Add candidate response trials

[njbbaer]

No description provided.

[Copilot]

🟡 Changes recommended

There are confirmed runtime edge cases (candidate alias overflow) and trial logging can currently omit image-only turns, producing incomplete logs.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR introduces a generalized “trial” mechanism to run multiple candidate variants for a stage (response and/or post-process), randomly select one, and log all candidate outputs for later inspection.

Changes:

• Added a trials runner (Stage, TrialRun, run) and a persistent per-conversation TrialLog writer.
• Integrated trials into Simulacrum generation flow, including per-candidate request recording and summed per-turn cost accounting.
• Added unit/integration coverage for candidate selection, override application, request recording behavior, and trial logging; removed the older dev-only experiment mode.

File summaries

File	Description
tests/unit/test_trials.py	Adds unit tests for trial execution, candidate overrides, aliasing, and trial log behavior.
tests/unit/test_request_recorder.py	Adds tests for request recording keys, reset behavior, and URL truncation.
tests/unit/test_context.py	Adds coverage for Context.with_overrides and runtime override layering behavior.
tests/integration/test_simulacrum.py	Adds end-to-end tests for response/post-process trials, logging, request recording, and cost aggregation.
src/trials/runner.py	Implements async candidate execution, alias assignment, and random selection for a stage.
src/trials/log.py	Implements trial log persistence/rebuild based on conversation history and recorded trial IDs.
src/trials/init.py	Exposes the trials API (Stage, TrialRun, TrialLog, run).
src/telegram/telegram_bot.py	Removes experiment-mode command/help and adjusts stats formatting to align with new cost model.
src/simulacrum.py	Refactors generation into staged trial runs, adds trial metadata/records, and writes/deletes the trial log on conversation mutations.
src/request_recorder.py	Makes request recording file path a default, adds reset(), and requires explicit keys per exchange.
src/lm_executors/experiment_executor.py	Removes the deprecated experiment executor implementation.
src/lm_executors/chat_executor.py	Requires explicit request_key, supports extra_messages, and uses the default RequestRecorder path.
src/lm_executors/init.py	Removes ExperimentExecutor from exports.
src/context.py	Adds runtime overrides layering + with_overrides, and exposes context/trials path helpers.
.gitignore	Removes experiment_log.txt ignore entry (experiment mode removed).

Review details

• Files reviewed: 14/15 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.