[pylrn]

Summary

Addresses several session synchronization and lifecycle edge cases identified across the Text-to-Speech WebSocket clients (as highlighted in Node SDK #200):

• Eager audio() binding: Captures target flush_id and checks session open state immediately at audio() call time rather than deferring evaluation to the first iterator step. This ensures that creating a reader before subsequent generate() and flush() calls does not inadvertently bind to an advanced or incremented flush_id.
• Prevent re-reading finished flush_ids: Tracks _done_through across consumed requests; calling audio(flush_id) on an id whose FlushDone was already consumed now immediately raises a ValueError rather than hanging indefinitely waiting on a dead queue.
• Purge and discard stale audio on cancel(): Purges pending audio frames from PullBuffer upon cancel() and drops in-flight audio frames arriving for cancelled flush_ids in _dispatch(), preventing cancelled audio from leaking into subsequent iterations.
• Activity-based terminate() drain: Replaces the flat cutoff wait in terminate() with an activity-based loop that continues waiting as long as the server is actively sending frames, timing out only if no message has arrived for terminate_timeout seconds.
• Speak backlog and chunking guard: Validates text_max_buffer_chars before sending text in speak() and chunks text exceeding single Generate frame limits.
• Version bump: Bumps SDK version to 1.8.3.

Verification

• Ran full test suite: 134/134 TTS session unit tests passing (tests/unit/test_tts_session.py, tests/unit/test_tts_session_async.py).
• Added unit tests for:
  • eager audio() request id binding across multiple flushes (sync and async)
  • re-reading finished flush_id error raising
  • speak() buffer backlog limit enforcement and chunking
  • purging and discarding in-flight audio on cancel()
• Linted with ruff and formatted with black.