feat(runs): persisted run history (T-K1) - #442

Summary

Runs now leave a persisted record that survives a restart: generations, workflows and mesh ops, including failed, cancelled and interrupted ones (T-K1).

Refs #411

Record shape

• Defined in api/schemas/runs.py (RunRecord) and mirrored in src/shared/types/runs.ts. Fields:
  • id, kind (generation | workflow | mesh_op);
  • status (pending | running | done | error | cancelled | interrupted);
  • model / workflow / op ids;
  • inputs, params, timed steps[], error, output_path;
  • created_at, started_at, ended_at, and a computed duration_s.
• This is the shape provenance (T-F1) should reuse, so there is no second store.

Storage (api/services/run_history.py)

• One JSON file at the workspace root, .modly-run-history.json. It is hidden from collections and the asset library, and it follows the workspace if it moves.
• Writes are atomic (temp file + fsync + os.replace).
• A missing file counts as an empty history. A corrupt file is set aside as .corrupt-<ts>.
• Retention cap: 500 records, oldest finished runs dropped first; at most 200 steps per run.
• Disk errors are logged and never raised, so the history can't break a generation.
• Runs still pending or running at shutdown, or found after a crash, become interrupted.

What gets recorded

• Generations: /generate/from-image, /generate/from-artifact and /workflow-runs/from-image, recorded as they are queued, start and end. Cancels are recorded immediately. Progress labels become timed steps.
• Mesh ops: /optimize/op/{id}, /optimize/mesh and /optimize/smooth. Report-only ops (inspect_mesh, suggest_fixes) are skipped.
• Workflows: the renderer reports these through PUT /workflow-runs/{id}, with one timed step per node. Writes are serialized so the final status always wins.

Endpoints (existing /workflow-runs router)

• GET /workflow-runs?status=&kind=&limit=&offset=: newest first, repeatable filters, limit at most 200.
• PUT /workflow-runs/{id}: creates or replaces a workflow record. Returns 409 if the id belongs to a backend-owned run.
• GET /workflow-runs/{id}: falls back to the history once the live job has been purged or the app restarted.

Known limits

• No Queue UI yet; that is T-K2.
• A workflow's model nodes also create their own generation records, not linked to the parent workflow record.
• /transform, /arrange and /patch-vertices are not recorded.
• interrupted runs have no ended_at or duration, since the real stop time is unknown.

Test plan

• python -m unittest discover -s tests: 27 new tests (store lifecycle, paging, filters, retention, restart and shutdown interruption, corrupt/missing file, atomic writes, endpoints, generation done/error/cancel, mesh-op recording)
• npm run test:node: 8 new tests for the workflow record helpers
• npm run lint
• In the app:
  • run one generation that succeeds, one that fails and one that is cancelled;
  • check GET /workflow-runs for each status and duration;
  • quit during a generation and confirm it is listed as interrupted after a restart.