test(optimize): 54 regression tests for all 6 confirmed bugs

Also fixes two source bugs uncovered by the new tests:

BUG-5 fix: OptimizeMetricAdapter.score() — wrap coro-creation inside
  the try block so sync-raising evaluate() is caught and returns 0.5
  neutral (instead of propagating the exception to the caller).

BUG-3 fix: PromptFitResult.to_dict() — include score_history in the
  serialized dict so it round-trips through JSON artefacts.

New test file: tests/test_optimization_bugs.py
  TestBug1AgentRunnerLocalCallable (7 tests)
    - async callable dispatched correctly
    - sync callable runs via asyncio.to_thread (verified with threading)
    - dict response extracted via _response_text
    - errors captured, no crash
    - prompt_override / context forwarded
    - without callable falls through to HTTP path

  TestBug2LLMCallerBaseUrl (5 tests)
    - configure() stores / resets class-level attrs
    - LLMCaller.call() forwards base_url/api_key to _call_openai
    - per-call args override class-level defaults
    - non-openai provider unaffected

  TestBug3PromptFitResultHistory (8 tests)
    - .history returns score_history list
    - derives from full_val trials when score_history empty
    - fallback to all trial scores
    - empty list when no data
    - stable across repeated calls
    - score_history in to_dict()
    - generate_fit_report does not crash
    - fitter.fit() populates score_history on result

  TestBug4CompositeMetricScore (9 tests)
    - has .score() method
    - returns float in [0,1]
    - custom always-1.0/0.0 sub-metrics
    - dict expected_output JSON-serialised and forwarded
    - sub-metric failure → 0.0 contribution, no exception
    - weighted average computed correctly
    - usable in agents.WeightedMetric (validates .score() on init)
    - usable in MetricLoss

  TestBug5OptimizeMetricAdapter (13 tests)
    - returns float not coroutine
    - correctly awaits async metric
    - query extracted from input["query"] and ["current_query"]
    - prediction dict JSON-serialised as response string
    - expected_output dict JSON-serialised
    - expected_output None passed as None
    - neutral 0.5 on async exception (judge unavailable)
    - neutral 0.5 on sync exception (evaluate() raises before await)
    - usable in agents.WeightedMetric
    - usable in MetricLoss
    - name inherited from metric / overrideable
    - score value propagated correctly for all values in [0,1]

  TestBug6PromptFitterBridgeLocalAgent (5 tests)
    - local_agent= param forwarded to PromptFitter
    - live agent used as fallback when local_agent=None
    - explicit local_agent wins over live agent
    - llm_base_url/api_key set on LLMCaller class defaults
    - injected fitter= returned as-is

  TestEndToEndLocalTraining (3 tests)
    - full PromptFitter.fit() with local callable: history, score_history,
      best_score, summary() all accessible
    - OptimizeMetricAdapter → WeightedMetric → MetricLoss end-to-end
    - CompositeMetric → MetricLoss with multiple sub-metrics

Author: piotrlaczkowski

src/agentomatic/agents/metrics.py

@@ -340,14 +340,18 @@ def score(
         )
 
         # --- run evaluate() synchronously ---
-        coro = self._metric.evaluate(query, response, expected)
+        # Wrap the entire call (including coro creation) so that metrics whose
+        # evaluate() raises synchronously (non-coroutine callables that throw)
+        # are also handled gracefully.
         try:
+            coro = self._metric.evaluate(query, response, expected)
             result = asyncio.run(coro)
         except RuntimeError:
             # Already inside a running event loop (e.g. notebook, FastAPI handler)
             import concurrent.futures
 
             try:
+                coro = self._metric.evaluate(query, response, expected)
                 with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
                     result = pool.submit(asyncio.run, coro).result(timeout=60)
             except Exception:

src/agentomatic/optimize/config.py

@@ -424,6 +424,7 @@ def to_dict(self) -> dict[str, Any]:
             "best_score": self.best_score,
             "baseline_score": self.baseline_score,
             "absolute_improvement": self.absolute_improvement,
+            "score_history": self.score_history,
             "metric_deltas": self.metric_deltas,
             "param_suggestions": {k: asdict(v) for k, v in self.param_suggestions.items()},
             "best_config": self.best_config.to_dict(),