Render ReActV2's last_request_note as a user message with no invented reply

[deepfates]

Closes #259

What I found

• ReActV2 keeps one Imp.History for the whole conversation. The caller's :history input is the start of it, each step adds an event (history_event/5: the step's pending inputs, next_thought, tool_calls, tool_call_results), and the prediction returns all of it in metadata.history. Dwell passes that back as the next call's :history.
• append_note/3 stored the note as %{first_input => note}. When the first step failed, note_after_inputs/3 also stored the unspent inputs as %{intent: ...}. Both entries have inputs and no outputs.
• Imp.Adapter.Chat.render_history_turns/3 has three branches:
  • An entry without tool_calls is rendered as a finished History exchange: a user message, then an assistant message where every missing output reads "Not supplied for this conversation history message." That filler is Imp's own, not DSPy's: DSPy 3.2.1's ChatAdapter.format renders a missing history output as None (checked in the pinned venv; "Not supplied" is DSPy's demo filler). This PR leaves that older divergence alone. For the note, it put a reply in the model's mouth that the model never gave. This happened in native and prompt tool modes, in the last request, and again whenever the history was replayed.
  • A native tool turn (render_native_tool_history_turn/3) already drops an assistant message that has no text and no calls.
  • A written tool turn (render_written_tool_history_turn/3, used in prompt mode) always rendered the assistant, with the filler in empty fields.

The fix

• ReActV2 now records its own user-only turns (the note, and unspent inputs before it) as step events that called nothing: the fields plus tool_calls: %ToolCalls{tool_calls: []} and tool_call_results: []. This is the same shape as every other step event, through the new append_user_turn/2.
• In written mode, the chat adapter now renders a turn with no calls and no other recorded output as its user message alone. That matches what native mode already did. A turn that recorded any output other than the tool-calls field (a generic Imp.Predict step whose signature sets tool_calls_field and stored an answer, say) keeps its assistant message. This also means a ReActV2 step that answered with nothing replays in prompt mode with no filler assistant message.
• History entries a host writes without tool_calls keep Imp's rendering, filler included. A new test pins that.

Breaking, for a host that reads the history's shape: the note entry, and the unspent-inputs entry before it, now carry tool_calls: %ToolCalls{tool_calls: []} and tool_call_results: []. A host that recognises the note by shape (Dwell does) must match the first input's key with an empty tool_calls, or the text. The CHANGELOG entry is under Changed, marked Breaking, with that migration.

Result:

• Native mode: the last request ends on the note as a user message.
• Prompt mode: the note is a user message followed only by the [[ ## tools ## ]] listing. Every prompt-mode step request already ends on that listing, also a user message. There is no assistant message between them.
• Replay: a returned history, dumped and loaded, renders the note as a user message followed by the reply the model actually gave (the submit call).

There is one knock-on change. extract_final's extractor also renders these entries through the native path, so the note no longer carries a filler there either.

Docs and records: I changed one sentence in the ReActV2 moduledoc, added two CHANGELOG entries under Unreleased > Changed (one marked Breaking), and added one line to decisions.md.

Tests and falsification

New tests in test/react_v2_last_request_note_test.exs:

• native tools: the note ends the forced request as a user message, with no invented reply
• prompt tools: the note ends the forced request as a user message, with no invented reply
• native tools: a returned history replays the note as the user turn the model answered
• prompt tools: a returned history replays the note as the user turn the model answered
• a host's input-only history entry keeps Imp's rendering: this one guards the unchanged behaviour, so it passes with or without the fix.

New tests in test/adapter_chat_written_history_test.exs:

• a Predict step with a tool-calls field keeps an answered turn's assistant message: fails with the first commit's adapter condition (thought-only), passes with the current one and on main.
• a ReActV2 step that answered with nothing replays with no assistant message: fails on main's adapter.

I also added a role assertion to ReActV2LastTextTest "a failed step takes the same last request, and the cause is recorded". It covers the unspent-inputs-then-note path.

Falsification, run on the three test files together (25 tests), each time restoring one file to origin/main:

• react_v2.ex restored to main, adapter change kept: 5 fail (the 4 note tests and the last-text test).
• chat.ex restored to main, ReActV2 change kept: 3 fail (the 2 prompt-mode note tests and the empty-answer replay test); native passes, since native replay already dropped an empty assistant.
• chat.ex at this PR's first commit: 1 fails (the generic Predict answered-turn test).
  So each half is needed, and each is covered. (A review note said the first two results were swapped in the earlier body; I re-ran them and the counts above are what they give.)

Other checks:

• mix check: 59 doctests, 9 properties, 3562 tests, 0 failures, 13 skipped (221 excluded)
• mix dialyzer.check: passed.
• mix docs --warnings-as-errors: passed.
• mix parity.check (golden trace against the pinned DSPy 3.2.1 venv): 1 test, 0 failures, so the golden trace is unchanged.

Unsure / open

• A note stored by an earlier Imp is an input-only entry, and it still renders with the filler. It can't be told apart from a host's own input-only entry. Dwell's stored histories from before this change keep that until they age out, unless GroveResearch/dwell#148's narrow renderer stays for them.
• mix differential.check (:dspy_parity) could not run in this worktree. Its pinned DSPy and GEPA sources (tmp/gepa-v0.1.4/src etc.) aren't provisioned, and 19 of 60 tests flunk on that. None of those tests touch ReActV2 or history rendering.