fix(gfql): preserve explicit whole-entity projection provenance

[lmeyerov]

Unlabeled whole-entity results can lack physical identity columns, leaving the adapter unable to distinguish RETURN x from property returns. This records projection kind and source presence independently of identity, preserving present entities with all-null properties, absent entities, OPTIONAL null insertion and reentry row alignment. Property/text projections clear provenance; input graphs remain unchanged.

Presence alignment uses native vectorized gather/reindex rather than a DataFrame per result segment. The paired adapter change is graphistry/tck-gfql#201; merge this product PR before that adapter PR. Tracks #2074 and graphistry/tck-gfql#200.

Final product: e1e22d196aa21b27fa748108cc49e31284252e47. Paired TCK: 9b5079eecc116dd334d32c4373903e56fc13b0a9. No scenario or expected-result snapshots changed.

Validation:

• All 85 exact-head product CI checks succeeded. Two consecutive clean source-review waves, including pandas2/3 adversarial checks for marker/property collisions, mutation isolation, duplicate gathers, stale presence and reentry slicing.
• Final DGX tests: cuDF25/26 each 50 passed / 14 skipped; polars-gpu 64 passed. Cases cover labels, renaming, duplicate properties with distinct IDs, DISTINCT/counts, properties, empty results and null presence.
• Final exact-pair TCK CI: 4169 passed, 13 skipped, 689 expected failures. Local mixed pandas/native-Polars runner plus provenance tests: 3840 passed / 5 skipped / 689 expected failures; 327 native Polars declines are not execution passes.
• Combined aggregate/validation/entity integration: 223 passed / 237 optional-backend skips. Focused pandas2/3 entity suites and supported device runs supplement this integration.

Performance: final OPTIONAL comparisons measured +6.4%/+1.58% at100/1000 seeds after replacing per-segment allocation. One separate-process ordinary-projection comparison exceeded the investigation threshold; an interleaved control measured +0.195ms/+6.12%, below the joint threshold (>20% AND >0.2ms). This retains measurement variance and modest overhead as limitations; no GPU throughput improvement is claimed.

Ruff/type-hygiene guards pass. Five local mypy dependency-profile diagnostics reproduce baseline. Exact commands, source/image pins, XML, review reports and raw timings are retained under local plans/gfql-2073-followups/. Ready for maintainer review; not merged automatically.

Current-base verification after2072

Product head10e3b4484d5f27715e10dc2a2e869e06df0101ca includes master2ca98922c; the Polars source-node-ID argument and entity presence/provenance are both preserved. Two clean catch-up source waves complete. Local entity/projection/reentry422passed; combinedTCK3840passed (327explicit Polars declines). Exact pairCI34733762397 verifies this product head with TCK9b5079eecc116dd334d32c4373903e56fc13b0a9:4169passed,13skipped,689xfail.

Earlier device/performance evidence above uses historical product heads. Product2079 is draft pending refreshed final device/performance and productCI gates. TCK201 remains dependent on2079; both remain unmerged.