Batch bounded sequences of small asset writes through Git

[flyingrobots]

Problem

A Think-sized git-warp v18-to-v19 rehearsal still showed short-lived Git children after commit reads and commit writes were moved onto persistent plumbing. The remaining hot path translates thousands of small patches and attachments through repeated assets.put() calls.

On git-cas v6.5.5, ChunkRepository.storeChunk() calls persistence.writeBlob() once per chunk, and GitPersistenceAdapter.writeBlob() runs one-shot git hash-object -w --stdin. git-cas already has a bounded writeBlobs() path backed by one scoped fast-import session, but ordinary asset ingestion does not consume it across independent small assets. This is distinct from #98, which preserves the one-shot fallback when duplex sessions are unavailable.

Desired outcome

Provide a git-cas-owned, bounded batch boundary for a sequence of independent asset writes so callers such as a migrator can submit many small streaming assets without receiving raw Git/session authority. The implementation may use putMany, an explicit scoped asset batch, or another semantic API, but process count should scale with bounded flush windows rather than one-chunk asset count.

Required invariants

• Sources remain streaming; batching has explicit item and byte bounds.
• Results preserve input order and produce the same manifests, trees, handles, and Git OIDs as individual put() calls.
• Chunk hashing, encryption, compression, deduplication, and integrity metadata remain unchanged.
• A failed item cannot silently expose partial success as a complete batch.
• Scoped fast-import checkpoints and closes before returned handles escape.
• Sessionless Plumbing continues to use the typed one-shot fallback from Fall back to one-shot Git I/O when Plumbing lacks duplex sessions #98.
• Callers never receive raw Git plumbing or a mutable persistence adapter.

Acceptance criteria

• A real-Git test stores a bounded sequence of one-chunk assets with O(windows) Git write processes rather than O(assets).
• A mixed-size test proves byte/item bounds and backpressure.
• Golden tests prove byte-identical manifests, trees, handles, and OIDs versus repeated put().
• Failure and cancellation tests prove deterministic cleanup and result semantics.
• SHA-1 and SHA-256 repositories are covered.
• A git-warp migration witness records child-process count and wall-time before and after adoption.

Non-goals

• A persistent repository-lifetime fast-import process.
• Exposing raw Git sessions publicly.
• Buffering an entire migration or large attachment in memory.
• Expanding git-warp PR #825; this is follow-up performance work.

[coderabbitai]

🔗 Related PRs

#62 - feat: add opaque application asset handles [merged]
#63 - Add structured bundle and page storage [merged]
#91 - perf: reuse bounded Git object sessions [merged]
#95 - Preserve coherent Git object sessions [merged]
#100 - Batch page retention in staging workspaces [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!