feat(regen): add Speak V2 streaming, agent update-listen, and Flux EOT tuning

[GregHolmes]

Summary

Fern SDK regeneration (2026-07-08). Pulls in new generator output, re-applies the hand-maintained patches, extends the socket-client convention to the new client, and adds test + example coverage for the new surface.

New public surface (Fern-generated this regen)

• Speak V2 — Flux TTS WebSocket (client.speak.v2.connect(...)): new speak/v2/ package (client, socket client, request/response types) plus top-level SpeakV2Encoding/MipOptOut/Model/SampleRate/Tag.
• Agent mid-session listen reconfigure: send_update_listen(...) with AgentV1UpdateListen/AgentV1UpdateListenListen, and the AgentV1ListenUpdated server acknowledgement.
• Flux end-of-turn tuning fields on the listen providers: eot_threshold, eager_eot_threshold, eot_timeout_ms.

Manual-patch reconciliation

The generator caught up on none of the existing back-compat concerns, so all 25 frozen patches were re-applied (nothing dropped from .fernignore), preserving the new generator additions:

• Legacy alias re-exports merged back into the package __init__.py files (kept alongside the new SpeakV2* / AgentV1* entries).
• language_hint → language_hints migration shim re-added to the three listen-provider models (kept alongside the new eot_* fields).
• Agent socket client: _sanitize_numeric_types, broad except Exception, optional send_keep_alive (kept the new send_update_listen).
• listen/v1, listen/v2, speak/v1 socket clients, query_encoder bool coercion, agent_v1settings* back-compat, create-key alias + wire test — all restored.

Judgment call — listen/v2 send_configure shim: kept

The generator now emits complete ListenV2Configure/ListenV2ConfigureSuccess models, but adopting them changes the public signature (raw dict → model, typed responses) — a breaking change. The typing.Any shim is kept for this maintenance regen; typed-model adoption is deferred to a future major/minor.

New: Speak V2 socket-client patch

The newly generated speak/v2/socket_client.py was made consistent with the other four socket clients and frozen:

• Broad except Exception (preserves the custom-transport error contract — a transport can raise arbitrary types; the narrow generated catch would leak them past the message loop).
• Optional send_flush() / send_close() for no-payload control messages (parity with speak/v1).

Test & example coverage

• tests/custom/test_speak_v2_socket.py — send serialization, response-type parsing, broad-except error emission, and forward compatibility: an unknown/future message type passes through untouched on all four receive paths (sync + async recv() and the start_listening loop), so a newer server never breaks an older client.
• tests/custom/test_speak_v2_connect_wire.py — connect request wire shape (sync + async): targets the /v2/speak path and serializes query params, incl. mip_opt_out=True → "true"; doubles as the wire-level guard on the frozen query_encoder bool coercion. Frozen in .fernignore since Fern never generates websocket-connect wire tests.
• tests/custom/test_eot_thresholds_feature.py — eot_* fields serialize on all three listen providers.
• tests/custom/test_agent_update_listen.py — send_update_listen serialization + AgentV1ListenUpdated resolution.
• tests/manual/speak/v2/connect/ — standalone scripts (main / async / with_auth_token / with_raw_response), parity with speak/v1; now accept a DEEPGRAM_BASE_URL override for non-prod (staging) targeting.
• examples/25-text-to-speech-streaming-v2.py (+ examples/README.md entry).
• reference.md — new "Speak V2 Connect" section and send_update_listen added to the Agent V1 send methods.

Verification

• mypy src/ — clean (844 files)
• mypy tests/typecheck — clean
• pytest — 318 passed, 1 skipped (pre-review baseline; the review follow-up tests below add 6 more and pass)
• pytest tests/custom — 205 passed, 1 skipped (includes the new forward-compat + connect-wire tests)
• ruff check / mypy on the review follow-up files — clean

Release telemetry — version drift fix

client_wrapper.py is a generic extra-file in release-please-config.json, but the generic updater only rewrites lines carrying an annotation — and the two version header lines had none, so release-please never bumped the wire telemetry version. It drifts from the published package version (the pre-existing pattern: main shipping 7.4.0 while sending 7.3.2, and this branch would send 7.4.1 against a 7.5.0 release).

• Added # x-release-please-version to the User-Agent and X-Fern-SDK-Version header lines so they bump in lockstep with the release.
• Froze client_wrapper.py in .fernignore — it is a live Fern-generated file, so a regen would strip the annotations otherwise — and documented the freeze + re-apply step in AGENTS.md.

Live end-to-end ✅ validated against staging

Prod /v2/speak (Flux) returns HTTP 400 at the handshake until GA, so the manual connect scripts accept a DEEPGRAM_BASE_URL override to point at staging:

DEEPGRAM_BASE_URL=wss://api.staging.deepgram.com DEEPGRAM_API_KEY=<staging-key> \
  poetry run python tests/manual/speak/v2/connect/main.py

Ran against staging (wss://api.staging.deepgram.com), both sync and async: Connect → Speak → Flush → Close completes end to end — Connected (model alexis, v 2026-07-02.2) → SpeechStarted → 20 binary audio frames (~77 KB) → Flushed → SpeechMetadata → SessionMetadata → clean close. The forward-compat and connect-wire unit tests remain the regression guards; this confirms the real handshake + streaming path.