Enable real-time Docent ingestion in custom_model

[oalmatov]

Summary

Wires real-time Docent ingestion into the finance-agent custom model: passes `docent_ingest=True` to `agent.run` so transcripts stream into Docent as the run executes (grouped under the platform's `run_id`).

Changes

• `pyproject.toml`: `model-library==0.1.19` → `0.1.23` (first PyPI release where `Agent.run` accepts the `docent_ingest` kwarg).
• `uv.lock`: refreshed to pick up new resolutions (`model-library 0.1.19 → 0.1.23`, `anthropic 0.86.0 → 0.97.0`, `+httpx-aiohttp 0.1.12`).
• `finance_agent/custom_model.py`: pass `docent_ingest=True` on the `agent.run` call. `question_id` and `run_id` were already forwarded from the platform.

Why

Companion to the benchmarks-side PR (vals-ai/benchmarks#1578) that enables real-time Docent ingestion across custom models. Finance-agent was blocked because its model-library pin predated the `docent_ingest` kwarg; bumping unblocks it.

Notes

• Independent of the in-flight `oa/docent-ingestion` branch (Valkyrie route). This is the smallest change to enable the real-time path; the Valkyrie work can land separately.
• Docent failures don't block the run (model-library swallows ingestion errors and warns).

Test plan

• `make install && make style && make typecheck` pass.
• On the next finance_agent benchmark run, verify transcripts appear in the Docent collection at `https://docent.transluce.org/dashboard/{run_id}\`.

[assert-app]

Review on Assert →