[cunninghamcard-bit]

Phase 1E unit E.4 (PHASES 1E, A7 §D25). Stacked on #28 (exp/e3-toolresult-enum @ 0f4aa9e), and the PR base is that branch. Commit a498558.

Change

• aimux-core/src/tool.rs: new shared pub enum ToolKind { Static, Dynamic }, re-exported, with is_dynamic() and from_dynamic(). Two crate-private serde modules handle the wire boolean:
  • tool_kind_bool always writes the boolean, and treats an absent field or false as Static.
  • tool_kind_option keeps all three states (absent / false / true).
• aimux-core/src/stream_part.rs:
  • Provider-layer StreamPart::ToolInputStart.dynamic: Option<ToolKind>: keeps the three states. This is V4 provider output, and AI SDK language-model-v4-stream-part.ts:58 passes it through.
  • Core TextStreamPart::ToolInputStart.dynamic: ToolKind: always writes the boolean, and absent equals false (AI SDK core stream-language-model-call.ts:809 writes dynamic: chunk.dynamic ?? tool?.type === 'dynamic').
  • into_text_stream_part maps absent to Static. Core has no tool set at this point, so the SDK's tool?.type === 'dynamic' fallback is not modelled.
  • provider_executed is unchanged (Option<bool>, passed through).
• TS: #[ts(optional, as = "Option<bool>")], so StreamPart.ts and TextStreamPart.ts have zero diff. gen_ts_types.py --check passes and bindings/ has no diff.
• Provider construction sites: one change, anthropic/stream.rs:1385 (dynamic.map(ToolKind::from_dynamic)). All other sites write dynamic: None and are unchanged.

Why StreamPart is not normalized

A7's D25 table lists "StreamPart/TextStreamPart → always write". However:

• StreamPart is the provider-layer (V4) stream type, and A7's own row for V4 DTOs says dynamic is passed through.
• AI SDK writes the boolean only in core.
• Normalizing StreamPart would add "dynamic":false to every provider stream part, which would break openai_upstream_e2e_test::responses_web_search_stream and the other upstream ports.

So this PR normalizes only the core TextStreamPart and keeps the three states on the provider-layer StreamPart. That is the same layer split as E.3. Please confirm, and A7's D25 row should be reworded to name TextStreamPart only.

Samples / tests

• fixtures/v4/enum-state/normalized/stream_tool_input_start.jsonl is now the identity, because the provider layer passes values through. gen_samples.py was updated. The 4 rows that changed are the absent/false rows that E.0 had normalized. stream_tool_input_start.jsonl still matches the 8455db9 baseline byte for byte.
• text_stream_tool_input_start_samples_match_baseline became text_stream_tool_input_start_samples_match_normalized: samples serialize to normalized/ (D25), and the normalized lines are a fixed point. baseline_equals_samples no longer lists it.
• New tool_input_start_kind:
  • in core, absent and false give the same JSON;
  • in the provider layer, absent round-trips as absent (None);
  • into_text_stream_part maps true to ToolKind::Dynamic.
• No other assertion changed. responses_web_search_stream passes unchanged.
• Fixture, cassette and contract-tests diff: none outside fixtures/v4/enum-state/ (the two files above).
• No untagged or flatten added.

Results (local linux x64)

	E.3 (#28)	E.4
cargo test --workspace	2894 / 0 / 7	2895 / 0 / 7
fmt / clippy -D warnings	ok	ok
boundary script	0	0
gen_ts_types --check	ok	ok, zero .ts diff
Node ava (node 22)	87 / 1 baseline fail	87 passed / same 1 baseline fail

Python was not run (no venv). The final clippy fix, which made the serde modules pub(crate), was made after the full test run; it only changes visibility, and clippy passes on it.