fix(server): avoid loading v2 transcripts for checkpoint effects

[juliusmarminge]

V2 checkpoint capture and rollback loaded the full thread projection, including conversation history and fork ancestry. Long conversations unnecessarily allocated and decoded large transcript payloads, and obsolete unrelated history could prevent checkpointing.

Add transactional reads for the exact checkpoint targets and switch capture/rollback to them. Finalization also uses checkpoint metadata for its workspace refresh instead of reloading the conversation after capture. Capture reads only its run, root node, scope, provider thread, and baseline readiness columns. Rollback retains the target attempt/provider turns, completed later runs/root nodes, bound session, and ready later checkpoints needed for existing rollback behavior. The in-memory store follows the same target selection.

Depends on #12338 and targets v2 from #2829.

Verification: two capture cases fail on the old full-read path and pass with these reads, while preserving concurrent delegated completion and at-least-once idempotency. SQLite/memory parity tests cover target selection, stale checkpoint scope/status filtering, and cross-thread ownership. An actual SQLite regression ignores obsolete transcript, unrelated run, and checkpoint file payloads. The seven focused projection/checkpoint/provider-control/finalization suites pass (52 tests). Server typecheck and focused lint pass.

A synthetic SQLite conversation with 200 × 128 KiB assistant items (25 MiB) measured these median read times over ten warmed reads on Node 24:

Read	Median
Full thread projection	139.25 ms
Capture context	0.44 ms
Rollback context	0.80 ms

This measures database projection work only, not Git capture/restore latency. No timing thresholds were added to tests. Backend-only change; no visual evidence applies.

Prepared with Codex.

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 05ed3bf.

This comment will update automatically after the next completed run.

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — The PR substantially replaces full transcript projection reads with new target-specific SQLite and in-memory selection logic across checkpoint capture, rollback, and finalization. Although the scope and tests are focused, these critical runtime and persistence-path changes merit human review.

Notes:

• Diff unchanged. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}