feat(session): harden per-session SQLite sharding

[sjawhar]

Issue for this PR

Closes #20935

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This adds per-session-tree SQLite sharding for message, part, todo, and event storage while keeping session and project metadata in the global database. Old sessions still fall back to the global database when no shard exists yet, so the new path can coexist with pre-sharding data.

It also hardens the global write path with SQLITE_BUSY* retry/backoff, WAL/checkpoint monitoring, and targeted routing plus load coverage. During soak verification on current dev, project bootstrap writes also needed the same immediate transaction wrapper so worker startup would not surface rare SQLITE_BUSY_RECOVERY failures before shard activity began.

How did you verify your code works?

• cd packages/opencode && bun typecheck
• cd packages/opencode && bun test test/storage/retry.test.ts test/storage/pragma.test.ts test/storage/db.test.ts test/storage/wal-monitoring.test.ts test/storage/checkpoint-monitoring.test.ts test/storage/write-metrics.test.ts test/session/concurrency-load.test.ts
• cd packages/opencode && OPENCODE_LOAD_SOAK_MS=300000 bun test test/session/concurrency-load.test.ts -t "soaks multi-process contention for the configured duration"

Screenshots / recordings

N/A (storage/runtime change)

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

The search results only return the current PR (21579) itself across all keyword combinations. The other PRs found in one search (#16730, #20344) are related to memory/database optimization and cross-session learning, but not directly addressing the same SQLite sharding and retry/backoff hardening work.

No duplicate PRs found

[sjawhar]

Windows-only CI note for reviewers:\n\n- is still hitting the same Windows-specific file-lock/concurrency class we saw earlier: fails on while unlinking a shard file, and has one worker exit non-zero under concurrent install.\n- is not the same EBUSY pattern. The current failures are app/UI-side: times out waiting for , cannot find the control, times out selecting agent sound, and the run shows repeated plus a Solid proxy error in app state.\n- Linux CI is green for , , and , and the local verification for this branch is green. From the DB sharding perspective, the remaining CI red is Windows-only.\n\nLeaving this note here so the remaining Windows failures are documented explicitly in the PR while mergeability is discussed.

[sjawhar]

Windows-only CI note for reviewers (corrected):

• unit (windows) is still hitting the same Windows-specific file-lock/concurrency class we saw earlier: resolve-routing.test.ts fails on EBUSY while unlinking a shard file, and install-concurrency.test.ts has one worker exit non-zero under concurrent install.
• e2e (windows) is not the same EBUSY pattern. The current failures are app/UI-side: session-undo-redo.spec.ts times out waiting for revert.messageID, session-review.spec.ts cannot find the Expand all control, settings.spec.ts times out selecting agent sound, and the run shows repeated [e2e:pageerror] Object plus a Solid proxy error in app state.
• Linux CI is green for typecheck, unit (linux), and e2e (linux), and the local packages/opencode verification for this branch is green. From the DB sharding perspective, the remaining CI red is Windows-only.

Leaving this note here so the remaining Windows failures are documented explicitly in the PR while mergeability is discussed.