Add Gemini 3.8 Live dialog support

[pragmatrix]

Why

Add support for Gemini 3.8 Live models, including their model-specific thinking, tool scheduling, interaction lifecycle, transcription, and client-content semantics.

Summary

• Add standard and Extended Thinking Gemini 3.8 Live model support.
• Add model-specific validation for thinking levels, tool behavior, and response scheduling.
• Add interaction-in-progress events and incremental client content.
• Add transcription modes and language hints.
• Update the Gemini Live submodule for optional tool-call IDs and scheduling fields.
• Bump the workspace version to 3.9.0.

Validation

• cargo fmt --all -- --check
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings

---

Note

This pull request was developed with assistance from an AI coding agent alongside the author.

[pragmatrix]

What's changed

• Fixed the gemini-live-rs workspace compile break: gemini-live-recorder and gemini-live-cli had not been updated for the optional tool-call ids (FunctionResponse.id → Option<String>) and the new scheduling field; crate checks pass for the whole workspace now.
• Corrected ADR 0006: omitted function-declaration behavior defaults to blocking, not non-blocking, per Google's Live API tool-use guide ("a declaration without behavior will still pause all interactions with the model"). The scheduling validation in google-dialog already matches the documented default, so only the ADR text changed.

What's open

Review findings noted but deferred:

• (High, gemini-live-rs) Two concurrent id-less tool calls collide on the empty-string dispatch key in the harness controller, losing the first call's cancellation handle.
• (Medium, gemini-live-rs) InteractionInProgress is not handled by the CLI TUI and recorder consumers — only TurnComplete flushes pending model text, which will stall on 3.8 models.
• (Medium, gemini-live-rs) An unknown interactionStatus value fails deserialization of the entire server message.
• (Medium, gemini-live-rs) A malformed GoAway time_left (NaN/negative/non-finite) panics the session runner; an unparseable one pends the reconnect deadline forever.
• (Medium, parent) On 3.8, output transcription is finalized at every interaction boundary, so one assistant turn can produce several final text events — consumers should expect that; worth documenting in ADR 0006.
• (Low) Dialog example transcription-override doc comments are stale ("Ok/Err", "enabled by default" no longer applies); gemini-live-rs resume-guard can't distinguish "no checkpoint yet" from "unsafe resumption".
• The parent-side submodule pointer bump to the gemini-live-rs fix is still uncommitted and its context-switch branch is unpushed. (submodule context-switch branch is pushed; parent pointer bump landed as cb02b3f)

Commits:

• 2d9ce02 — dialog: rename stdin input enum ServiceEvent to InputEvent
• b4903c8 — adr: omitted function behavior defaults to blocking, not non-blocking
• cb02b3f — gemini-live: update submodule for optional tool call ids
• gemini-live-rs commit in the submodule fork, pushed to pragmatrix/gemini-live-rs branch context-switch (cross-repo, won't auto-link here): 0505325 — recorder, cli: track optional tool call ids after FunctionResponse changes

---

This comment was created entirely with AI assistance.