perf(runtime): batch Recall session reads to avoid N+1 queries and repeated decoding

[liuxiaocs7]

What happened

Recall narrows candidates to Sessions, then reconstructs each candidate's complete message view before ranking passages. A long Session therefore pays for repeated event reads and per-run SQL even when only ten passages match.

At ed38ccbb6bac474b93e4517f2a12a30209c108d5, a 20,000-Turn fixture returning ten passages takes about 1.60 seconds. One complete RuntimeReadModel read executes 80,002 SQL reads and 160,000 JSON.parse calls:

• The invocation inventory reads openings and separately fetches each invocation's terminal event.
• The ordinal reader loads and decodes all event payloads just to build an event-ID-to-ordinal map.
• The read model then rereads each run's events, partial snapshots and partial segments, including empty partial queries for completed runs.

There is also a fallback duplication: if the candidate path's corpus count returns null, collectHits restarts a full scan after already reading its candidates. A two-Session fixture reads first, first, second.

Expected: batch the complete-view reads within a consistent snapshot, decode each durable event once, and resolve count fallback without rereading candidate Sessions. Preserve ranking, redaction, passage contents/navigation, event order, legacy openings, partial presentation and corruption checks.

How to reproduce

1. Seed one Session with 20,000 completed Turns. Each Turn contains an invocation-opening event, one user text event and a terminal event, with Session ordinals in commit order.
2. Put deploy target in the first ten user events and unrelated text in the rest.
3. Run Recall with { terms: ['deploy'], limit: 10 } through SQLite and RuntimeReadModel.
4. Count executed all/get/iterate calls and JSON.parse calls during a separate getSessionMessages read. Counts are deterministic: 4 * turns + 2 SQL reads and 8 * turns parses for this fixture. Even two Turns require ten SQL reads and sixteen parses.
5. For the fallback case, offer only the first of two eligible Sessions as a candidate and return null from countSearchableMessages; record readMessages calls.

Implementation PR #5880 includes regression tests; benchmark results and measurement methodology are recorded in its description. SQL counts exclude transaction-control statements, candidate selection and corpus counting; timings cover Core Recall, SQLite and projection, excluding IPC, UI and fixture setup.

Environment

• Commit: ed38ccbb6bac474b93e4517f2a12a30209c108d5
• macOS 15.7.3, Apple M4 Pro, arm64
• Node.js v24.14.0; in-memory SQLite synthetic fixture
• Surface: Runtime Host Recall, shared by model tools and Desktop history search

Logs, screenshots, or additional context

Relevant sources: complete read model, invocation inventory, and Recall collection/fallback.

Related background: #5368, #5531 and #4677. This is separate from the long-term-memory item batching fixed by #5263. Search indexing and Host request cancellation are outside this issue; stopping after the first ten matches would change the existing ranking contract.

Prepared and submitted with Codex assistance on behalf of @liuxiaocs7.

[liuxiaocs7]

take