Claude Code 2.1.208: screen reader mode, memory-leak and long-session performance work

• Screen reader mode: claude --ax-screen-reader, CLAUDE_AX_SCREEN_READER=1, or "axScreenReader": true
• New vimInsertModeRemaps setting for two-key insert-mode sequences (e.g. jj to Escape)
• New CLAUDE_CODE_PROCESS_WRAPPER: routes every Claude Code self-spawn through a required wrapper executable
• Mouse clicks work in fullscreen multi-select menus and "Other" input rows
• Memory: MCP stdio stderr capped (was up to 64 MB per server), LSP documents LRU-capped at 50, file edit read cache bounded to 16 MB, async hook output and large headless tool results no longer retained
• Transcript size cut up to 79x in edit-heavy sessions by pruning superseded file-history backups; checkpoint disk usage bounded
• Up to 7x faster tool rounds in print/SDK sessions with many MCP tools; permission rule matchers compiled once and cached
• Fixed the context window resetting to 200k after an auto-update, causing false "100% context used" on resume
• Fixed truncated stream-json output from claude -p, blank CRLF lines killing stream-json sessions, and hangs on non-string set_model control requests
• Fixed CLAUDE_CODE_MAX_OUTPUT_TOKENS and similar vars reading 1e6 as 1
• Fixed the Edit tool failing on files modified after read when the target text still matches uniquely
• Fixed Read/Grep/Glob edge cases: empty-file offset error, invalid regex reported as no matches, paginated count under-reporting, null bytes in paths
• Fixed /release-notes injecting the whole changelog into subsequent requests
• Fixed apiKeyHelper failures hidden behind a generic 401 after ~10 silent retries
• Markdown tables over 200 rows render the first 200 with a truncation notice
• Background agents: completed agents stay in /tasks until cleanup, cold attach shows the transcript immediately, older daemons no longer restart newer workers on the old binary

Note: the source entry is truncated mid-bullet at the end.