perf(gfql): row-evaluator text sniffers decide on a sample before scanning a column (#2116 3c) - #2128

Real-GPU receipt (dgx-spark, graphistry/test-rapids-official:26.02-gfql-polars, cudf 26.02.01 / cupy 13.6.0 / polars 1.35.2) at a47a332: 102 passed (sniffer pins, ORDER BY null placement, temporal coverage, test_engine_polars_gpu); CHANGELOG-only commit since (853ec1a) changes no code.

Read-only review (parallel session) — sample-first sniffers, head 853ec1a

Nothing here touches the branch; findings only, fixes deferred.

What it changes: order_detect_temporal_mode tries each all-rows regex conjunction on head(16) of the text first and
runs the full column scan only when the sample passes; _gfql_series_is_list_like / _gfql_series_is_mapping_like
return False as soon as a sampled non-null value is an actual str.

Correctness check (by reading the full functions on the head, not just the diff):

• Temporal: each mode is all() over rows, so a failing sample is a sound negative; a passing sample still runs the
  full scan, so a column that fails only past row 16 keeps its old answer. Mode ORDER is preserved
  (date, datetime, time, date_constructor, datetime_constructor, time_constructor). Sound.
• List/mapping: the old code computed actual_string = (series == series.astype(str)) and forced those rows False
  before .all(); a str value always equals its own astype(str), so any sampled str already made the
  old answer False. The early return is exactly equivalent. Stringified lists ("[1, 2]") were never list-like
  by this probe before (they are actual strings) and still are not. Sound.
• Engine parity: uses .head, series_str_fullmatch, _gfql_series_to_pylist — all already engine-polymorphic;
  cuDF Series.head exists. No new pandas-only call. Not run on cuDF here (no runnable cuDF locally) — the PR
  body's local numbers are pandas; a dgx cuDF smoke of the three probes would close that.

Tests (test_sniffers_sample_first_3c.py, 18): answers preserved per family with and without a trailing None;
failure past the sample still counted (both orders); native dtypes untouched; counting pin via monkeypatched
series_str_fullmatch asserting max(calls) <= _GFQL_TEMPORAL_SNIFF_SAMPLE (structured, not message-matching). Good.

Findings

• SUGGESTION (testing): the body cites a 176-series old-vs-new equivalence sweep with 0 mismatches, but the sweep is
  not committed. Commit it as a parametrized test (or a scripts/ check) so the equivalence stays pinned when the
  regexes change.
• SUGGESTION (perf evidence): the win is only realised when the sniffer is reached, i.e. the cross-alias OR
  shape. A pyg-bench pin is not needed for this PR by itself, but the colleague's [FEA] typecheck invalid api=... value #285 sentinel set already
  targets 3a/3c — confirm it includes an IN ... OR row-form point so the 545 ms → ~? ms claim has a receipt.
• SUGGESTION (scope note): the body says the unseeded MATCH ... RETURN b LIMIT 5 binding-table cost is filed
  separately — link the issue number in the PR.
• Repo conventions: no new broad excepts (the two except Exception around series == text / astype(str) are
  pre-existing), vectorized, pure. No new source file (no coverage-baseline entry needed).

Recommendation: merge once CI is green; nothing blocking found.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ropeBMLJUuy6ViYwy15ud

Your three suggestions plus my own pass found six items; all six are fixed at 0c741d0, and two of them were larger than reported.

The "identical answers" claim was false, and the CHANGELOG entry was in a shipped release. Both confirmed. The entry has moved from ## [0.59.1 - 2026-10-03] into [Development]. The claim is now a measurement: an A/B of master against this head over 30 adversarial column shapes (nulls first, a failing value past the sample, all-null, empty, shorter than the sample, category, string dtype, bytes, Decimal, nested lists, mixed list and map, tuples, paren text, arrays) gives one divergence, and it is a correction rather than a regression:

master: list_like([np.array([1,2]), "[3, 4]"]) = True
head:                                          = False

Master answered True only because series == text raises on the array and the error was swallowed, which left the real string row undetected. A column mixing arrays with list-looking text is not a list column. Pinned in test_an_array_column_with_list_looking_text_is_not_a_list_column, and stated in the CHANGELOG as the one changed answer.

Sample-first was incomplete, and on the expensive half. The temporal sniffer still ran astype(str) over the whole column before taking head(16) — rendering is the cost being avoided, so the sample is now rendered on its own and the column only when a mode survives. The list and mapping probes now regex their sample before rendering the column too. Measured on 100k rows, warmed, median of the same harness:

No cuDF pin for a compute/gfql/row/ change. Added: the three sniffers answer the same on six column shapes, parametrized pandas and cuDF.

List and mapping engagement rested on timing. Now mechanism-pinned the way the temporal one was: test_the_list_and_mapping_probes_render_only_the_sample counts the rows handed to the regex and asserts the column is never rendered when the sample rules it out.

Bench attribution. Standing: the only arm with an A/A control is the combined 3a+3c run (1908.8 ms against master's 3297.8 and 3244.8, A/A spread 1.00 to 1.03), and the threshold is a master times 1.5 regression bound rather than an improvement pin. I am not claiming an isolated 3c number from it; the four local before/after pairs above are what this PR stands on.

Also fixed from your list: the comment run the density guard flagged in ordering.py is one line.

GPU receipt at the current head.

dgx-spark, GB10, graphistry/test-rapids-official:26.02-gfql-polars, head c9e75d5:

graphistry/tests/compute/gfql/row/test_sniffers_sample_first_3c.py
graphistry/tests/compute/gfql/test_engine_polars_gpu.py
  -> 92 passed, 0 failed   (EXIT=0)

The first pass at 0c741d0 failed one cell, and it was my own test's fault rather than the code's: the cuDF parametrization asserted pandas' answer for a mapping column. Probing a master-tip tree and this head side by side on the same GPU gives the identical result on both:

maps: temporal=RAISE NotImplementedError  list=RAISE NotImplementedError  map=True

cuDF holds mappings as a struct column and cudf 26.02 cannot render a struct to text, so the text probes raise there on master too. The shared row could never pass, so at c9e75d5 it is replaced by a per-engine pin of the real divergence: pandas renders and declines, cuDF raises, and the mapping probe answers True on both.