fix: decompose pipeline KeyError when constraint strings don't exactly match between stages

[planetf1]

Problem

docs/examples/m_decompose/python/python_decompose_example.py fails with a KeyError in finalize_result. Found when explicitly running slow tests on a Mac M1 (32 GB).

KeyError: '* The blog post must have a catchy title'

Traceback points to cli/decompose/pipeline.py:431:

"val_strategy": constraint_val_data[cons_str]["val_strategy"],

Root Cause

constraint_val_data is keyed on the raw strings returned by constraint_extractor (e.g. "The blog post must have a catchy title"). Later, subtask_data.constraints is populated by a separate model call which can return the same constraints with slight formatting differences — in this case a * bullet prefix. The dict lookup then fails.

There's no normalisation of constraint strings between the two stages.

Suggested Fix

Strip leading punctuation/whitespace when building or looking up constraint_val_data keys, or normalise constraint strings at extraction time so both stages agree on the format.

Flagging for @AngeloDanducci as the main author of the pipeline.

[AngeloDanducci]

I'm the author of the example but not the primary author of the pipeline 😅

I'll take a look at the changes in the upgraded pipeline last week - I think we solved a similar issue previously in #569