E.0: enum-state shared wire samples + 8455db9a baseline; add A7 (phase 1E)

[cunninghamcard-bit]

Phase 1E unit E.0 (A7; D2/D25/D26). Branches from exp/restart@8455db9a and is independent of the 1R stack (#19–#23): it touches no code those PRs change. Test and fixture additions only; no library code changed.

Added

• docs/amendments/A7-enum-state-modeling.md: a verbatim copy of the plan's A7 (cmp identical), including the D25/D26 decisions and the 2026-10-11 corrections.
• fixtures/v4/enum-state/*.jsonl shared samples. Each line is one compact JSON value, with keys in the Rust serialization order so the comparison can be byte for byte.
  • tool_call (15 rows): dynamic absent/false/true × providerExecuted absent/true/false; invalid:false with and without dynamic; Dynamic+invalid:true+error (also with providerExecuted/providerMetadata); and providerMetadata.
  • tool_call_rejected (3 rows, D26): invalid:true without dynamic, with dynamic:false, and without error. They are accepted at 8455db9; E.2's try_from must reject them.
  • tool_result (8 rows): dynamic absent/false/true; isError and preliminary true/false; and providerMetadata.
  • raw_tool_call (18 rows): the RFC-0042:84-91 rows 1–6 as provider input text, reused from 1.3's tool-input-vectors.json (on exp/u1.3-tool-input-vectors, not on 8455db9). Includes "" and {"a":, plus dynamic × providerExecuted three-state.
  • stream_tool_input_start / text_stream_tool_input_start (10 rows each): dynamic × providerExecuted three-state, plus title.
• normalized/ is generated per the A7 D25 table:
  • ToolCall: dynamic:false and invalid:false are dropped.
  • ToolInputStart: dynamic is always written as a boolean.
  • providerExecuted passes through everywhere.
  • RawToolCall is identity.
  • ToolResult is identity for now. At 8455db9 ToolResult has no providerExecuted field, so the client-vs-provider split in A7's E.3 correction can't be read from it. E.3 must decide that and update normalized/tool_result.jsonl.
• baseline/ holds the round-trip output recorded at 8455db9 (AIMUX_BLESS_ENUM_STATE=1).
• gen_samples.py documents how the samples and normalized/ were produced.
• aimux-core/tests/enum_state_wire_test.rs (7 tests): serialize(deserialize(line)) is compared byte for byte with baseline/, and normalized/ is checked to be a round-trip fixed point. baseline_equals_samples shows that at 8455db9 every sample round-trips unchanged. Later units point the comparison at normalized/ where D25 changes the output.

Results vs R.0 (local linux x64; the 7 new test names appear in the log)

	R.0 (8455db9)	E.0 (3a63e66)
cargo test --workspace	2871 / 0 / 7	2878 / 0 / 7
fmt --check / clippy -D warnings	—	ok / ok
gen_ts_types.py --check / bindings/node/src/types diff	—	n/a (no type change) / empty
hand-written bindings diff	—	empty
untagged/flatten/new tag in new code	—	none

The Python and Node bindings were not re-run because no library code changed. Go/Java/Kotlin/Swift/Flutter were not run.

Update (684c8b9): review fixes

• Core-side RFC-0042:84-93 samples. tool_call.jsonl adds the parsed values:

  • string "Tokyo"
  • string "{\"x\":1}", which stays a string and is not turned into an object
  • 123, true, null
  • last-wins {"a":2} and 100.0
  • empty text and {"a": as invalid dynamic calls with an error, never {}

• Comparison rules. Provider raw text (raw_tool_call.jsonl) is compared byte for byte. tool_call_semantic.jsonl holds the non-canonical spellings ({"a":1,"a":2} and 1e2) and is compared by JSON semantics with normalized/tool_call_semantic.jsonl (new test tool_call_semantic_samples_match_by_json_value). The baseline was re-blessed at 8455db9 code: existing lines are unchanged and the new lines are appended.

• check_normalized now asserts the line counts are equal before zipping.

• Bless mode (AIMUX_BLESS_ENUM_STATE=1) skips baseline_equals_samples, because the other tests are rewriting the baseline at that moment.

• Deferred to E.3a:

  • the tool_result D25 normalization: dynamic is always written as a boolean for client-executed results and passed through for provider-executed ones;
  • providerExecuted ToolResult samples, including a provider-executed result without dynamic.

  At 8455db9 ToolResult has no providerExecuted field, so normalized/tool_result.jsonl stays identity here.

• enum_state_wire_test: 8 tests pass. The full workspace wasn't re-run for this fix (expected 2879 / 0 / 7 = 2878 + 1).