chore: various fixes, improvements for optimization package

[andrewklatzke]

Requirements

• I have added test coverage for new or changed functionality
• I have followed the repository's pull request submission guidelines
• I have validated my changes against all supported platform versions

Describe the solution you've provided

Provides some fixes and improvements to the optimization method after some QA-ing on my end.

• ensures tools are captured and passed in through the final payload; don't get lost when posting back to LD
• better matching for model configs (defaulting to global models as those will be by far the most used; custom keys will still work for custom models)
• fixes how the state updating works, previously were getting less results than expected for "Ground truth" iterations (only the last result, not intermediaries)
• fixed a bug where the validation phase was advancing the "iteration" and blowing up the iteration number unnecessarily
• late addition: adds a shared dataclass for the LLM call so that they share the same call signature (same fn can be provided to both)
• late addition: removes context_choices as a necessary arg, defaults to an anonymous context
• late addition: creates a protocol class for the config and LLMCallConfig so that the same handler can be used for both if the user wants

Describe alternatives you've considered

These are fixes for inconsistencies/bugs in the package. No alternatives considered.

---

Note

Medium Risk
Touches core optimization run lifecycle and LaunchDarkly API persistence semantics (new PATCH flow, iteration/validation folding), so regressions could affect UI progress reporting and stored run data; changes are well-covered by expanded tests.

Overview
Fixes result persistence for optimize_from_config runs by creating one result record per logical iteration (POST) and incrementally updating it (PATCH) with status, telemetry (latency/tokens), scores (including judge thresholds), and variation metadata, while also auto-closing prior iterations so ground-truth sample records aren’t left stale.

Corrects validation-phase behavior by folding validation sub-iterations back into the parent iteration’s persisted record and ensuring terminal events (success/failure/turn completed) use the main iteration context (avoids mismatched userInput vs completionResponse).

Improves auto-commit payloads by retaining the original tool keys from the fetched agent variation and by resolving modelConfigKey via a new _find_model_config helper that prefers global model configs; also prefetches model configs once and patches the final result with createdVariationKey after commit.

Adds LLMCallConfig/LLMCallContext Protocols (exported in __init__) so one handler can be used for both agent and judge calls, renames judge context variables to current_variables, and makes context_choices optional (defaults to an anonymous context).

^{Reviewed by Cursor Bugbot for commit 8f3468f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 55674ae. Configure here.}