repair=True can rescale an entire price table by 100x and invert the Repaired? flags #2923

Every claim below is backed by an offline reproduction that runs without touching the network, so none of it needs to be taken on trust.

When _fix_unit_switch detects a genuine unit change in part of a table, the correction is applied to the whole table and the Repaired? flag array is inverted wholesale rather than being restricted to the affected block.

Reproduction, using a fixture frame shaped like pence-quoted LSE data with one genuine 100x block:

• all 90 rows come back 100 times too large
• the 20 genuinely broken rows are flagged Repaired? = False
• 40 untouched rows are flagged Repaired? = True
• a control run with that branch skipped returns 0 of 90 rows wrong

Three smaller defects fall out of the same audit:

• history.py calls to_numpy() and then writes in place, raising ValueError: assignment destination is read-only on pandas 3.0.5. Your own fixture tests/data/SSNLF-1d-bad-div.csv triggers it.
• the crude 100x fallback rescales OHLC but not Adj Close, and default auto_adjust then reapplies the factor.
• a fund paying capital gains with no dividend on the same day hits a ZeroDivisionError.

I have a patch that fixes all four. Against your 42 committed golden fixtures it leaves 41 byte-identical and fixes the one that currently raises. Happy to open a PR if useful, or to leave the repro here if you would rather fix it your own way.