perf: preserve Git object sessions across coherent writes

[flyingrobots]

Problem

@git-stunts/git-cas@6.5.2 opens typed cat-file and mktree sessions, but GitPersistenceAdapter retires both sessions after every one-shot blob write, retires both after every bounded bulk write, and retires cat-file after every tree write. In a 128-node git-warp retained materialization this converts nominally persistent protocols back into process-per-object behavior.

Measured baseline:

Metric	6.5.2
Total Git child processes	558
cat-file processes	80
mktree processes	82
hash-object processes	140

Audited Git coherence matrix

Raw protocol tests were run against Git 2.50.1 and 2.39.5, in SHA-1 and SHA-256 repositories.

Existing session	Subsequent write	Safe to preserve?	Evidence
cat-file --batch-command	loose hash-object -w	yes	same process reads the new object
cat-file --batch-command	checkpointed fast-import pack	yes	same process refreshes and reads the new packed object
cat-file --batch-command	loose mktree output	yes	same process reads the new tree
mktree --batch -z	loose hash-object -w	yes	pack-primed process validates the new loose object
mktree --batch -z	checkpointed pack above fastimport.unpackLimit	no	pack-primed process rejects the later packed object

This matches upstream Git implementation: cat-file object lookup permits a second read that refreshes packed state, while mktree requests OBJECT_INFO_QUICK.

Proposed single change

Preserve:

• cat-file across all successful CAS object writes;
• mktree across one-shot loose blob writes;
• both sessions across loose tree writes.

Retire only:

• scoped fast-import after every bounded bulk write, as today;
• mktree after a bounded bulk write, because the output may be a newly created pack.

Measured candidate

The audited candidate, applied only to an ignored installed package for A/B testing:

Metric	6.5.2	Candidate	Delta
Total Git child processes	558	401	-28.1%
cat-file processes	80	4	-95.0%
mktree processes	82	1	-98.8%
Three-run median materialization test time	14.24s	12.72s	-10.7%
Three-run median process CPU (user + sys)	15.74s	13.83s	-12.1%

Process counts are deterministic; timings are local evidence and must not be represented as a cross-platform guarantee.

Acceptance criteria

• Unit tests prove the exact retirement matrix using protocol doubles.
• Real-Git integration proves both safe reuse and the required post-pack mktree retirement.
• SHA-1 and SHA-256 object identity remains unchanged.
• Individual blob writes remain one-shot hash-object operations.
• Scoped fast-import still checkpoints and closes before OIDs escape.
• Lifecycle, idle retirement, invalidation, and explicit close semantics remain intact.
• Existing full unit/integration/platform gates pass.
• A committed witness records process-count and timing scope.

Non-goals

• No public API change.
• No persistent fast-import process.
• No git-warp cache or CAS ownership change.
• No workspace RootSet coalescing or retention-policy change.
• No attempt to eliminate the remaining one-shot hash-object, ref, or commit processes.

[coderabbitai]

🔗 Related PRs

#62 - feat: add opaque application asset handles [merged]
#73 - release: prepare git-cas v6.3.0 [merged]
#78 - release: prepare git-cas v6.4.0 [merged]
#91 - perf: reuse bounded Git object sessions [merged]
#93 - docs: record git-cas v6.5.2 publication [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!

[flyingrobots]

Publication follow-through is complete.

• The single runtime correction landed in Preserve coherent Git object sessions #95 and preserves coherent long-lived Git object sessions without changing public API, retention policy, object identity, or lifecycle bounds. [cite: src/infrastructure/adapters/GitPersistenceAdapter.js#93-175@7bdcbf1f9eccd16acd324c94d576e1ecd2e11d98]
• Unit and real-Git integration coverage prove the selective retirement matrix in SHA-1 and SHA-256 repositories. [cite: test/unit/infrastructure/adapters/GitPersistenceAdapter.sessionCoherence.test.js#6-45@7bdcbf1f9eccd16acd324c94d576e1ecd2e11d98] [cite: test/integration/git-object-session-coherence.test.js#24-60@7bdcbf1f9eccd16acd324c94d576e1ecd2e11d98]
• Release candidate release: git-cas v6.5.3 #96 merged as 00df6077f1f9c111b9d0d9b636b7d746df0d2aad; signed tag v6.5.3, release workflow 29696131557, npm provenance, package identity, and final GitHub Release are recorded immutably. [cite: docs/design/0053-git-object-session-coherence/witness/release-publication.md#7-61@d446d97e0436afccdfc0398091ba18888375d8b0]
• Current shipped posture and the downstream git-warp evidence obligation are explicit. [cite: STATUS.md#3-25@d446d97e0436afccdfc0398091ba18888375d8b0] [cite: docs/design/0053-git-object-session-coherence/witness/release-publication.md#63-79@d446d97e0436afccdfc0398091ba18888375d8b0]
• The release-state contract now rejects regression from published posture or drift in immutable artifact facts. [cite: test/unit/docs/release-state.test.js#112-134@d446d97e0436afccdfc0398091ba18888375d8b0] [cite: test/unit/docs/release-state.test.js#236-266@d446d97e0436afccdfc0398091ba18888375d8b0]

Independent verification: 14/14 release gates, 6,829 observed tests; hosted release and publication CI green; npm latest is 6.5.3; GitHub Release is final.