experiment(rlm): use Monty as the default interpreter

[isaacbmiller]

Purpose

Measure the CI and test-suite impact of replacing RLM's default Deno/Pyodide interpreter with a first-party Monty interpreter.

This is an experiment, not a merge proposal. The temporary base branch is pinned to Stanford DSPy main at c69136b2 so the diff and CI comparison stay focused.

Changes

• add a Monty-backed CodeInterpreter with persistent sessions, host tools, typed SUBMIT, lifecycle/error handling, mounts, and resource-limit passthrough
• make MontyInterpreter the default for dspy.RLM
• make pydantic-monty>=0.0.19 a required dependency for this experiment
• move RLM integration coverage from the Deno-only phase into the normal test phase using the Monty default
• retain PythonInterpreter and its explicit Deno integration tests

Local validation

• 130 passed, 18 skipped for RLM + Monty interpreter tests in 16.58s
• 1131 passed, 352 skipped, 2 xfailed for the normal repository suite in 39.85s with xdist
• focused Ruff checks and git diff --check pass

CI result

The experiment matrix passed on Python 3.10–3.14. Compared with the exact-base upstream run:

Job	Deno default baseline	Monty default experiment	Delta
Python 3.10	198s	200s	+2s
Python 3.11	205s	185s	-20s
Python 3.12	201s	194s	-7s
Python 3.13	163s	195s	+32s
Python 3.14	208s	210s	+2s
Median test job	201s	195s	-6s
Mean test job	195.0s	196.8s	+1.8s
Workflow wall time	211s	215s	+4s

Conclusion: total CI time is effectively neutral in this single-run sample; differences by Python version are dominated by normal hosted-runner noise. Monty adds no visible critical-path regression despite moving Monty-backed RLM integration cases into the normal test phase.

The dependency-range workflow's minimum-direct job also passed. Its latest-dependencies job failed in LiteLLM startup because the latest FastAPI removed get_flat_dependant; the same exact-base upstream dependency run failed independently with the same timeout, so this is not attributable to Monty.