Adopt persistent Git object sessions and close them safely

docs/topics/git-perf.md

@@ -35,8 +35,8 @@ only as a process-startup control; the retained harness invokes it once after
 timed scenarios to record the exact Git build in result metadata. No production
 recommendation invokes it.
 
-The current git-cas adapter starts a fresh process for each blob write, tree
-write, blob stream, tree lookup, and uncached object-info lookup:
+Before git-cas v6.5.2, the adapter started a fresh process for each blob write,
+tree write, blob stream, tree lookup, and uncached object-info lookup:
 
 - `hash-object -w --stdin` is invoked per blob. [cite: `git-cas/src/infrastructure/adapters/GitPersistenceAdapter.js#57-65@49b7d5cb9d589d73fa17d393e48d40bd6f139e57`]
 - `mktree` is invoked per tree. [cite: `git-cas/src/infrastructure/adapters/GitPersistenceAdapter.js#68-79@49b7d5cb9d589d73fa17d393e48d40bd6f139e57`]
@@ -49,6 +49,13 @@ On the measured Apple M1 Pro, one Git process had an approximately 18-20 ms
 startup floor. The problem is therefore both process overhead and our use of
 the process boundary: paying that floor once per object is self-inflicted.
 
+git-cas v6.5.2 shipped bounded persistent object sessions, and v6.5.3 corrected
+their visibility/retirement policy: `cat-file` survives successful immutable
+writes, `mktree` survives loose writes, and `mktree` is retired after a bounded
+`fast-import` pack write. The original process-per-object measurements below
+remain the diagnosis and baseline, not a description of the current adapter.
+[cite: `git-cas/src/infrastructure/adapters/GitPersistenceAdapter.js#93-175@7bdcbf1f9eccd16acd324c94d576e1ecd2e11d98`]
+
 The 128-operation profile makes the distinction concrete:
 
 | Packed workload               | One process per operation | Persistent Git | NodeGit |
@@ -64,6 +71,41 @@ Persistent Git matches or beats libgit2 for object metadata and packed blob
 reads. Parsing one bounded tree page and caching its immutable entry index beats
 re-entering either Git implementation for every path lookup.
 
+### Published v6.5.3 consumer checkpoint
+
+git-warp consumed the published npm artifact `@git-stunts/git-cas@6.5.3` with
+`@git-stunts/plumbing@3.2.0` and repeated the temporary 128-node cold retained
+materialization diagnostic three times. The injected plumbing counter began
+after repository initialization, so it observed 398 workload Git children per
+run; adding the excluded `git init` and two `git config` setup commands yields
+the same 401 total recorded for the release candidate.
+
+Every run produced the same major command counts:
+
+| Command        | Processes |
+| -------------- | --------: |
+| `hash-object`  |       140 |
+| `rev-parse`    |        87 |
+| `symbolic-ref` |        43 |
+| `update-ref`   |        43 |
+| `commit-tree`  |        39 |
+| `rev-list`     |        37 |
+| `cat-file`     |         4 |
+| `mktree`       |         1 |
+
+The three local cold-materialization timings were 7.02 s, 7.18 s, and 6.67 s.
+They are diagnostics only: this rerun used Node 26, did not capture process-tree
+CPU/RSS, and intentionally did not reproduce the old timing harness. The
+structural process counts are the comparable result. Focused real-Git retained
+materialization, materialization-store, and trie integration tests passed 11/11
+against the registry artifact.
+
+This checkpoint proves that npm v6.5.3 matches the audited session topology. It
+does not satisfy the versioned cold/warm/incremental benchmark contract in
+[#759](https://github.com/git-stunts/git-warp/issues/759): that issue still
+requires a committed corpus, process-tree CPU/RSS, repeated base/head CI runs,
+and blocking regression thresholds.
+
 ## Bounded page reads
 
 The resource workload used random 4 KiB objects so compression could not hide

package.json

@@ -117,8 +117,8 @@
   },
   "dependencies": {
     "@git-stunts/alfred": "^0.10.4",
-    "@git-stunts/git-cas": "^6.5.0",
-    "@git-stunts/plumbing": "^3.1.0",
+    "@git-stunts/git-cas": "^6.5.3",
+    "@git-stunts/plumbing": "^3.2.0",
     "@git-stunts/trailer-codec": "^2.1.1",
     "@noble/hashes": "^2.2.0",
     "boxen": "^7.1.1",

test/unit/scripts/v18-v17-public-read-legacy-reading-builder.test.ts

@@ -46,10 +46,10 @@ describe('v18 v17 public-read legacy reading builder', () => {
       'alice',
       'alice',
     ]);
-    // git-cas package versions are stamped into asset manifests, so dependency bumps
-    // intentionally advance this migration-reading golden handle.
+    // git-cas stamps its package version into the manifest formatVersion, so dependency
+    // bumps intentionally advance this migration-reading golden handle.
     expect(reading.facts.find((fact) => fact.factKey === 'node:alpha:_content')?.value)
-      .toBe('git-cas:1:asset:manifest-tree:cbor:sha1:8851a151d100f2faad3a93892dfb83fa6f1345f7');
+      .toBe('git-cas:1:asset:manifest-tree:cbor:sha1:88b6deeb1a9e0e364f80d244731170f29fbe09a4');
   });
 
   it('fails closed when a restored v17 writer ref drifts after restore', async () => {