Add ModelRun.total_cost and total_data_rows

[apollonin]

What

Adds two lazy properties to ModelRun:

model_run = client.get_model_run(model_run_id)
model_run.total_cost       # float | None  (USD)
model_run.total_data_rows  # int | None    (data rows processed)
model_run.refresh_cost_and_usage()  # bust the cache

Why

Model Foundry users (DoubleVerify) currently log cost and dataset size by hand. This lets them pull both straight off the ModelRun they already fetch — matching the workflow in their existing scripts (client.get_model_run(...)).

How

On first access, a single modelFoundryModelRunInfo query is issued (which proxies to model_service's already-computed per-job cost + data-row count) and cached on the instance. Nothing is persisted on the run — it's real-time rehydration. Runs not backed by a Foundry model job return None (the LabelboxError is swallowed).

The get_model_run query itself is unchanged, so existing fetches stay cheap; the extra round trip only happens when cost/usage is actually read. (A server-side ModelRun.totalCost field resolver was rejected because DbObject selects all fields on every fetch, which would hit model_service on every get_model_run.)

Part of a 3-PR stack (deploy in order)

1. python-monorepo model_service — populates total_data_rows (Labelbox/python-monorepo#2502)
2. intelligence — exposes totalDataRows on GraphQL (Labelbox/intelligence#29662)
3. this PR — SDK properties

Test

tests/unit/test_unit_model_run.py — fetch + cache, refresh re-fetch, and None for non-Foundry runs. 3 tests passing (pytest tests/unit/test_unit_model_run.py).

---

Note

Low Risk
Read-only additive SDK properties with opt-in GraphQL cost; depends on upstream Model Foundry GraphQL being deployed in the stack.

Overview
Adds lazy, cached ModelRun.total_cost (USD) and ModelRun.total_data_rows, populated on first read via a modelFoundryModelRunInfo GraphQL call (not on get_model_run). refresh_cost_and_usage() clears the cache for a live re-fetch.

Non–Model Foundry runs and missing payloads return None; ResourceNotFoundError / InternalServerError and empty responses are treated as no usage data. Transient errors (e.g. NetworkError) are not cached so a later read can retry.

Unit tests in test_unit_model_run.py cover fetch/cache, refresh, non-Foundry/missing payload behavior, and transient error propagation.

^{Reviewed by Cursor Bugbot for commit 9f91a2b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ff2aadd. Configure here.}