coldstart/

Cold-start probes

These deliberately tiny programs measure the cost of importing the core, voice, or ElevenLabs package graph. Go's linker retains only the selected graph, so each binary is a realistic lower bound for that deployment shape.

Build without the compiler cache influencing process measurements:

mkdir -p .bench
CGO_ENABLED=0 go build -trimpath -o .bench/core ./benchmarks/coldstart/core
CGO_ENABLED=0 go build -trimpath -o .bench/voice ./benchmarks/coldstart/voice
CGO_ENABLED=0 go build -trimpath -o .bench/elevenlabs ./benchmarks/coldstart/elevenlabs
CGO_ENABLED=0 go build -trimpath -o .bench/measure ./benchmarks/coldstart/measure

CGO_ENABLED=0 go build -trimpath -ldflags='-s -w' -o .bench/core.stripped ./benchmarks/coldstart/core
CGO_ENABLED=0 go build -trimpath -ldflags='-s -w' -o .bench/voice.stripped ./benchmarks/coldstart/voice
CGO_ENABLED=0 go build -trimpath -ldflags='-s -w' -o .bench/elevenlabs.stripped ./benchmarks/coldstart/elevenlabs

Measure fresh processes (the warmups affect only filesystem caches):

.bench/measure -binary "$PWD/.bench/core" -samples 200 -warmups 10
.bench/measure -binary "$PWD/.bench/voice" -samples 200 -warmups 10
.bench/measure -binary "$PWD/.bench/elevenlabs" -samples 200 -warmups 10

Pass -json to retain every sample in nanoseconds for a checked baseline:

.bench/measure -json -binary "$PWD/.bench/core" -samples 200 -warmups 10 > cold-core.json

For peak RSS, run one probe through the platform's time executable. On macOS use /usr/bin/time -l; on Linux use /usr/bin/time -v. Record the OS, hardware, power mode, Go version, build flags, and raw samples with published results. Do not compare results collected on different hosts. See the ../baselines directory for the checked release evidence.