream:
  • the recording with its preamble evicted (alternate screen, any mouse);
  • normal output before a retained ?1049h, which fails the reverted "current modes" design (mutation-checked);
  • mouse-protocol exclusivity and ESC c;
  • a sequence cut by a boundary;
  • a byte-exact replay when no mode was set.
• sessionManager.screenGate.test.ts: through the real SessionManager at the real 512 KiB cap, attachAgentPty gives a remounted xterm the alternate screen and any mouse tracking. This fails with read() in the attach path.

src/main/sessions and sessionManager pass (196), and npx tsc -b is clean.

🤖 Generated with Claude Code

[Juliusolsson05]

Review (APPROVE) findings folded in 37e83e08:

1. ?2031 (theme-change notifications) is tracked.
2. A sequence straddling the eviction boundary is replayed whole (prefix + held fragment + tail), with a test.
3. An ESC that cuts a sequence short restarts parsing on it, with a test.
4. Soft reset (DECSTR) is documented as not modelled.
5. ESC c keeps the cursor-visibility flag, as the shipped xterm does.
6. A shell attach-path test at the 256 KiB cap. It fails with read().
7. The fixture README now says exactly what was edited: the user name only. The wrapped temporary path can't be string-replaced, and it isn't sensitive.

Tests: src/main/sessions plus both SessionManager suites pass, and tsc is clean. I'll merge once CI is green.