[Bug] GSI pagination with duplicate index sort keys returns no items on a composite base table

[Iamrodos]

Describe the bug

A Query on a GSI whose index sort key values are duplicated returns page one plus a LastEvaluatedKey, but resuming with that same key returns zero items. Nothing errors, so a paginating client reads page one and concludes it has everything.

The base table needs a range key. TestGSIOnHashOnlyBaseTable::test_paginate_duplicate_gsi_sort_keys already covers this index shape on a hash only base table and passes; adding a range key to that fixture reproduces the failure. Same class as #145.

Reproduced on both the sqlite and the default postgres backend at aefd1b6.

To Reproduce

EP=https://localhost:18443   # or your endpoint

aws dynamodb create-table --table-name esk \
  --attribute-definitions \
      AttributeName=pk,AttributeType=S AttributeName=sk,AttributeType=S \
      AttributeName=gsipk,AttributeType=S AttributeName=gsisk,AttributeType=S \
  --key-schema AttributeName=pk,KeyType=HASH AttributeName=sk,KeyType=RANGE \
  --global-secondary-indexes '[{"IndexName":"gsi","KeySchema":[{"AttributeName":"gsipk","KeyType":"HASH"},{"AttributeName":"gsisk","KeyType":"RANGE"}],"Projection":{"ProjectionType":"ALL"}}]' \
  --billing-mode PAY_PER_REQUEST --endpoint-url $EP
aws dynamodb wait table-exists --table-name esk --endpoint-url $EP

# Five items sharing one index sort key value, differing only in the base keys.
for i in 1 2 3 4 5; do
  aws dynamodb put-item --table-name esk --endpoint-url $EP \
    --item "{\"pk\":{\"S\":\"item$i\"},\"sk\":{\"S\":\"S\"},\"gsipk\":{\"S\":\"P\"},\"gsisk\":{\"S\":\"DUP\"}}"
done

# Page one: Count 2, plus a LastEvaluatedKey.
aws dynamodb query --table-name esk --index-name gsi --limit 2 --endpoint-url $EP \
  --key-condition-expression "gsipk = :p" \
  --expression-attribute-values '{":p":{"S":"P"}}'

# Page two, resuming from that LastEvaluatedKey verbatim.
aws dynamodb query --table-name esk --index-name gsi --limit 2 --endpoint-url $EP \
  --key-condition-expression "gsipk = :p" \
  --expression-attribute-values '{":p":{"S":"P"}}' \
  --exclusive-start-key '{"gsipk":{"S":"P"},"gsisk":{"S":"DUP"},"pk":{"S":"item2"},"sk":{"S":"S"}}'

Expected behavior

Page two returns the remaining items, as it does for the same index shape on a hash only base table.

Actual behavior

Page two returns nothing, with no error:

{"Count": 0, "ScannedCount": 0, "Items": []}

Environment

• ExtendDB version: 0.1.2, commit aefd1b6
• Operating system: macOS 15 (Darwin 24.6.0)
• Rust version (if building from source): 1.97.1
• Client SDK/driver and version: AWS CLI v2
• Deployment method (binary, container, source): source

Additional context

A regression test is the existing gsi_on_hash_only_base_table fixture with a range key added to its key_schema; the surrounding test body needs no change. Happy to open a PR.

DISCLAIMER: this issue was diagnosed with an LLM agent. I've reviewed and reproduced it.

Checklist

• I have searched existing issues for duplicates
• I have included the ExtendDB version and environment details
• I can reproduce this on current main (aefd1b6)