LLM answering agent for free-form questions

[slowdini]

Part of #244 (full-codebase eval testbed). Depends on #257 (heuristic responder).

Why

#257 ships deterministic responder heuristics and stops the conversation with a distinct reason
when it meets a question it cannot classify. #244 asks for the general case:

For more complex questions, we may need some kind of "question answering agent", designed to
quickly generate a comprehensible response, always tending towards "choose whatever is
recommended/simplest/easiest".

This ticket fills that gap. It is deliberately separate from #257 so the loop is proven before a
second model is introduced into the attribution picture.

Scope

• Dispatch a small answering agent through the same harness as the run, prompted to choose the
  recommended, simplest, or easiest path.
• Trigger it only on the "heuristics could not classify" branch from Dynamic conversation turns with a heuristic responder #257 — the deterministic path
  stays the default, both for cost and for reproducibility.
• Attribute it separately in conversation.json: a turn produced by the answering agent must be
  distinguishable from a heuristic turn, and the responder model must be recorded alongside the
  agent and judge models already tracked in conditions.json.

Files

• src/cli/run/conversation.rs
• schema/conversation.schema.json
• src/core/types.rs (responder model provenance)

Acceptance criteria

• A free-form question the heuristics reject is answered by the responder agent and the run
  continues.
• Responder-produced turns are distinguishable from heuristic turns in conversation.json.
• The responder model is recorded as run provenance.
• With the responder disabled, behaviour is identical to Dynamic conversation turns with a heuristic responder #257 alone.
• Responder failure degrades to Dynamic conversation turns with a heuristic responder #257's stop reason rather than failing the run.

Cross-cutting requirements

• Mode B parity. Acceptance must hold for both Mode::NewSkill and Mode::Revision
  (src/core/types.rs:225), and the snapshot/promote path (src/workspace/snapshot.rs,
  src/workspace/promote.rs) must keep working against codebase-backed iterations.
• Provenance reaches the report. Source + resolved SHA must land in conditions.json, each
  run.json, benchmark.json, and the BASELINE.md built by promote.rs:231 — not just in the
  workspace.

Verification

cargo fmt --check && cargo build && cargo test && cargo clippy --all-targets -- -D warnings

Implementation requires the slow-powers:working-with-tdd skill.

Parent: #244