fix(pty): strip terminal control sequences from pty_read and pty_exited output

[NAnD71]

Summary

pty_read results and <pty_exited> notifications pass raw terminal output to the host. Escape sequences in that output are useless to the agent, and the OpenCode TUI executes them because it writes text content to the terminal verbatim (anomalyco/opencode#47687). This affects both V1 and V2. On Windows it hits almost every session, because ConPTY prefixes output with mode, clear-screen and window-title sequences.

Fixes #75. Supersedes #76.

Changes

• src/plugin/pty/sanitize.ts: sanitizeTerminalText(), built on node:util stripVTControlCharacters, plus removal of OSC/DCS/APC strings and leftover C0/C1 control characters. Tabs are kept. Trailing whitespace is dropped, because ConPTY renders an erased line as text padded with spaces to the terminal width. Lines without control characters skip the regex passes. No Bun-only APIs, so it also works under V2's Node build.
• pty_read: lines are sanitized before formatting. Pattern search matches against the sanitized text and returns it.
• <pty_exited>: Last Line is sanitized, and lines that are empty after sanitizing are skipped. The buffer is now split once instead of once per scanned line.
• The raw buffer, the Web UI and /api/sessions/:id/buffer/raw are unchanged.
• PTY_SANITIZE_OUTPUT=0 / false turns this off and keeps the raw output (documented in the README).

Testing

• test/pty-output-sanitize.test.ts (end to end): spawns a real PTY child (the current Bun binary, so it runs on Linux and Windows) that writes colours, a window title, line erases and \r. It then checks pty_read, pattern search and the exit notification for control characters, and checks that PTY_SANITIZE_OUTPUT=0 returns raw output. On Windows, this also covers the ConPTY preamble and padding.
• Unit tests in test/sanitize.test.ts, pty-tools.test.ts and notification-manager.test.ts cover the ConPTY preamble, OSC/DCS/APC, \r, padding, search on sanitized text, and the env switch.
• bun lint, bun typecheck: pass.
• Performance (Bun 1.4.2, 1 MB buffer = the default max size):
  • Default pty_read (500 lines): ~0.1 ms for plain output, ~0.7 ms for coloured output.
  • Pattern search over the full buffer: 0.6 → 3.8 ms for plain output, 0.5 → 17.6 ms for coloured output.
  • Pathological single lines (unterminated OSC/CSI, 300k chars of bare ESC[1) stay linear, worst ~20 ms.
• On my Windows machine, the existing tests that spawn echo/sleep/bash fail with PTY spawn failed (those binaries don't exist there). The failing set is identical on main.

[BananaAcid]

While playing around this myself ( https://gist.github.com/BananaAcid/e0374ff83a2d774d4cf8d76d54ea10da ) I realized, I needed to see the escape sequences as normal text, but just not have them applied to not ruin the TUI. and pty_read should get the raw to be able to have the LLM understand it (like understanding by colored logs or debugging a TUI application).

[NAnD71]

While playing around this myself ( https://gist.github.com/BananaAcid/e0374ff83a2d774d4cf8d76d54ea10da ) I realized, I needed to see the escape sequences as normal text, but just not have them applied to not ruin the TUI. and pty_read should get the raw to be able to have the LLM understand it (like understanding by colored logs or debugging a TUI application).

You're right,I realized escaping is the better default than stripped. The LLM keeps the color/formatting context, and the TUI stays safe.

The change should be straightforward: same sanitize path, just swap the strip call for an escape function (e.g. ESC → ␛, CSI/OSC shown as readable text). Raw buffer and Web UI stay untouched.

I won't have time to get to this soon,so if you're up for it,it's okay to push directly or open a PR against my fork.