Fix SQLite bounded adjacency prefix parity

[forhappy]

Summary

• Order directional SQLite adjacency entries by canonical edge ID.
• Merge requested relationship-kind buckets before applying query bounds, matching the graph.json engine.
• Add an explicit adjacency capability marker so older endpoint-ordered sidecars fail with a rebuild instruction instead of silently returning a different truncated subset.
• Filter pinned discovery adjacency by relationship kind before truncation.

Motivation

On the Django repository, SimpleTestCase::assertRaisesMessage has 2,312 callers. With max-edges 500 and max-nodes 500, graph.json and SQLite both reported truncation but returned different edge subsets because SQLite bounded each endpoint-ordered index scan before the global edge-ID merge.

This change makes bounded selection backend-neutral and fails closed for sidecars that cannot provide those semantics.

Verification

Against the final compass-v0.3.22 base:

• cargo fmt --all -- --check
• cargo test -p compass-query --test store_engine truncated_callers_use_the_same_edge_id_prefix_for_store_and_json --locked -- --exact
• cargo test -p compass-graph --lib snapshot_without_edge_ordered_adjacency_requires_rebuild_for_directional_reads --locked
• cargo clippy -p compass-graph -p compass-query --all-targets --all-features --locked -- -D warnings

Additional qualification with the same source patch before the final rebase:

• cargo test -p compass-query --test store_engine --locked: 27 passed
• cargo test -p compass-graph --test store_snapshot --locked: 21 passed
• cargo clippy --workspace --lib --bins --locked -- -D warnings
• cargo test --workspace --lib --bins --locked
• sh scripts/check_product_boundary.sh
• ./scripts/qualify_code_graph_v1.sh --fixtures-only, including scale ceilings and 57-language deterministic qualification

Real-repository verification used Django commit c9eb16a87e60c305fb3651459639f647cce498db: 76,570 nodes and 192,081 edges. The formerly divergent bounded callers query produced byte-identical 500-edge responses from both engines with SHA-256 a11a5acbffb1e0c04a25f92d2e6bfdc73c660855f39d51efe1aa1edf730518df. SQLite completed in 11.13 seconds at about 127 MiB peak RSS; cold JSON completed in 102.09 seconds at about 1.20 GiB peak RSS.

The existing macOS large unwind-table linker warning was observed; it did not affect the checks.

Compatibility and documentation

The canonical graph.json contract is unchanged. New sidecars carry the edge-ID-ordered adjacency capability. Existing sidecars remain usable for status, validation, backup, export, and recovery, but directional store queries require rebuilding with:

compass update --force --store sqlite
compass store validate compass-out --format json

MIGRATION.md, COMPATIBILITY.md, the store operations guide, the query-engine implementation guide, and CHANGELOG.md document the behavior.

Checklist

• The change is focused and excludes unrelated formatting or generated files
• Tests cover changed behavior, or this pull request changes documentation only
• User-facing commands, flags, limits, and examples are documented
• Compatibility or migration effects are described
• No credentials, private source code, or sensitive report details are included
• I agree to license my contribution under MIT OR Apache-2.0
• I followed the Compass code of conduct

[forhappy]

CI follow-up: dependency-policy failed because crates.io yanked chacha20 0.10.1 after the existing lockfiles were published. Commit 9939af8 updates only the main and fuzz lock entries to non-yanked 0.10.2. Both cargo-deny policy invocations now report advisories, bans, licenses, and sources OK; cargo check --workspace --locked also passes.