fix: safely sync large session files

[iHildy]

Summary

• replace oversized opencode.db bundles and legacy storage/message files with a versioned, content-addressed chunk representation
• validate strict count/size/mode/SHA-256 metadata, reject symlinks and corrupt/missing/extra parts, and reconstruct through staged atomic replacement
• protect the reserved chunk namespace with a versioned ownership marker, bounded known-location cleanup, and realpath containment
• detect oversized blobs in unpushed Git ancestry and fail closed with exact backup-branch/rebuild commands instead of silently rewriting history

This is a fresh replacement for the already-closed #48. It preserves plain legacy session files and normal non-session files.

Closes #45

Safety and compatibility

• chunks start above 50 MiB and are at most 40 MiB each
• v1 pointers cap reconstructed files at 4 GiB and 128 parts
• database, WAL, and SHM files validate and install or roll back as one bundle
• file modes are retained in pointer metadata
• stale/corrupt pointers, source mutation, filename collisions, namespace collisions, and ancestor symlink escapes fail closed
• old failed oversized commits are not rewritten automatically; the error first instructs creation of an exact backup branch, then rebuilds only unpushed history when the user elects to run the commands

Verification

• bun run check
• bun test — 129 passed, 0 failed
• bun run build
• independent final focused audit — 33 passed, 0 failed; no blocker
• realistic production-threshold two-home exercise: 52 MiB opencode.db and 51 MiB legacy message round-tripped A→B and modified B→A with matching SHA-256 and modes 0600/0640; largest committed blob was 41,943,040 bytes
• isolated two-instance GitHub E2E — passed; sentinel replicated and ephemeral private repo deleted
• isolated GitHub E2E with --enable-sessions — passed; repo session titles verified after link and pull, with expected restart requirement for the live local database; ephemeral private repo deleted

Expected harness-only warnings were the unsecured loopback test server, duplicate copied sandbox skill, live local DB restart requirement, and local AGENTS.md lag while repo replication was confirmed.