perf(gfql): answer a numeric datetime searchAny on cuDF from the component index too

[lmeyerov]

What

The datetime component index was pandas-only; a numeric searchAny term over a cuDF datetime
column still rendered every row to text on every search. The index needs nothing pandas-specific
— integer components, comparisons, a gather — so it now lives where the column lives: numpy
arrays for a pandas column, cupy arrays on the device for a cuDF one, chosen from the column's
own module. The "cudf" not in type(s).__module__ gate in search_any.py is gone.

cuDF stays UTC-only, for the reason the render is: libcudf's strftime ignores tz_convert,
so no other zone can be named on the GPU. The index declines any other zone with the same
CudfTemporalTzUnsupported the render raises, rather than indexing a false one.

The cache key gains the device: identical bytes on host and GPU are the same instants but not
the same arrays. The cuDF digest is sha256 over a host copy — measured on dgx, a device-side
order-sensitive digest is 30.8 ms but needs a 480 MB powers table, twice the index it protects,
when memory is already the binding constraint; the copy+sha256 is 124.7 ms and needs nothing.

Why this matters: the production table search runs on cuDF frames

The viz's remote table search (GRAPHISTRY_TABLE_QUERY_REMOTE, FEP compute_table_query) works
on cuDF frames, so the pandas index never reached it. Measured on dgx at 30M rows, same frame,
FEP's own match logic lifted verbatim, every term asserted equal:

per keystroke	FEP cuDF (GPU render + scan)	pygraphistry cuDF before	pygraphistry cuDF after
2 (five fields)	861 ms	3,964 ms	133 ms
2024	848 ms	3,935 ms	255 ms
typing sequence	4,259 ms	19,745 ms	783 ms

30× on our own cuDF path; 5.4× faster than the production FEP path on FEP's frames. The
residual ~125 ms is the host-copy digest; a caller-supplied cache key (FEP has dataset ids) would
take a keystroke to ~3–70 ms.

Typing

compute/typing.py's ArrayLike/ArrayNamespace protocols gain the members the kernel calls
(equal, logical_or, logical_and, take, int8/int16; __mod__, all, any, copy),
so the module-agnostic code is typed against the same protocols gfql/index/engine_arrays uses.
The type-hygiene guard rejected cast(...) in favour of engine-agnostic annotations with
localized type: ignores — followed.

Tests

The cuDF class (TestOnCuDF, 9 tests) is skipped without a GPU and validated on dgx: 80 passed
— exactness against render_datetime_cudf over every shape and term, the route pinned on a cuDF
frame, the non-UTC decline, and that host and device columns with the same bytes do not share an
entry. CPU suites unchanged. Changed-line coverage vs the stack top 52/61 = 85%, no pragmas:
the nine uncovered lines are the GPU-only arms, which dgx covers.

Stack

PR 4 of 4, base fix/gfql-datetime-index-cache-thrash (#2112). Lands after #2110 → #2111 →
#2112; GitHub retargets the base on each merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ropeBMLJUuy6ViYwy15ud

Check count

This PR will show 68 checks rather than the 70 on #2110: codeql-analysis.yml runs on pull_request only for branches: [master], so a stacked PR whose base is another branch skips it until the base retargets to master after the PRs below it merge. The 68 are the full CI Tests matrix.