V25: Add an open metadata bag (:map → nvarchar(max) via TDS) to cyclium_episodes and cyclium_episode_steps.

Every existing :map column on these tables has a specific, framework-owned purpose (trigger_ref, budget, classification, error_detail, dry_run_opts, args_redacted, result_ref), so there was no place to hang per-run attributes — most importantly which model/LLM actually ran. The model is actor-level config resolved at runtime and observable in Datadog (llm_model), but it was not queryable in the DB.

metadata is a general key/value bag so new per-run attributes don't each need a migration:

• On cyclium_episodes it holds the primary model the episode ran under (e.g. %{"model" => "gpt-5.4-2026-03-05"}), stamped once.
• On cyclium_episode_steps it records a model only when that step diverges from the episode's primary model, so a single-model episode doesn't repeat the model on every synthesis step.

Nullable with no default, matching the other :map columns (a SQL default of %{} isn't portable to the nvarchar(max) representation); the app layer treats NULL as the empty bag.