perf(pm): make resolver registry stateless

[elrrrrrrr]

Summary

• Remove registry cache hooks and memory cache layer.
• Make UnifiedRegistry stateless now that the mainloop owns cache/inflight state.
• Update cache ownership docs and resolver diagnostics.

Review Focus

• No remaining resolver-side OnceMap/MemoryCache ownership in registry.

[gemini-code-assist]

Code Review

This pull request refactors the manifest caching and resolution logic by removing the global MemoryCache and the OnceMap-based single-flight mechanism from UnifiedRegistry. Instead, the BFS resolver now owns the in-memory manifest maps for the duration of a resolution run, making the UnifiedRegistry service more stateless. A potential performance regression was identified regarding persistent storage I/O: the removal of the stored_version deduplication set means that multiple dependency specifications resolving to the same version will now trigger redundant writes to the ManifestStore.

[gemini-code-assist]

The removal of the stored_version deduplication set in UnifiedRegistry introduces a performance regression for persistent storage I/O.

In the previous implementation, this set ensured that multiple dependency specifications resolving to the same version (e.g., ^1.0.0 and ^1.1.0 both resolving to 1.2.3) would only trigger a single write to the ManifestStore. With its removal, every ManifestJob::Version or ManifestJob::ExtractVersion that completes will now issue a redundant, fire-and-forget write to the store for the same version manifest.

While the BFS loop's FetchQueues deduplicates jobs with the same FetchKey, different ranges have different keys, so they will still result in multiple provider jobs that resolve to the same exact version. Consider re-introducing a deduplication mechanism, perhaps within the ManifestState of the BFS loop if the goal is to keep the UnifiedRegistry service itself stateless.

[github-actions]

📊 pm-bench-phases · 018a43a · linux (ubuntu-latest)

Workflow run — ant-design

PMs: utoo (this branch) · utoo-npm (latest published) · bun (latest)

npmjs.org

p0_full_cold

PM	wall	±σ	user	sys	RSS	pgMinor
bun	9.30s	0.20s	10.53s	10.36s	762M	337.1K
utoo-next	8.56s	0.91s	10.61s	12.34s	998M	137.0K
utoo-npm	8.22s	0.31s	10.75s	12.29s	962M	117.9K
utoo	9.17s	1.00s	11.55s	12.43s	986M	147.8K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	14.9K	19.7K	1.22G	6M	1.94G	1.82G	1M
utoo-next	124.6K	86.1K	1.19G	5M	1.77G	1.77G	2M
utoo-npm	122.8K	83.1K	1.19G	5M	1.77G	1.77G	2M
utoo	125.2K	65.7K	1.19G	6M	1.77G	1.76G	2M

p1_resolve

PM	wall	±σ	user	sys	RSS	pgMinor
bun	1.92s	0.16s	4.04s	1.01s	512M	161.6K
utoo-next	2.80s	0.03s	5.26s	1.66s	619M	87.1K
utoo-npm	2.98s	0.05s	5.51s	2.04s	626M	80.3K
utoo	2.44s	0.07s	6.07s	1.67s	647M	124.2K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	8.0K	4.7K	205M	3M	109M	-	1M
utoo-next	47.3K	70.0K	202M	2M	7M	3M	2M
utoo-npm	71.9K	88.7K	202M	2M	7M	3M	2M
utoo	15.9K	20.6K	205M	3M	7M	3M	2M

p3_cold_install

PM	wall	±σ	user	sys	RSS	pgMinor
bun	6.85s	0.12s	6.44s	10.01s	654M	211.3K
utoo-next	7.00s	1.77s	5.00s	11.07s	466M	62.4K
utoo-npm	7.63s	2.45s	5.18s	11.28s	532M	61.0K
utoo	6.67s	1.62s	5.04s	10.94s	511M	62.2K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	5.0K	7.3K	1.02G	3M	1.82G	1.82G	1M
utoo-next	107.0K	49.3K	1018M	3M	1.76G	1.76G	2M
utoo-npm	118.1K	49.9K	1018M	3M	1.76G	1.76G	2M
utoo	100.8K	52.9K	1018M	3M	1.76G	1.76G	2M

p4_warm_link

PM	wall	±σ	user	sys	RSS	pgMinor
bun	3.36s	0.05s	0.20s	2.36s	136M	33.3K
utoo-next	2.33s	0.12s	0.50s	3.78s	81M	18.5K
utoo-npm	2.31s	0.04s	0.49s	3.74s	80M	18.3K
utoo	2.32s	0.04s	0.50s	3.82s	80M	18.6K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	472	24	5M	17K	1.98G	1.81G	1M
utoo-next	43.9K	19.2K	6K	6K	1.76G	1.76G	2M
utoo-npm	41.3K	18.9K	6K	22K	1.76G	1.76G	2M
utoo	45.0K	20.8K	7K	7K	1.77G	1.76G	2M

npmmirror.com: no output captured.