perf(pm): clone first waiter after extract

[elrrrrrrr]

Summary

This is a p3 scheduler experiment on top of #2989.

When a tarball download completes, the scheduler already knows which clone placements are waiting for that package. Today extract completes, reports cache readiness to the scheduler, then the scheduler queues clone/materialize work onto rayon. This PR moves the first ready clone waiter into the extract worker:

• download completion removes one primary clone waiter from download_waiters;
• extract worker writes the cache and immediately materializes that primary target;
• extract completion still records the cache and queues remaining waiters normally;
• if extract fails, the primary and remaining waiters receive the same error;
• if primary clone fails after cache succeeds, remaining waiters can still use the cache.

Hypothesis

• p3 may improve by removing one scheduler/rayon handoff per downloaded package with a ready waiter.
• p4 should be neutral because warm link does not use the extract path.
• ctx may move either way: fewer clone tasks, but longer extract workers.

Validation

• cargo fmt
• cargo test -p utoo-pm install_scheduler
• cargo clippy -p utoo-pm --all-targets -- -D warnings --no-deps

Bench Results

Baseline #2989 integrated head (fd38cfff):

• run 26094747554: p3 5.30s, ctx 52.1K/35.0K; p4 2.07s, ctx 7.5K/4.5K
• full-green run 26095851789: p3 5.66s, ctx 54.7K/37.6K; p4 2.16s, ctx 7.8K/4.6K

This PR #2995 (d1acae9, run 26100667673):

• p0 7.71s, ctx 77.8K/50.2K
• p1 2.52s, ctx 17.5K/22.0K
• p3 5.75s, ctx 59.7K/33.9K
• p4 2.09s, ctx 7.6K/4.3K

Conclusion: do not integrate for now. p4 is neutral as expected, but p3 does not beat the #2989+#2991 baseline and vCtx increases. The extra coupling between extract and first materialize is not justified by this result.

[gemini-code-assist]

Code Review

This pull request optimizes the package installation process by integrating a primary clone operation directly into the extraction phase. It introduces the ExtractOutcome struct and updates DownloadedPackage and OpDone::Extract to track a primary clone target. By selecting a waiter to be processed immediately after extraction, the scheduler reduces overhead. The changes also include a new helper method take_primary_extract_clone and updated unit tests to verify the concurrent extraction and cloning logic. I have no feedback to provide.

[github-actions]

📊 pm-bench-phases · 7d94cb1 · 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	9.27s	0.32s	10.41s	10.32s	699M	327.6K
utoo-next	8.24s	0.05s	10.56s	12.29s	963M	130.5K
utoo-npm	7.82s	0.22s	10.78s	12.31s	954M	116.6K
utoo	8.28s	1.14s	11.11s	11.26s	923M	139.1K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	14.9K	18.1K	1.19G	6M	1.84G	1.73G	1M
utoo-next	120.5K	82.0K	1.16G	5M	1.71G	1.70G	2M
utoo-npm	120.7K	82.5K	1.16G	5M	1.71G	1.70G	2M
utoo	79.0K	44.6K	1.16G	6M	1.71G	1.70G	2M

p1_resolve

PM	wall	±σ	user	sys	RSS	pgMinor
bun	2.03s	0.02s	4.02s	1.03s	511M	166.1K
utoo-next	3.04s	0.06s	5.51s	2.09s	616M	82.0K
utoo-npm	3.06s	0.02s	5.55s	2.13s	604M	81.6K
utoo	2.38s	0.01s	6.05s	1.67s	654M	120.5K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	9.3K	4.3K	203M	3M	107M	-	1M
utoo-next	72.5K	88.4K	200M	2M	7M	3M	2M
utoo-npm	72.7K	90.6K	200M	2M	7M	3M	2M
utoo	14.1K	18.2K	202M	3M	7M	3M	2M

p3_cold_install

PM	wall	±σ	user	sys	RSS	pgMinor
bun	6.74s	0.27s	6.29s	10.01s	601M	205.6K
utoo-next	6.85s	0.54s	5.06s	11.33s	545M	68.2K
utoo-npm	9.12s	2.66s	5.15s	11.52s	477M	58.8K
utoo	5.16s	0.18s	4.85s	9.69s	517M	55.7K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	4.1K	7.0K	1019M	3M	1.76G	1.76G	1M
utoo-next	109.9K	50.6K	988M	3M	1.70G	1.70G	2M
utoo-npm	132.8K	50.1K	988M	3M	1.70G	1.70G	2M
utoo	52.5K	34.3K	988M	2M	1.70G	1.70G	2M

p4_warm_link

PM	wall	±σ	user	sys	RSS	pgMinor
bun	3.31s	0.01s	0.20s	2.38s	134M	32.7K
utoo-next	2.38s	0.28s	0.48s	3.81s	79M	18.8K
utoo-npm	2.24s	0.08s	0.50s	3.85s	79M	18.2K
utoo	2.07s	0.06s	0.36s	3.31s	50M	11.2K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	262	22	5M	32K	1.91G	1.75G	1M
utoo-next	42.1K	19.1K	1K	6K	1.70G	1.70G	2M
utoo-npm	42.4K	18.9K	1K	5K	1.70G	1.70G	2M
utoo	7.4K	4.3K	2K	6K	1.71G	1.70G	2M

npmmirror.com: no output captured.