luxbench — one bench harness, one way

luxbench is the single driver for every benchmark in the luxfi stack.

Before this, /tmp had 5+ ad-hoc bench dirs each with its own orchestration, output format, and metric names. This is the decomplected replacement.

Usage

# Build:
GOWORK=off go build -o /tmp/luxbench ./cmd/luxbench

# List modes:
/tmp/luxbench -list

# Run a mode:
/tmp/luxbench --mode=zap-vs-codec --benchtime=2s --count=3 --output=/tmp/result.json

Common flags (every mode sees these via bench.Config):

Flag	Default	Purpose
--mode	(req)	Mode name; see -list.
--benchtime	1s	go test -benchtime value (or mode equivalent)
--count	1	Runs per subtest (for median / quantile)
--output	(req)	Output JSON path (always written)
--verbose	false	Stream underlying tool output

Result file

Every mode writes the same envelope:

{
  "version": 1,
  "mode": "zap-vs-codec",
  "records": [
    {
      "mode": "zap-vs-codec",
      "subtest": "BenchmarkRejectBadMagic_ZAP",
      "architecture": "spark",
      "timestamp": "2026-06-04T04:37:53Z",
      "go_version": "go1.26.3",
      "iterations": 1000000000,
      "median": 2,
      "p50": 2, "p95": 2, "p99": 2,
      "bytes_per_op": 0,
      "allocs_per_op": 0,
      "ops_per_sec": 5e8,
      "extra": { "MB/s": 4087460.62 }
    }
  ]
}

Schema: result/schema.go (ResultRecord).

Modes

Mode	Status	Source	What it measures
zap-vs-codec	LIVE	modes/zap_vs_codec/	8 head-to-head claims: linearcodec vs zap zero-copy
gpu	scaffolded	modes/gpu/	luxfi/accel + lux-private/gpu-kernels (TODO: port)
multi-validator	scaffolded	modes/multi_validator/	10v / 100v / 500v E2E (TODO: port)
parity	scaffolded	modes/parity/	crypto-primitive parity audits (corona/pulsar) (TODO)
matrix	scaffolded	modes/matrix/	7-profile sig-pack matrix (TODO)

The canary zap-vs-codec is fully migrated from /tmp/zap-vs-codec/. Other source dirs in /tmp/ are reproducible outputs and stay there until their migration sessions land.

Adding a new mode

1. Create modes/<name>/ (snake_case dir, snake_case Go package).
2. Implement bench.Mode (see mode.go):
   • Name() — kebab-case (--mode=<name>)
   • Description() — one line for -help
   • RegisterFlags(*flag.FlagSet) — mode-specific flags
   • Run(ctx, bench.Config) ([]result.ResultRecord, error)
3. Register in cmd/luxbench/main.go (alphabetical).
4. Document mode-specific flags in this README's table.

Two integration approaches for wrapping benchmarks:

• Wrap go test -bench (recommended when the bench is already a Go test, like zap_vs_codec). The mode shells out to go test, parses the standard bench output, collapses repeated runs into one ResultRecord per subtest with median / p50 / p95 / p99. The parser is reusable — copy modes/zap_vs_codec/mode.go's parseBenchOutput.
• Native sampling (for benches that aren't *_test.go, e.g. GPU kernels driven by a custom Go harness). The mode times its own iterations and constructs ResultRecord values directly.

Rule

Every new benchmark MUST go through this harness. No new ad-hoc /tmp/<thing>-bench/ directories. If the harness is missing something your bench needs, extend the harness — don't bypass it.

The /tmp/ output dirs are still fine — they're where you store result JSON. They are no longer where you write code.

Analysis (TODO)

Future: cmd/luxbench-analyze/ to consume result.File JSONs and emit comparison tables / regression flags. Stub goes in cmd/ when the first analysis task arrives.

Migration plan (other 4 modes)

Source bench → target mode → effort estimate (real measure: hours of porting + verification, not lines of code).

Source	Target mode	Effort	Notes
/tmp/spark-gpu-bench/	modes/gpu/	M	CGo bench with CUDA kernels. Already has its own go.mod. Port runner + parse MEDIANS.txt.
/tmp/multi-validator/	modes/multi_validator/	L	Network harness — already emits JSON close to ResultRecord shape. Map field names + collapse the 3 profiles into Extra.
~/work/lux/threshold/protocols/parity/	modes/parity/	S	Single Go test file; thin wrapper around go test -run TestParity.
/tmp/zap-vs-codec/ (matrix part)	modes/matrix/	M	The 7-profile sig-pack matrix is partially in zap-vs-codec but conceptually a different mode; split it cleanly.
(no source yet)	modes/gpu_vtbl/	S	GPU dispatch wiring bench — light port once luxfi/accel vtbl lands.

Each mode lands as one PR. The decomplection is done in steps; the harness contract is now frozen.