Hold MLOps artifacts on Unity Catalog volumes through the Files API - #114

**Stack:** main (#53 merged) → #112 → #114 → **#50** → #60 → #64 → #115.
Merge bottom to top.

## Summary

Every graph snapshot, training table and model bundle now carries a
self-contained `_provenance.json` that leads back to the source table
versions (and locations, wherever the source backend can name them),
query, plan and engine build that produced it. Today a model has no id,
an artifact records no lineage, an index works only on the machine that
built it, and scoring needs the training job's `Graph`.

This PR adds `kumo.MLOps` / `with kumo.mlops(...)` to store each
artifact once under one artifact root (`s3://`, `gs://`, local, or a
`(connector, path)` file store from #112); `GraphSnapshot.export` /
`load_from_connector` to move a checksummed index between machines;
`snapshot.source_changes(graph)` to see which tables changed since a
snapshot; and `Trainer.predict(snapshot)` to score a stored snapshot
with the trained query. `MLOps(on_error='raise')` turns a failed store
into a `RuntimeError` after the call finishes. Off unless asked for.

**Design decision:** for this release MLOps is opt-in per call or per
block (`mlops=` or `with kumo.mlops(...)`). A call outside both runs
without MLOps, and `on_error="raise"` does not detect that. A model- or
workflow-bound association with an explicit `mlops=False` was considered
and is left as a follow-up. See "Per call, not per model" in
`docs/mlops.md`.

## Usage

```python
cfg = kumo.MLOps("s3://<bucket>/kumo")             # store=True: snapshots, tables, models
with kumo.mlops(cfg):                               # or mlops=cfg on one call
    snapshot = graph.snapshot()
    result = trainer.fit(pq.generate_training_table(snapshot), snapshot=snapshot)

# Later, on another machine with no Graph code:
snapshot = GraphSnapshot.load_from_connector(cfg, id=result.snapshot_id)
trainer = Trainer.load_from_connector(cfg, path=f"models/{result.id}")
scores = trainer.predict(snapshot)
```

## Back-compat

Without a block or `mlops=`, existing calls behave as before.

| Situation | Behaviour |
|---|---|
| Index built before this PR | Reused, not rebuilt; its manifest is
backfilled with `"reconstructed": true`. |
| Bundle published before this PR | Loads and predicts identically; id
computed at load, `manifest` is `None`. |
| Artifact from this PR read by an older engine | Extra files are
ignored. |
| A single S3 or GCS object source | Versioned by VersionId or entity
tag, so its snapshot gets a new id once. |
| `verify_model_bundle` return value | 64 hex instead of a 16-hex digest
no caller used. |
| `DatabricksConnector(identity=...)` | Unchanged; `identity` is now
optional (ambient auth). |
| Manifest, sidecar or root write fails | A warning; the call completes
as before. A failed store raises a `RuntimeError` instead, once the call
has finished, with `MLOps(on_error='raise')`. |

## Testing

- kumo-engine: `test_snapshot_export.py` (round trip, reuse, interrupted
and concurrent exports, corrupt markers, `source_changes`, `on_error`),
`test_snapshot_manifest.py`, `test_training_table_export.py`,
`test_holdout_df.py` (published model's holdout, loaded and fetched
holdout, empty test split), `test_bundle_provenance.py`,
`test_query_from_snapshot.py`, `test_filestore_sink.py` (identical
bundle reused), `test_mlops.py`, `test_graph_mismatch.py`,
`test_artifacts.py`.
- kumo-core typed source versions, bundle ids and export sidecars;
kumo-s3 `test_sink_delete.py`; Snowflake, Databricks and BigQuery
identity tests assert `version_kind`.
- DiskGraph: Rust
`range_checksums_cover_every_listed_file_and_the_manifest`; pybind
`test_exception_exports.py`.
- Env-gated live S3 and GCS round trips in
`test_snapshot_export_live.py`.

Change map: see the `kumo-pr-change-map` comment.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>