Allocate child runtime resources lazily

[SandroMaglione]

Summary

• create the child scope and observable child registry only when child-management capabilities are first used
• serialize first allocation against process shutdown with one synchronized lifecycle state
• preserve immediate child lookup, race-free child change observation, dynamic spawning, and child cleanup
• avoid holding the lifecycle lock while closing child scopes so child finalizers can unregister safely
• run one benchmark harness against both revisions with five interleaved processes and process-level variability
• measure observed bursts, child lookup/delivery, parent-child lifecycle, and parent-with-child memory for Effect Machine and both XState baselines

Changeset

• Added or updated for a library or package-metadata change
• Not required because this PR does not change src/ or package.json

Validation

• pnpm check
• Relevant example checks, when examples changed (no examples changed)
• Reviewed the automated type-performance report, when the public TypeScript API or inference changed (no public type changes; local pnpm perf:types passed)
• Reviewed the automated runtime-performance report, when runtime behavior changed

pnpm check and pnpm perf:types pass. The 181 machine runtime tests also passed five additional consecutive stress runs before the benchmark expansion.

The final local A/B used the PR harness for both revisions and the CI ordering: five interleaved full processes per revision. Relative process variability is reported as median absolute deviation.

• ordinary planning: +0.1% (base 1.4% MAD, PR 2.0% MAD)
• unobserved burst: +1.1% (base 0.9% MAD, PR 0.9% MAD)
• observed burst: +1.7% (base 1.1% MAD, PR 0.5% MAD)
• child lookup and delivery: +0.4% (base 0.8% MAD, PR 0.3% MAD)
• no-child start/stop: +2.5% (base 1.2% MAD, PR 0.7% MAD)
• parent-with-child start/stop: -2.6% (base 0.9% MAD, PR 0.4% MAD)
• idle machine heap: 25.8 KiB to 25.1 KiB (-2.7%)
• idle parent-with-child heap: 74.5 KiB to 75.2 KiB (+1.0%)

An earlier independent five-process pass measured the same memory changes, child lookup/delivery at -1.0%, and parent-with-child lifecycle at -3.6%. The combined result identifies a narrow tradeoff: machines that never use child capabilities save 2.7% idle heap and start slightly faster; machines that create a child pay about 1% more combined heap and 2.6-3.6% during family allocation/teardown, while steady-state child lookup and delivery remain flat.

The GitHub runner reproduced the result across five processes: planning +1.3%, unobserved burst +2.2%, observed burst +0.9%, child lookup/delivery -0.0%, no-child start/stop +3.0%, parent-with-child start/stop -1.5%, idle heap -2.7%, and parent-with-child heap +1.0%. All required checks pass.

[github-actions]

Runtime performance

Median of 5 independent benchmark processes on AMD EPYC 7763 64-Core Processor with Node v24.18.0.

Pull request baseline

Scenario	Effect Machine	XState 5	XState 6 alpha
Plan counter transitions	27,377 transitions/s	132,804 transitions/s	123,281 transitions/s
Drain burst with terminal fence	15,655 increments/s	400,727 increments/s	111,428 increments/s
Drain burst with a change observer	14,898 increments/s	398,271 increments/s	111,663 increments/s
Lookup and send to one child	14,535 increments/s	395,440 increments/s	106,516 increments/s
Start and stop a machine	6,409 machines/s	184,502 machines/s	73,073 machines/s
Start and stop a parent with one child	2,599 families/s	67,627 families/s	26,596 families/s

Idle memory	Heap per machine
Effect Machine	25.0 KiB
XState 5	3.8 KiB
XState 6 alpha	6.1 KiB

Effect Machine change from base

Metric	Base	PR	Difference
Plan counter transitions	27,013 transitions/s	27,377 transitions/s	+1.3%
Drain burst with terminal fence	15,322 increments/s	15,655 increments/s	+2.2%
Drain burst with a change observer	14,761 increments/s	14,898 increments/s	+0.9%
Lookup and send to one child	14,541 increments/s	14,535 increments/s	-0.0%
Start and stop a machine	6,220 machines/s	6,409 machines/s	+3.0%
Start and stop a parent with one child	2,639 families/s	2,599 families/s	-1.5%
Idle heap per machine	25.7 KiB	25.0 KiB	-2.7%

Versions and interpretation

• Effect Machine: 0.3.0
• XState 5: 5.32.5
• XState 6 alpha: 6.0.0-alpha.27

Higher throughput is better; lower idle heap is better. Runtime measurements on shared GitHub-hosted hardware remain informational, so small differences should be confirmed across multiple workflow runs.

[github-actions]

Type performance

Measured with TypeScript 6.0.3 and skipLibCheck=true.

Scenario	Base	PR	Difference
Effect only	55	55	0 (0.0%)
Import effect-machine	55	55	0 (0.0%)
Machine.defineStates (3 states)	2,965	2,965	0 (0.0%)
Machine.make (3 states, 2 events)	8,731	8,731	0 (0.0%)
machine.handle (3 states, 2 transitions)	23,634	23,634	0 (0.0%)
machine.handle (depth 8)	113,023	113,023	0 (0.0%)
machine.handle (depth 12)	125,661	125,661	0 (0.0%)
machine.handle (depth 16)	141,003	141,003	0 (0.0%)
machine.handle (depth 24)	179,799	179,799	0 (0.0%)
machine.handle (wide depth 16)	210,137	210,137	0 (0.0%)
machine.handle (parallel/history/choice)	128,876	128,876	0 (0.0%)
machine.handle (4 successive calls)	129,498	129,498	0 (0.0%)
machine exact input/output/error/services	107,371	107,371	0 (0.0%)
execution adapter readiness	122,890	122,890	0 (0.0%)

Marginal instantiations are measured against the matching setup without that API call:

Scenario	Base	PR	Difference
Import effect-machine	0	0	0
Machine.defineStates (3 states)	2,910	2,910	0 (0.0%)
Machine.make (3 states, 2 events)	5,758	5,758	0 (0.0%)
machine.handle (3 states, 2 transitions)	14,903	14,903	0 (0.0%)
machine.handle (depth 8)	105,546	105,546	0 (0.0%)
machine.handle (depth 12)	116,984	116,984	0 (0.0%)
machine.handle (depth 16)	131,126	131,126	0 (0.0%)
machine.handle (depth 24)	167,522	167,522	0 (0.0%)
machine.handle (wide depth 16)	198,169	198,169	0 (0.0%)
machine.handle (parallel/history/choice)	110,514	110,514	0 (0.0%)
machine.handle (4 successive calls)	114,497	114,497	0 (0.0%)
machine exact input/output/error/services	95,819	95,819	0 (0.0%)
execution adapter readiness	96,968	96,968	0 (0.0%)

Check times (informational)

Scenario	Base	PR
Effect only	0.03s	0.03s
Import effect-machine	0.03s	0.03s
Machine.defineStates (3 states)	0.10s	0.10s
Machine.make (3 states, 2 events)	0.15s	0.15s
machine.handle (3 states, 2 transitions)	0.23s	0.22s
machine.handle (depth 8)	0.52s	0.50s
machine.handle (depth 12)	0.55s	0.56s
machine.handle (depth 16)	0.58s	0.57s
machine.handle (depth 24)	0.67s	0.66s
machine.handle (wide depth 16)	0.73s	0.74s
machine.handle (parallel/history/choice)	0.60s	0.60s
machine.handle (4 successive calls)	0.58s	0.57s
machine exact input/output/error/services	0.50s	0.51s
execution adapter readiness	0.53s	0.53s

Type instantiations are the comparison metric. Check time varies with runner load and is informational only.