feat(experiment): add multi-episode orchestration

[Staaaaaaaaar]

Motivation

PR 07 established the authoritative single-Episode lifecycle and PR 08-10 added PointNav, evaluation, and durable recording. PR 11 connects those boundaries into an ordered multi-Episode Experiment with explicit continuation and infrastructure-failure policies.

Scope

• execute every configured Episode serially without restarting Env, Agent, or Experiment processes
• recreate and dispose Episode-scoped Task and Evaluator resources between Episodes
• reset Env before Agent, republish the immutable task, and reuse the existing manual/automatic start path
• bind evaluator termination callbacks to the originating Episode ID and reject delayed callbacks
• freeze evaluator metrics and trajectory during safe finalization
• persist each Episode and the final Experiment summary through ResultRecorder
• publish EpisodeResult after durable Episode artifacts are committed
• preserve the SingleEpisodeOrchestratorNode compatibility import

Failure Policy

• SUCCESS, TIMEOUT, ABORTED, navigation FAILURE, and INVALID_TASK finish the current Episode and continue when another Episode remains
• component ERROR/staleness, clock freeze, and Env/Agent reset failures stop the Experiment
• no process or container is automatically restarted

Architecture Changes

• add an ROS-independent ExperimentController for ordered Episode lifecycles
• add a recorder-ready EpisodeEvaluationResult boundary
• make Task publishers and Evaluators explicitly Episode-scoped and closeable
• keep Experiment lifecycle mutation serialized on the existing control thread
• retain all existing ROS interfaces without wire changes

Testing

• make dev-lint
• make dev-check
• 16 packages built
• 207 tests passed, 0 errors, 0 failures, 0 skipped
• protocol-stub tests cover three Episodes, unchanged process ID, reset ordering, independent artifacts, EpisodeResult publication, stale callback rejection, INVALID_TASK continuation, and reset-failure stop policy
• a reference vertical test runs Mock Env + Mock Agent + PointNav + Simple Evaluator + Recorder for three consecutive simulation-time timeout Episodes without process restart

Acceptance Criteria

• at least three Episodes execute without restarting services
• Episode-scoped resources and state do not leak between runs
• each Episode produces independent artifacts and aggregate counts are correct
• success, timeout, abort, invalid task, and reset/component failure paths are covered
• status and simulation-clock watchdog behavior remains active
• interrupted Experiments retain parseable complete=false results

Known Limitations

• Episodes remain serial; parallel execution is out of scope
• no automatic process/container restart or distributed recovery
• results_root must be configured to enable filesystem recording and experiment_id must be unique for immutable result directories
• three-container Compose and CI black-box execution remain PR 12

Follow-up

PR 12 will run this vertical path across the production-like Env, Agent, and Experiment container boundaries.

Roadmap: PR 11 - Multi-Episode and Failure Policy