fix(postgres): batch pg push comparison reads (#331)

[rodboev]

Summary

• Keep the candidate-window and boundary-session behavior in internal/postgres/push.go unchanged for this PR, and batch the PostgreSQL-side comparison reads used to decide whether a candidate session can be skipped.
• Implement new batched loaders in internal/postgres/push_fingerprint.go for message aggregates, message content hashes, role/time fingerprints, message flags, message system ordinals, token fingerprints, tool-call aggregates, tool-call fingerprints, and usage fingerprints, with chunking inside the helper when session counts exceed ANY($1) practicality.
• Use the preloaded message and tool-call aggregates on the hot no-op path, and retry any comparison-preload SQL failure in a fresh transaction without the batched preload instead of continuing inside an already-aborted transaction.
• Add targeted regression tests in internal/postgres/push_test.go and internal/postgres/push_fingerprint_test.go to cover the new batch-driven skip decision path and helper behavior with empty inputs.

Scope

• Files changed are internal/postgres/push.go, internal/postgres/push_fingerprint.go, internal/postgres/push_test.go, and internal/postgres/push_fingerprint_test.go.
• No boundary/windowing semantics, no schema changes, and no changes to PR feat(sync): local-first multi-machine artifact sync #731 or broader sync-work areas.

Notes

• A focused PG comparison query-count assertion was not added because the existing harness does not expose a stable helper-call/query metric for this exact path without adding brittle test-only instrumentation.
• The review-driven follow-up keeps the existing non-batched fingerprint fallback, but now that fallback only runs from a clean transaction after preload failure instead of on the poisoned transaction that raised the preload error.

Fixes #331

[roborev-ci]

roborev: Combined Review (3fb4275)

No issues found.

---

Panel: ci_default_security | Synthesis: codex | Members: codex_default (codex/default, done, 4m59s), codex_security (codex/security, done, 1m49s) | Total: 6m48s

[wesm]

awesome, thanks for working on this!