ReActV2: submit only for signatures that need it

[deepfates]

What changes

ReActV2 offers submit only to signatures that need it.

• One output of type :string: no submit tool is offered. A step that writes text and calls no tool is the answer (termination_reason: :answered). Every interruption takes one path: the step limit, a failed request (:prediction_error, :parse_error) and a step that calls nothing and says nothing (:empty_completion, or :invalid_answer when the output's constraints reject the prose) each make one more request. That request offers the same tools with tool_choice: "none" and carries last_prose_note if the caller gave one. Its text is the answer, with termination_reason: :last_prose and termination_cause naming the interruption. It keeps the tools for two reasons: Anthropic refuses a request whose history holds tool_use/tool_result blocks when no tools are declared, and changing the tool list changes the prompt prefix the provider caches. If the model calls a tool anyway, the call is not run and is not added to the history, where it would be replayed without a result. The completion's text is the answer, and the calls are listed in unexecuted_tool_calls. If Imp.Deadline has already passed, no request is made and the run ends with termination_reason: :deadline_exceeded (and termination_cause). A full context window still ends the run at once, as before.
• Several outputs, or one non-text output: submit and the forced submit are unchanged from DSPy's react_v2.py: same tool description, same guidance text, same tool_choice forcing, plus Imp's existing required-only fallback and typed extraction.

Removed options: prose and on_max_iters. Neither has a choice left: the signature decides. forced_submit_notice applies only to signatures with submit, and last_prose_note only to signatures without it; each is refused at construction for the other kind. Dumps no longer carry prose/on_max_iters; a dump that has them still loads, and the keys are ignored.

Also changed:

• The answered and last-prose history events carry the output field (answer), as a submit event does. Context projection uses these to find where episodes end, and ACP session replay reads the answer from them.
• A prose answer is no longer also emitted as a :reasoning event. Before this, an ACP client showed the answer twice: once as a thought and once as the message.
• If the first step fails before any step has used the inputs, the inputs go into the history before the note, so the note is the last thing the model reads.
• Imp.Adapter.Chat guidance: with finish_tool: nil, the finish line says "When the final answer is ready, write it as plain text without calling a tool." With a finish tool, the text is DSPy's, word for word. The divergence and its reason are written beside put_submit/2 and with_guidance/2.
• Imp.Observability.prediction_status now counts :answered, :last_prose and :finished_by_tool as complete. Before, it reported them as incomplete/failed. This fix is in its own commit.

Demos, the workspace agent example, the deployment agent-optimization example and the tests that scripted a submit for question -> answer now answer in prose. Tests that exercise submit itself use question -> answer, confidence: float.

tool_choice "none" on the wire

test/react_v2_last_request_wire_test.exs runs each provider path through the real ReqLLM stack against a local server and reads back the encoded body: OpenAI "tool_choice": "none", OpenRouter "none", Anthropic {"type": "none"}, each with the same tools as the first request. The test passes on ReqLLM 1.17.1 (this branch). I also ran it on 1.24.0 by cherry-picking these commits onto #138's branch, and the ReActV2 test files (50 tests) passed there too. Both versions have the same normalize_tool_choice clauses, and both omit tools and tool_choice together when the tool list is empty. Reverting to "no tools" fails all three wire tests.

.dialyzer_ignore.exs conflicts with #138. Both PRs move the chat.ex fetch_meta filter line. Whichever merges second needs {"lib/imp/adapter/chat.ex", :pattern_match_cov, {786, 8}}, and #138's own multimodal_runner line.

Evidence

• Mutation checks, each run against the named tests: always adding submit fails 4 tests; routing interruptions back to the forced submit fails 9; removing the deadline check fails the deadline test; dropping the output from the answered event fails the no-submit test and the ACP MCP answer test; putting the note before unspent inputs fails the failed-step test; emitting the answer as reasoning fails the ACP live-updates test; dropping the tools from the last request fails 10 (7 last-prose, 3 wire). The observability test failed before its fix.
• Local mix check on 7f62248b: format and --warnings-as-errors pass; 2,873 tests, 1 failure (MLXLMTrainerTest readiness-timeout output capture), which only fails under machine load and is unrelated. CI: every check passes on 7f62248b.

Not covered here

• No live provider run.
• Imp does not add a fallback for a provider that ignores tool_choice: "none". Such a call is kept out of the history and listed in unexecuted_tool_calls.