perf(host-core): replace session updated_at index with composite index

[xiaobaZeo]

Fixes #1434

Motivation

When listing sessions, the ORDER BY s.updated_at DESC, s.id DESC clause could not fully utilize the existing idx_sessions_updated (which only covers updated_at DESC). This forces a TEMP B-TREE sort in SQLite memory.

Changes

• Bumped SCHEMA_VERSION to 22
• Added a database migration from v21 to v22 that replaces idx_sessions_updated with idx_sessions_updated_id(updated_at DESC, id DESC)
• Updated schema.rs and database startup routing in repositories.rs
• Added unit tests for the migration.

🤖 Generated with Claude Code

[muzimu217]

Verified the perf claim against today's main (6881a2bfd) — the principle is sound and the composite index matches the real query shapes:

• Hot session-list paths really do ORDER BY s.updated_at DESC, s.id DESC (crates/host-core/src/plugin_sessions.rs:861-865, session_collaboration/projections.rs:274) — exactly what idx_sessions_updated_id(updated_at DESC, id DESC) serves without a temp b-tree. This is where the reported latency comes from.
• crates/host-core/src/sessions.rs:1291 orders by updated_at DESC alone — still served by the new index as a prefix, so dropping idx_sessions_updated is safe there.
• crates/host-core/src/session_search.rs:114 uses updated_at DESC, id ASC — the composite can't serve that id ASC tiebreak, but the old single-column index didn't either; no regression.
• Migration wiring is correct as far as I can read it: the v21→v22 step follows the v20→v21 pattern (backup + transaction + user_version bump), the legacy-path re-read of migrated_version after v20→v21 keeps chained upgrades on the loop path, and updating the v20_database_migrates_... fixture to pre-create the old index name is the right call now that fresh schema creates the composite directly.

Two things standing between this and mergeable, both visible in CI:

1. cargo fmt --check is the only Rust failure (the failed step is "Check formatting"; tests/lint themselves aren't the blocker) — the new test has trailing-whitespace blank lines. cargo fmt fixes it.
2. Base gate: needs a rebase onto latest main.

One suggestion, not a blocker: add an EXPLAIN QUERY PLAN regression asserting the session-list query uses idx_sessions_updated_id without USE TEMP B-TREE FOR ORDER BY. That pins the actual invariant #1434 is about (query plan), not just the DDL — a future query reshuffle that silently reintroduces the memory sort would otherwise pass this test suite.

[muzimu217]

Update with local toolchain verification (rustup installed here, so we can now run the Rust gates directly):

• cargo fmt --check reproduces exactly two spots: crates/host-core/src/db/migrations.rs:966 — the execute_batch string literal needs a trailing , after ...id DESC);" — and crates/host-core/src/db/tests.rs:1669-1683 — the execute_batch call needs rustfmt's method-chain indentation. cargo fmt -p host-core fixes both.
• The migration logic itself is correct locally: migrates_v21_to_v22_replaces_session_index and the updated v20_database_migrates_the_session_checklist_with_a_backup fixture both pass on this head (cargo 1.99.0, stable).

So after cargo fmt + rebase onto latest main, everything we can check locally is green. The EXPLAIN QUERY PLAN regression suggestion from the review stands as a non-blocking follow-up.