perf(pm): cap install worker pressure

[elrrrrrrr]

Summary

Experiment for p3 cold install scheduling pressure.

• cap install clone workers at 4 instead of scaling to 6/10 on larger runners
• cap extract concurrency at 4 instead of 8
• keep the scheduler architecture unchanged; only tune worker pressure around clone/extract materialization

Verification

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

GHA Benchmark

Conclusion: negative on GHA; do not integrate as-is.

Compared with #2989 latest Linux/npmjs run (26102983933), this worker-cap experiment (26105816781) regresses p0 and p3:

phase	#2989 utoo	#2996 utoo	note
p0 full cold	7.34s, 65.1K/45.7K ctx	9.85s, 82.6K/52.8K ctx	regression
p1 resolve	2.39s, 14.9K/19.3K ctx	2.77s, 15.5K/19.8K ctx	run/network noise, no benefit
p3 cold install	5.30s, 52.4K/33.8K ctx	5.68s, 62.3K/40.6K ctx	regression
p4 warm link	2.20s, 7.9K/4.6K ctx	2.17s, 7.9K/4.6K ctx	neutral

The cap can improve one high-ctx internal environment, but GHA does not support making it the default.

[gemini-code-assist]

Code Review

This pull request reduces the concurrency and worker limits for installation tasks, capping them at a maximum of 4. The changes affect the clone_worker_limit and extract_concurrency_limit functions, and the unit tests have been updated accordingly. A review comment suggests simplifying the clone_worker_limit implementation, as the current logic is unnecessarily complex and mathematically equivalent to a simple clamp operation.

[gemini-code-assist]

The logic in clone_worker_limit has become unnecessarily complex with the nested clamping. Since the goal is to cap the worker limit at 4 while ensuring it doesn't exceed the total clone_limit (and stays at least 1), the entire expression is mathematically equivalent to clone_limit.clamp(1, 4).

For all $x \ge 1$, the formula $\lfloor x/2 \rfloor + 2$ is always greater than or equal to $\min(x, 4)$, meaning the previous scaling logic is now completely shadowed by the new cap. Simplifying this improves readability and maintainability.

Suggested change

	clone_limit
	.saturating_div(2)
	.saturating_add(2)
	.clamp(1, clone_limit.max(1))
	.clamp(1, clone_limit.clamp(1, 4))
	clone_limit.clamp(1, 4)

[elrrrrrrr]

Closing this draft as a negative experiment: the GHA AB regressed p0 and p3 versus #2989, so the worker cap should not be integrated as-is.

[github-actions]

📊 pm-bench-phases · e6f9edc · 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	15.43s	2.74s	9.92s	9.81s	571M	323.4K
utoo-next	10.44s	0.51s	10.70s	13.08s	984M	119.6K
utoo-npm	13.02s	2.75s	10.95s	13.37s	976M	121.0K
utoo	0.00s	0.00s	7.28s	6.23s	795M	117.0K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	34.0K	12.0K	1.19G	7M	1.86G	1.75G	1M
utoo-next	169.1K	114.6K	1.16G	6M	1.71G	1.70G	2M
utoo-npm	184.2K	128.3K	1.16G	6M	1.71G	1.70G	2M
utoo	37.3K	17.5K	688M	5M	958M	952M	-

p1_resolve

PM	wall	±σ	user	sys	RSS	pgMinor
bun	10.26s	3.29s	3.70s	1.14s	465M	217.5K
utoo-next	6.62s	1.91s	5.72s	2.28s	610M	88.0K
utoo-npm	8.01s	2.24s	5.80s	2.26s	597M	81.4K
utoo	8.49s	1.95s	6.11s	1.73s	632M	120.7K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	19.8K	1.6K	202M	3M	107M	-	1M
utoo-next	98.0K	79.6K	200M	3M	7M	3M	2M
utoo-npm	100.7K	79.3K	200M	3M	7M	3M	2M
utoo	21.6K	18.6K	201M	3M	7M	3M	2M

p3_cold_install

PM	wall	±σ	user	sys	RSS	pgMinor
bun	9.66s	4.94s	6.41s	10.11s	425M	184.1K
utoo-next	13.40s	2.25s	5.15s	11.38s	424M	61.3K
utoo-npm	11.41s	1.54s	5.25s	11.35s	442M	60.4K
utoo	10.02s	1.61s	5.14s	9.54s	462M	52.3K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	10.4K	9.8K	1014M	5M	1.76G	1.76G	1M
utoo-next	140.0K	52.7K	990M	4M	1.70G	1.70G	2M
utoo-npm	135.0K	62.3K	990M	4M	1.70G	1.70G	2M
utoo	79.2K	44.5K	990M	4M	1.70G	1.70G	2M

p4_warm_link

PM	wall	±σ	user	sys	RSS	pgMinor
bun	3.36s	0.07s	0.20s	2.34s	135M	33.4K
utoo-next	2.53s	0.15s	0.52s	3.77s	79M	18.1K
utoo-npm	2.34s	0.05s	0.48s	3.77s	78M	17.9K
utoo	2.17s	0.05s	0.34s	3.32s	50M	11.1K

PM	vCtx	iCtx	netRX	netTX	cache	node_mod	lock
bun	547	24	5M	21K	1.91G	1.73G	1M
utoo-next	41.8K	18.7K	7K	10K	1.70G	1.70G	2M
utoo-npm	40.4K	17.3K	6K	5K	1.70G	1.70G	2M
utoo	8.1K	4.5K	7K	9K	1.71G	1.70G	2M

npmmirror.com: no output captured.