pandas2df: unify integer classification, add bigint option

[jefferis]

What

Replaces the two separate integer-conversion paths in pandas2df (the typed int64/uint64 path and the object-column path) with a single classify_integer_strings() decision point, so a column classifies identically regardless of how pandas typed it.

Tiering (mirrors reticulate::py_to_r() where it is faithful)

Each column maps to the narrowest base R type that holds it exactly, reserving bit64::integer64 only for values base R cannot:

magnitude	type
≤ 2³¹−1	integer
< 2⁵³	double (exact)
≥ 2⁵³	integer64

This drops two warts:

• the int64/uint64 asymmetry — small uint64 count columns were being force-promoted to bit64; they are now plain integer like int64.
• the object-vs-typed-path divergence — object-int columns used to become double and silently left overflow alone, while typed columns stop()-ped on overflow.

New bigint argument

pandas2df(x, bigint = c("auto", "integer64", "character")):

• "auto" (default) — the tiering above.
• "integer64" — every integer column → integer64, for a stable schema regardless of magnitude.
• "character" — the large (≥ 2⁵³) tier → character, as data.table::fread() offers.

uint64 values beyond the signed 64-bit range cannot be held by integer64 (which is signed); they are returned as character — the only faithful option — with a warning unless bigint = "character".

Testing

• New tests cover every tier boundary and all three modes plus the overflow warning.
• Verified end-to-end against a real pandas frame spanning small/mid/large int64, small uint64 counts, and overflowing uint64.
• R CMD check clean (0/0/0); full matrix (Python 3.9 / reticulate-default / 3.13) run on the branch.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com

🤖 Generated with Claude Code