feat!: finish the data layer — blessed ModelTurn conversion, counter/newtype ids, owned dispatch, serde across the protocol surface

[gold-silver-copper]

Interstitial PR between #2405 and the rig-bevy plugin (bevy-prep series, one-protocol-two-drivers). An audit of main against the ECS-driver requirements found the protocol extraction complete but left data-layer gaps that would force a second driver to duplicate runner logic or fight the borrow checker at every task-spawn site. This PR closes all of them so the plugin PR is pure assembly. Ordered commits, each green.

1. ModelTurn::from_response — the conformance anchor

The only CompletionResponse → ModelTurn conversion was ~15 inlined lines in rig-agent's runner, with two inputs hand-assembly gets wrong: the tool-name sets must come from the prepared request (never re-derived from the spec), and the finish reason must go through the normalized finish_reason() accessor. A second driver copying those lines can silently drift — which defeats "one protocol, two drivers, conformance by construction".

rig_run::ModelTurn::from_response(resp, &PreparedRequest) (plus from_response_parts for drivers carrying the name sets by value) is now the one blessed conversion; the runner is rewritten over it. Also drops the stale "Arc clone" comment (CompletionResponse::raw is a plain serde_json::Value).

2. InternalCallId counter id + ConversationId newtype (breaking)

• internal_call_id was a 21-char nanoid String cloned at every boundary. It is now a Copy + Hash + Ord, serde-transparent NonZeroU64 counter id (the existing counter_id! macro, same shape as RunId), threaded through rig-core's streaming assembler and stream items, rig-run's pending-call/policy types, and rig-agent's hook events (which drop their &'a str borrow for a by-value id).
• Persisted runs re-emit saved ids after resume, so the macro gains advance_past (an AtomicU64::fetch_max) and PendingToolCall's deserialize advances the mint counter past any persisted id — a resumed process cannot mint an id its consumers already saw (regression-tested).
• ConversationId is a string-backed, serde-transparent newtype in rig_core::id; the ConversationMemory trait, rig-memory adapters, and the conversation(..) setters (now impl Into<ConversationId>, so string call sites compile unchanged) use it.
• Serialized stream items and persisted runs carry internal_call_id as a number now — documented in MIGRATING; run state has no cross-version stability contract.

3. Owned-future entry points

ToolCatalog::execute_owned(self, name, args, context) -> (ToolResult, ToolContext) and ConversationMemoryExt::{load_owned, append_owned, clear_owned} (blanket impl over Arc<Self> + owned ConversationId) give hosts that spawn tasks Send + 'static futures without per-site clone-into-async move ceremony. The mutated ToolContext returns by value, preserving execute's publish contract (the runner's dispatch observes context metadata the same way). The borrowed methods stay the implementation surface.

4. Serde across the step/policy types

AgentRunStep, PreparedRequest, RequestPatch, InvalidToolCallContext, RetryRequest, InvalidToolCallAction, StreamedTurnEvent derive Serialize + Deserialize; ModelTurnOutcome gains Clone (+ serde). AgentRun could re-derive its step, but a component caching an in-flight step/prepared request could not round-trip through a saved world. Round-trip tests for each.

5. StreamedTurnAssembler: Clone + Serialize + Deserialize

A mid-stream streamed turn was the one piece of run state that could not persist. The assembler and its internals now derive Clone + serde; the ExclusionCount drop-guard serializes as a transparent count (loudness contract survives a resume) and Clone copies it so each lineage warns for its own drops. Dropping an assembler mid-turn (an ECS despawn) stays silent unless replayed assistant content was actually excluded. A midway serialize/resume test proves the restored assembler produces the identical turn.

(ModelRef: Deserialize, flagged by the audit, turned out to already exist — no change.)

Verification

• cargo nextest run --workspace --all-features: 6239 passed, 0 failed; git diff --stat main -- tests/cassettes empty (commit 1 is byte-neutral on requests — full provider suites replay clean).
• clippy/fmt/doc (-D warnings) clean; cargo check --target wasm32-unknown-unknown green for rig-core (incl. --all-features), rig-run, rig-agent, rig.
• rig-run keeps zero async/tokio/futures deps; dependency-graph guard green.
• New tests: from_response field-for-field equality, counter-id/ConversationId round-trips + advance_past, persisted-id counter advance on deserialize, owned-future Send + 'static pins with behavior parity, step/policy serde round-trips, assembler midway resume.
• grep ModelTurn::new in rig-agent: only test-fixture construction remains.