Preserve edge IDs for featureless graph sampling

[kmontemayor2-sc]

Summary

Restore edge-ID sampling for featureless graph partitions without requiring callers to switch to with_edge=False. Retain implicit edge IDs by default, with an explicit memory-saving opt-out for callers that do not need them.

Failure and fix

1. Partitions without explicit IDs, weights, or edge features enter the compact graph construction path.
2. That topology previously set _edge_ids = None. Existing callers requesting edge IDs could therefore reach the native sampler with a null pointer and crash. Absence of edge features did not make those sampling requests invalid.
3. The builder now retains original COO-position IDs in CSR order. A native guard also rejects nonempty ID requests when IDs were explicitly omitted, while allowing empty results.
   For example:
   Input COO edges: (1,7), (0,5), (1,6), (0,4)
   Implicit edge IDs: 0, 1, 2, 3

Sampling nodes [0,1]:
Neighbors: [4,5,6,7]
Counts: [2,2]
Edge IDs: [3,1,2,0]
The featureless-partition regression test exercises this through actual dataset construction and native sampling.

API and memory

• retain_edge_ids=True preserves compatibility and costs 8 bytes per edge. Callers using with_edge=False can explicitly disable retention.
• Task-config construction reads this option from trainer or inferencer arguments.
• Explicit IDs, features, and weights retain their existing construction path.
• build_csr_from_coo returns CsrBuildResult with named indptr, indices, and optional edge_ids fields. Direct tuple-unpacking callers must migrate to named fields.
• Construction writes into shared output tensors with bounded chunk temporaries.

Verification

63 focused CSR and topology tests passed, along with scoped lint, formatting, and type checks. The featureless edge-ID regression was also rerun independently and passed. Full-suite validation is deferred to CI.

[kmontemayor2-sc]

/all_test

[github-actions]

GiGL Automation

@ 17:02:10UTC : 🔄 E2E Test started.

@ 18:24:30UTC : ❌ Workflow failed.
Please check the logs for more details.

[github-actions]

GiGL Automation

@ 17:12:40UTC : 🔄 Python Unit Test started.

@ 19:01:54UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 17:14:23UTC : 🔄 Scala Unit Test started.

@ 17:24:48UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 17:15:15UTC : 🔄 Lint Test started.

@ 17:24:13UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 17:16:27UTC : 🔄 C++ Unit Test started.

@ 17:18:28UTC : ✅ Workflow completed successfully.

[github-actions]

GiGL Automation

@ 17:21:21UTC : 🔄 Integration Test started.

@ 19:04:38UTC : ✅ Workflow completed successfully.

[mkolodner-sc]

Thanks Kyle! Approving to unblock proved my comments are addressed.

[dsaini2-sc]

Thanks Kyle, I have a question -- why can't the default be to retain_edge_ids=False? loaders only set with_edge when the graph has edge features (see base_dist_loader.py:464), and those graphs already take GLT's build with edge ids. Your 0003 guard now anyways turns any other caller's request into a clear error rather than a crash. With True by default, every featureless graph pays 8 B/edge: with int32 columns from #788 the stored graph goes from 4 to 12 B/edge, and the build peak from 1.5x to 2.5x. A caller that does need edge ids could pass True explicitly. wdyt?

[dsaini2-sc]

also, #788 was just merged, so we will need to rebase this

[kmontemayor2-sc]

> why can't the default be to retain_edge_ids=False

Because there are some usages of with_edge being set to True explicitly, and I do not want to break those usages.

We will migrate those usages to use this opt-out, and then we can swag the flag after.

#788 should be merged in now :)

[dsaini2-sc]

Because there are some usages of with_edge being set to True explicitly
ahh, ok -- makes sense. Changes lgtm! thanks!