Reconcile time zones before the as-of join, so trade enrichment stops failing silently - #97

… failing silently

enrich_trades_with_signals joins trade times against signal timestamps, and the two sides
reach it from different places: trade times come out of the engine carrying the bar clock
without its zone, while a prediction panel is read from parquet with its zone intact.
join_asof refuses a naive left key against an aware right key, and BacktestProfile wraps
this call in a try that degrades to an empty frame with a UserWarning - so the reader does
not get a traceback, they get a tear sheet quietly missing its ML trade-alignment charts.

Found in case_studies/nasdaq100_microstructure/20_strategy_analysis.ipynb, whose committed
output carries 'Prediction enrichment failed: datatypes of join keys don't match -
entry_time: datetime[us] on left does not match timestamp: datetime[us, UTC] on right'.

A naive column here is the same wall clock with the zone dropped, so it is localized and
never shifted. Measured on that case study: reading the naive trade times as UTC matches 703
of 905 trades to an exact (symbol, instant) row in the prediction panel, and reading them as
America/New_York matches 199; the residual 202 are trades entered between two prediction
instants, which is what a backward as-of join is for. Shifting instead of localizing would
still return rows, so getting this backwards would be silent rather than loud. Where both
sides are aware and disagree the signals are converted, because two aware columns already
name instants.

The negative control neutralizes the helper with a monkeypatch and asserts the join still
raises, so the other six tests are evidence that this path is load-bearing rather than
assertions polars would satisfy on its own.

…source

The timezone-alignment change moves the engine tree digest, and provenance.py stamps
that digest into every retained record, so all 68 read stale against the working tree.
Correctness is the kind of evidence that is genuinely re-derived rather than certified:
it is cheap and tolerant of a contended machine.

66 pass, 2 unsupported, unchanged from the previous matrix. Compared leaf by leaf against
the committed copy: 231,740 values, 201 differ, and every one of the 201 is a digest, a
commit or a timestamp. No correctness value moved.

Every record is stamped dirty:false, so this was produced from a clean tree.