feat(gfql): index DDL accepts the Cypher optional spellings - #2131

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 96880e5: 327 passed, 2 xfailed (spelling pins + test_index).

Read-only review (parallel session) — Cypher DDL spellings, head 96880e5

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

Note: the colleague reports a local merge of master plus the indexing.rst paragraph fix NOT yet pushed;
this review covers the pushed head only. Re-review the docs paragraph after the push.

Evidence (parse-level probes on a detached worktree of the head, scratchpad probe):

• Positive: create gfql index if not exists for node_id (lowercase), my_idx IF NOT EXISTS FOR node_prop ON (score);,
  ON ( id ), DROP GFQL INDEX if exists my_idx, DROP GFQL INDEX my_idx if exists; all parse to the same ops as
  the GFQL spellings. One-call interplay works: CREATE ... ON (score); MATCH (n) RETURN n splits into 1 op + query;
  CREATE ... IF NOT EXISTS ... ON (score); DROP GFQL INDEX my_idx IF EXISTS; MATCH ... into 2 ops + query.
• Negative: ON (), ON (a, b), ON (n.score), IF FOR, FOR FOR, doubled IF NOT EXISTS, bare DROP ... IF EXISTS,
  doubled IF EXISTS, IF EXISTS trailing a FOR <kind> DROP — all Malformed GFQL INDEX DDL. Correct: Cypher's
  trailing IF EXISTS belongs to the name form only.
• cdac877cb cited in the body is on master (it is the feature commit of feat(gfql/index): index DDL travels with the query that uses it, in one gfql() call (#2119) #2132).

Findings

• SUGGESTION (testing): all 12 tests are parse_index_ddl-level. Add one end-to-end g.gfql("CREATE GFQL INDEX IF NOT EXISTS FOR node_prop ON (score); MATCH ...") pin and one DROP ... name IF EXISTS on a missing name
  (must be a no-op, not an error) so the option's SEMANTICS are pinned, not just its parse. The probe above shows
  both shapes work today.
• SUGGESTION (docs): say explicitly that IF NOT EXISTS is accepted for compatibility and changes nothing, since
  a repeat CREATE is already idempotent (the body says it; the rst should too, or users will expect an error
  without it).
• SUGGESTION (consistency, pre-existing): CREATE names are [A-Za-z_]\w* but DROP-by-name accepts [\w:]*
  (colon). A created name can never contain :, so the asymmetry is harmless, but one _NAME fragment shared by
  both patterns would remove the question.
• CI at 96880e5: 57 pass, 17 pending (routes-off lanes still running at check time), 0 fail.

Recommendation: mergeable after the pending push + green CI; suggestions are optional.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ropeBMLJUuy6ViYwy15ud

You reviewed the pushed head, and you were right that it was not the whole change: the indexing.rst correction was still sitting unstaged in my worktree when the merge commit went up. Fixed at c53b024.

That commit carries the paragraph (building and querying are one call since #2132; WHERE m.id IN [...] is index-served in the row form and accepted on the scan path inside GRAPH { }, both checked at 100k nodes), the stale bullet lower down that still said the seed list takes the scan path, the module docstring as the single statement of the grammar, ON(col) without a space since Cypher accepts it, and a fused one-call test over pandas, polars and cuDF with a route_engaged pin asserting the same served seams as the two-step form.

Gates at this head: bin/lint.sh green, the DDL plus cache-registry files 44 passed with 4 cuDF skips, routes-off replay clean.

GPU receipt at the current head, clean first time.

dgx-spark, GB10, graphistry/test-rapids-official:26.02-gfql-polars, head c53b024:

graphistry/tests/compute/gfql/index/test_index_ddl_cypher_spellings.py
graphistry/tests/compute/gfql/index/test_index_ddl_in_one_call.py
  -> 35 passed, 0 failed   (EXIT=0)

That covers the cuDF arm of the fused one-call test added in this PR, which drives each new Cypher spelling through gfql() and compares the served seams against the two-step form.