refactor: make v2 the only production query engine

[azimafroozeh]

What & why

Production reads still default to v1, leaving v2's planner and query fixes opt-in. This makes v2 the only production query engine in v0.12.0.

• Keep engine=v2 as the default and only accepted value across embedded sessions, CLI, HTTP and process settings.
• Move v1 into the frozen, unpublished omnigraph-reference-engine, with GQT as its only dependent. Compiler, core and catalog remain shared.
• Add --- expect same as v1 alongside written row expectations. Reference errors or different results fail the step; DST skips the comparison.
• Convert 28 traversal and 22 search tests to GQT and remove the redundant engine-v2 CI mode.

Compatibility changes: engine=v1 is rejected; bm25() and nearest() ordering on traversal destinations returns a bad-request error; BM25 ties use identity order; multiple text predicates on one scan are conjoined.

Backing issue / RFC

• Amends GQ logic tests, Session settings and Shared expression model. Their recorded status is still draft; acceptance is unconfirmed.

Checklist

• Change is focused on production v2 and explicit v1 reference comparisons.
• Tests added/updated for behavior changes.
• Public docs, OpenAPI and release notes updated.
• Invariant review complete; the current branch still lacks the multi-hop limit fix from PR 790.

Local verification

Recorded on September 27; not rerun for this description:

• cargo test -p omnigraph-gqt --locked: passed, including 214 corpus cases and harness suites.
• cargo test --locked -q from crates/omnigraph-dst: 108 passed, 30 ignored.
• cargo test --workspace --exclude omnigraph-gqt --exclude omnigraph-dst --locked --features omnigraph-engine/failpoints,omnigraph-cluster/failpoints --no-fail-fast -q: three CLI failures, reproduced individually; not established as pre-existing.
• cargo clippy --workspace --all-targets --locked -q --message-format=short -- -D warnings -W clippy::dbg_macro: passed.
• cargo fmt --all --check, bash scripts/check-agents-md.sh and python3 scripts/check-docs.py: passed.

Notes for reviewers

• The dependency-alias guard and reference-refusal advice findings are fixed. The multi-hop fix is merged upstream but absent from this tree.
• Release performance remains unverified. The recorded comment audit is non-green.
• Cost goldens need owner review: warm-read data reads change from 2 to 4; DST Lance GET totals decrease.