experiment: run DSPy without Deno

[isaacbmiller]

Purpose

Measure the CI and compatibility impact of removing Deno/Pyodide entirely and using Monty as DSPy's only first-party code interpreter.

This is an experiment, not a merge proposal. The base is pinned to Stanford DSPy main at c69136b2 for comparison with both upstream CI and the Monty-default/Deno-compatible experiment in #76.

Changes

• use MontyInterpreter by default in RLM, ProgramOfThought, and CodeAct
• remove PythonInterpreter, its Deno runner, public export, backend-specific tests, and docs
• port module-level interpreter contract tests to Monty
• remove Deno setup and pytest flags from test, dependency-range, and release workflows
• continue running optional-dependency tests separately

Net diff: 111 insertions, 2,516 deletions.

Validation

• focused Monty/RLM/ProgramOfThought/CodeAct suite: 171 passed, 2 skipped
• local full normal suite: 1164 passed, 247 skipped, 2 xfailed in 40.23s
• local extra suite: 35 passed in 7.52s
• CI test/build matrix passed on Python 3.10–3.14 twice
• docs build passed
• dependency-range minimum-direct passed; latest has the same current LiteLLM/FastAPI incompatibility as upstream main
• Ruff on touched Python files and git diff --check: pass

CI measurement

Comparisons:

• exact upstream Deno baseline: https://github.com/stanfordnlp/dspy/actions/runs/30657262551
• Monty-default while retaining Deno coverage: experiment(rlm): use Monty as the default interpreter #76
• Monty-only run 1: https://github.com/cmpnd-ai/dspy/actions/runs/30823626183
• Monty-only run 2: https://github.com/cmpnd-ai/dspy/actions/runs/30824092323

Matrix test job	Deno baseline	Monty + retained Deno	Monty-only run 1	Monty-only run 2
Python 3.10	198s	200s	123s	121s
Python 3.11	205s	185s	248s*	106s
Python 3.12	201s	194s	116s	116s
Python 3.13	163s	195s	107s	125s
Python 3.14	208s	210s	135s	242s*
Median	201s	195s	123s	121s

Across both Monty-only runs, the median test job is 122s: 79s (39%) below the Deno baseline and 73s (37%) below the Monty-default/retained-Deno experiment.

Step timing explains the gain. The normal suite remains approximately flat (median 92s baseline vs 96s Monty-only on the second run), while the optional/Deno phase falls from a baseline median of 91s to 13s after the 105 Deno tests are removed. The speedup is therefore specifically the eliminated Deno compatibility suite, not faster unrelated DSPy tests.

* Each run had one unrelated ~120s normal-suite outlier on a different Python version. Those outliers make total workflow wall time misleading (258s and 246s respectively), but do not affect the repeatable per-job median or the 78s reduction in the former extra+Deno phase.