fix: reject cross-segment ExclusiveStartKey in parallel Scan

[LeeroyHannigan]

What

Reject a cross-segment ExclusiveStartKey in a parallel Scan.

A parallel Scan (TotalSegments/Segment) accepted an ExclusiveStartKey belonging to a different segment, silently returning a truncated or empty page. This validates the start key against the same segment-assignment function used
for the scan and rejects a mismatch with a ValidationException naming the correct Segment, so a key returned as a LastEvaluatedKey for one segment is only valid when the same segment is re-scanned.

Why

DynamoDB validates that an ExclusiveStartKey supplied to a parallel scan belongs to the requested segment, and rejects it otherwise; the engine instead accepted any key and combined it with the segment predicate, silently yielding a truncated or empty page rather than an error. This is a correctness/parity gap for parallel-scan pagination. The behavior was verified against the real DynamoDB service.

Closes #

Testing done

• Scan integration test: a key that belongs to a segment is accepted as that segment's ExclusiveStartKey, and rejected with a ValidationException when supplied for a different segment.
• Verified verbatim against the real DynamoDB service (us-east-1): a same-segment ExclusiveStartKey paginates within the segment; a cross-segment key returns ValidationException: The provided starting key is invalid: Invalid ExclusiveStartKey. Please use ExclusiveStartKey with correct Segment. TotalSegments: 4 Segment: N.
• cargo test -p extenddb-storage-postgres unit pass; scan integration test passes; cargo clippy --workspace --all-targets and cargo fmt --all -- --check clean.

Checklist

• I have read CONTRIBUTING.md
• All tests pass (cargo test --workspace)
• Code is formatted (cargo fmt --check)
• Clippy is clean (cargo clippy -- -W clippy::pedantic)
• I have added or updated tests for new functionality
• I have updated documentation if behavior changed
• Breaking changes are noted below (if any)
• If this changes the wire protocol, Storage trait, auth model, on-disk
  format, or public CLI surface, an RFC has been accepted or is linked
  below. Otherwise, an ADR captures the decision (link below).

ADR / RFC: n/a -- no change to wire protocol, Storage trait, auth model,
on-disk format, or public CLI surface. Request-validation correctness only.

Breaking changes

None. A request that was previously mis-accepted (a cross-segment ExclusiveStartKey) now returns the DynamoDB-correct ValidationException; valid same-segment pagination is unchanged.

---

By submitting this pull request, I confirm that my contribution is made under
the terms of the Apache License 2.0 and I agree to the Developer Certificate of
Origin (DCO). See CONTRIBUTING.md for details.

[jcshepherd]

It's a little bizarre that we're executing a SQL query to do a mathematical calculation. I get why but in a perfect world I wonder if we'd be taking a dependency on a Postgres internal hash function for assigning segments.

[LeeroyHannigan]

Pushed 6d5d55f, which fixes the MongoDB integration failure by moving the refusal to where it can be uniform.

The failing test was correct. parallel_scan_rejects_cross_segment_exclusive_start_key got three items back on MongoDB where it expected a ValidationException, because the check lived inside the Postgres scan path only. MongoDB and SQLite fabricated a page from whatever their own segment predicate matched.

It can't be shared logic in the engine, because each backend assigns segments differently and privately:

backend	assignment
Postgres	(hashtext(pk)::bigint & 2147483647) % total
MongoDB	crc32(pk_text) % total
SQLite	rowid % total, not a function of the key at all

So it follows the ADR-0005 split: the engine owns the rule and the measured service message, and the storage trait gains one primitive, scan_key_in_segment, that each backend answers for itself. Postgres keeps its existing hashtext query (moved out of query_scan behind the trait method, no behaviour change), MongoDB is a pure CRC32 computation with no round trip, and SQLite resolves the key to its rowid.

@jcshepherd this also speaks to your point about the SQL query doing a mathematical calculation. It's still a query on Postgres, but the dependency on hashtext is now contained behind a trait method that documents it as backend-private, rather than being the shape the shared path assumes. If we ever want to move Postgres off hashtext, that's now a one-backend change with no engine edit.

Two decisions worth flagging for review:

Ok(true) means "cannot be proven foreign", not "proven local". SQLite's assignment depends on storage identity rather than key content, so a key it can't resolve has to be admitted: DynamoDB permits an ExclusiveStartKey for an item that no longer exists, and refusing it would reject a legitimate resumption after a delete.

The trait method is required, not defaulted. A default returning true would silently disable the refusal for any future backend, which is the failure mode this PR exists to fix.

Verification, live against a running server for each backend:

• sqlite 13/13 scan integration tests
• postgres 13/13
• mongodb 13/13

Each backend also has a negative control: sabotaging its scan_key_in_segment to return Ok(true) unconditionally makes the discriminating test fail with the fabricated page, so the test distinguishes the fix from its absence rather than passing incidentally.

fmt clean; clippy --all-targets -D warnings clean on the sqlite and postgres feature sets; 943 lib tests passed, 0 failed, 0 filtered on both. The mongodb clippy failure (install_backend dead code in crates/bin/src/main.rs) is pre-existing on this branch's base and reproduces identically with these changes stashed, so it's not from this work, but it will need addressing separately for the mongodb job to go green.