fix: allow opt-out of session_state_changed via env var

[vincentye38]

Summary

• Fixes prompt() hangs when Claude Code binary doesn't support CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS #497
• When CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=0 is set in the environment, the prompt loop returns the PromptResponse directly from the result message handler instead of waiting for session_state_changed(idle) which will never arrive on older Claude Code binaries
• The env var defaults to "1" (enabled) to preserve existing behavior
• Respects user's env var instead of hardcoding "1", allowing deterministic opt-out

Changes

1. Session.useSessionStateEvents — new boolean flag on the session, derived from the env var at session creation
2. createSession env override — CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS now respects process.env instead of always hardcoding "1"
3. Result handler fallback — after the inner result subtype switch, returns immediately when useSessionStateEvents is false

Test plan

• With CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS unset (default): behavior unchanged, waits for session_state_changed(idle)
• With CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS=0: prompt returns immediately after result(success), no hang
• With older Claude Code binary + env var set to 0: client receives PromptResponse and does not deadlock

[josevalim]

Do we want to support different versions of Claude Code or should we instead document that you need a minimum version? Especially because this seems to fix the problem only by explicitly setting an env var, which I assume has low discovery?

[vincentye38]

CLAUDE_CODE_EMIT_SESSION_STATE_EVENTS Is introduced in v2.1.83. The default is set to 0. IMO, We should maintain this acp agent downward compatible with 0 while supporting 1 until Claude Code SDK changes default setting to 1.

[vincentye38]

@benbrandt Could you take a look at this PR. Thank you!