feat(m3): prime multi-turn dialogue history so the last turn is a real follow-up

[haroldship]

Summary

Split out of PR #127 — this wasn't part of issue #125's ask, it's a new
agent-behavior change discovered while integrating the capability_4
dataset. Stacked on #175 (targets that branch, not main) because it
restructures the same function #175's policy-text wiring (item 5) touches —
review/merge #175 first.

Problem

VAKRA's capability_4 dialogue samples with 2+ turns hold every turn except
the last one already answered (each prior turn carries a gold "answer").
The eval loop live-invoked the agent on every turn in sequence, so the
agent had to actually re-solve turns whose answers were already known —
never a genuine follow-up question, always a fresh multi-step task.

Changes

• sdk_eval_helpers.py: optional history_messages param on
  evaluate_task_with_langfuse; when given, prepended to the live turn's
  message so agent.invoke() receives the whole primed conversation in one
  call (the agent only ever answers the last message — everything before it
  is context, not re-solved).
• eval_m3.py: evaluate_multiturn_task splits 2+-turn samples into prior
  turns (query + gold answer, stringified into synthetic
  HumanMessage/AIMessage history) and the final live turn; only the live
  turn is invoked and scored.
• eval_m3.py: cuga-agent's ## User Context injection only fires on a
  thread's very first message — a primed thread never satisfies that, so
  policy text would silently never reach the model. Embedded directly on
  the live turn's message instead.
• eval_m3.py: expected_output now reads gold_sequence/
  answer_per_turn/tool_response_per_turn from index -1 (the live/last
  turn) instead of 0 — equivalent for single-turn samples, correct for
  multi-turn.
• CodeRabbit review fixups: GenericReactAgent fails loudly instead of
  silently dropping primed history (not currently reachable — M3's
  --agent react never calls this path — but a shared helper other
  benchmarks call); guard gt_by_turn construction against malformed
  ground-truth entries.

Tests

uv run --no-sync pytest benchmarks/m3/tests/ benchmarks/helpers/tests/ -m sanity:
180 passed, 2 skipped (pre-existing/unrelated).

Test plan for reviewers

• Merge/review feat(m3): add capability_4 dataset, wire policy text, remove unlicensed legacy data #175 first
• Run uv run --no-sync pytest benchmarks/m3/tests/ benchmarks/helpers/tests/ -m sanity

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 1c25d0fc-3967-4b93-b0e7-224a7145f745

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}