Summary

The default autoclean config generates a safe_int cleaning op for columns of
unhashable values (lists / dicts / sets).

base_summary_stats short-circuits value_counts to an empty Series for
unhashable columns (buckaroo/customizations/pd_stats_v2.py — restores #843;
value_counts on a list column is meaningless and falls back to an O(n²)
pairwise compare). pd_cleaning_stats then derives its int-parse stats from that
empty value_counts:

coerced_ser = pd.to_numeric(vc.index.values, errors='coerce', ...)   # empty
nan_sum     = (...).sum()                                            # 0
int_parse   = (length - nan_sum) / length                            # 1.0

so int_parse comes out 1.0, clears the 0.3 threshold in
cleaning_gen_ops, and a safe_int op is emitted for a column that is entirely
lists. Casting a list column with safe_int is meaningless.

Reproduction

import pandas as pd
from buckaroo.pluggable_analysis_framework.stat_pipeline import StatPipeline
from buckaroo.customizations.pd_stats_v2 import PD_AUTOCLEAN_DEFAULT_V2

df = pd.DataFrame({'lists': [['a'], ['b'], ['a', 'c'], ['d']]})
sd, errs = StatPipeline(PD_AUTOCLEAN_DEFAULT_V2, unit_test=False).process_df(df)
col = list(sd)[0]
sd[col]['value_counts']   # empty
sd[col]['int_parse']      # 1.0
sd[col]['cleaning_ops']   # [{'symbol': 'safe_int', ...}, {'symbol': 'df'}]

The same value_counts → int_parse interaction predates #876 (v1
DefaultSummaryStats also emptied value_counts and PdCleaningStats
computed the same); it became reachable on the default pandas widget once #876
routed it through @stat.

Scope

pd_cleaning_stats × base_summary_stats in pd_stats_v2.py. Gated to the
"default" autoclean config (CleaningConf / PD_AUTOCLEAN_DEFAULT_V2); the
default display path (PD_ANALYSIS_V2) doesn't compute cleaning ops, so this
only fires when a user selects default cleaning. The polars path (pl_cleaning_stats
casts to Int64) returns 0 parses for a list column, so polars likely does not
misfire — worth confirming and aligning the two paths.

Impact

• safe_int on an unhashable column is a no-op at best, an error at worst, and
  surprising either way.
• The root fragility is int_parse deriving from value_counts rather than the
  raw series — an empty value_counts should mean "can't decide", not "100% parses".

Found during review of #876; fix is out of scope there.