feat: persist pragmata-owned train metadata

[saschagobel]

Summary

Persists pragmata-owned metadata for completed eval training runs.

This is needed because tlmtc owns the training artifacts but has no concept of pragmata eval tasks. By writing a small pragmata_train.meta.json sidecar with the run_id, task, and optional annotation_export_id, later predict workflows can select the latest compatible model for a task and validate explicit run IDs against the requested task.

Key changes

• Add EvalTrainMeta to src/pragmata/core/schemas/eval_output.py.
• Add resolve_eval_train_meta_path to src/pragmata/core/paths/eval_paths.py.
• Add src/pragmata/core/eval/export.py.
• Persist pragmata_train.meta.json after successful run_tlmtc_train.
• Keep the public train_evaluator return value unchanged.
• Store the sidecar beside the tlmtc train run artifacts under eval/train_outputs/<run_id>/.

Tests

• Add unit tests for the eval train metadata schema.
• Add unit coverage for the train metadata path resolver.
• Add unit coverage for exporting the metadata sidecar.
• Update eval API tests to cover metadata persistence orchestration.
• Update eval integration tests to assert the sidecar is written for direct input, explicit export selection, and latest export selection.

[lramir14]

Thanks for the nice work. I am only flagging a potential issue with the integration between Pragmata and TLMTC. "atomic_write_json" requires its parent directory to already exist, and every other caller in the repo satisfies that itself via an explicit ensure_dirs() on a pragmata-owned path bundle. export_eval_train_meta is the exception — it writes into train_outputs/<run_id>/, a directory pragmata never creates itself; it only exists because tlmtc happens to create it as a side effect of training. Would it be worth adding a defensive path.parent.mkdir(parents=True, exist_ok=True) before the write, so pragmata satisfies its own utility's contract directly instead of relying on tlmtc's internal behavior?

[saschagobel]

Thanks for the nice work. I am only flagging a potential issue with the integration between Pragmata and TLMTC. "atomic_write_json" requires its parent directory to already exist, and every other caller in the repo satisfies that itself via an explicit ensure_dirs() on a pragmata-owned path bundle. export_eval_train_meta is the exception — it writes into train_outputs/<run_id>/, a directory pragmata never creates itself; it only exists because tlmtc happens to create it as a side effect of training. Would it be worth adding a defensive path.parent.mkdir(parents=True, exist_ok=True) before the write, so pragmata satisfies its own utility's contract directly instead of relying on tlmtc's internal behavior?

Good point about making the atomic_write_json precondition explicit. I avoided creating the parent directory here because train_outputs/<run_id>/ is the tlmtc-owned run directory; if it is missing after training returns, pragmata should fail clearly rather than create an orphan run directory. I added an explicit parent-dir check before writing the pragmata sidecar so the contract is local without taking over tlmtc directory ownership.