Add reflector instrumentation, pluggable payload mapper, v2 templates

[Teq2412]

Summary

Three small, independent additions to the agentic optimization pipeline:

• BaseAgenticOptimizer._invoke_agent — wraps the inner Strands agent(...) call to record wall-clock time and accumulated metrics (tokens, cycles, tool calls, latency) onto self.last_wall_clock_s / self.last_metrics. Lets callers read step cost/latency without subclassing. ContrastiveReflectionOptimizer._run_reflection routes through it.

• AgentCoreRolloutEngine.payload_mapper — the engine still builds the canonical {"data_sample": ..., "params": ...} payload. New payload_mapper: Callable[[dict], dict] init kwarg is a transform applied to that canonical dict before the HTTP call, so deployments expecting a different shape (flat fields, renamed keys, nested envelopes) can remap without subclassing the engine. Default is None (identity), fully backward-compatible.

• New built-in templates at harness_optimizer/templates/contrastive_reflection_v2/. Differences from v1:

  • Forbids the ## Learned Behaviors appendix; requires holistic integration of insights into the original prompt's existing structure.
  • Adds a non-negotiable safety invariant (require user confirmation before consequential actions; remove guidance that tells the agent to act autonomously without confirmation).
  • Adds a length guardrail (<= max(1.1 × original, original + 500 chars)).
  • Submission flow still uses the existing submit_optimized_params tool contract.

Test plan

• pytest tests/rollout_engines/ tests/optimizers/ passes locally (excluding tests/optimizers/test_multi_agent.py, which is broken on main independently of this PR — it imports context_alchemy.datamodels).
• Runtime smoke: load_builtin_template("contrastive_reflection_v2/...") loads + renders both new templates.
• Runtime smoke: AgentCoreRolloutEngine(..., payload_mapper=fn) stores the mapper; default constructor leaves _payload_mapper = None.
• Runtime smoke: ContrastiveReflectionOptimizer exposes last_metrics, last_wall_clock_s, _invoke_agent.
• End-to-end reflector run on a real benchmark — verified separately on AppWorld (Sonnet 4 reflector produced a valid optimized prompt; metrics captured wall=224.7s, in=758644 / out=15092 tokens).

🤖 Generated with Claude Code