[cunninghamcard-bit]

Unit E.6 (PHASES 1E). Docs only. Stacked on #31 (E.4, a498558a), and the PR base is exp/e4-tool-input-start-kind.

E.5 has no remaining changes, so it has no PR

The E.5 acceptance check is rg -n '\.dynamic\b|\.invalid\b' aimux-* tools. At the E.4 head it finds only:

• Wire conversions: tool.rs:369-370, :686, :690.
• V4-layer RawToolCall fields (provider-layer three-state data):
  • parse_tool_call.rs:107, :192
  • tool_input_test.rs:30, :212
  • mistral_upstream_e2e_test.rs:203
• Accessor method calls to ToolResult::dynamic(): provider_executed_marking_test.rs:88, data_loss_regression_test.rs:275.
• Two test.invalid URLs.

No source code reads wire fields such as ["dynamic"], ["invalid"], ["providerExecuted"] or get(...), either. Every consumer listed in PHASES E.5 already uses accessors or match: GenerateContent/TextContent, ToolApprovalRequestOutput.tool_call, aimux-stream, aimux-ffi, providers and tools. The compiler forced that when E.2–E.4 changed the types. The two internal functions in A7 (parsed_tool_call_arguments, response_tool_call_input) already take &ToolCall.

E.6

I appended a conclusion table to docs/amendments/A7-enum-state-modeling.md. Every A7 row now has a conclusion and a file:line:

• Already done: ToolCall, ToolResult, RawToolCall, ToolKind, StreamPart and TextStreamPart (E.1–E.4).
• Already an enum, no change: ContentPart, LanguageModelMessage and its parts, ToolResultOutput/ToolResultContent, GenerateContent and Source.
• Plain data, no change: ToolCallPart, ToolApprovalResponsePart, RawToolApprovalRequest, ToolApprovalRequestOutput and GenerateTextResult.
• Can't be met, no change: MessageContent/ModelPrompt (message.rs:28, :82) use untagged.
• V4 DTOs on u1.6 (b837aac), no change: aimux-provider/src/content.rs:84, :99.

The table also records the E.4 decision: D25's "always write" rule applies only to core TextStreamPart, and the provider-layer StreamPart passes the value through.

No code changed, so no tests were re-run. The numbers from #31 still apply.