[cjpais]

What

scripts/bench/memory.py, the peak-memory harness that the memory work keeps quoting numbers from.

Every memory branch is accepted on a number, and those numbers are only comparable if they come from one harness run one process at a time (two concurrent runs on a 16 GB host corrupt peaks). Until now that was a throwaway shell script in tmp/, which is why the numbers in the audit, in #220 and in the moonshine work were hand-copied between files.

# one cell
uv run scripts/bench/memory.py --models granite-4.0-1b-speech --samples jfk

# a sweep at the plan's reference length, warm (2 runs, keep the last)
uv run scripts/bench/memory.py --models granite-4.0-1b-speech,canary-qwen-2.5b \
    --samples dots-full --repeat 2

# duration scaling (cuts and caches the sample with ffmpeg)
uv run scripts/bench/memory.py --models sensevoice --samples dots-full --duration 1200

# A/B a branch switch; --env for library env vars, `--` for CLI flags
uv run scripts/bench/memory.py --models granite-4.0-1b-speech --samples jfk \
    --label copy --env TRANSCRIBE_NO_GATE_UP_ALIAS=1

Reuses run.py's model/quant resolution and machine detection, so variants are spelled the same way as the perf matrices, and writes reports/memory/<machine-slug>/<label>.csv.

Why these columns

After #195 peak RSS stopped being the number to trust: it counts mapped file pages the OS can reclaim and mostly misses Metal-owned buffers. cohere sits at 490 MB RSS with 4.4 GB of scheduler scratch; qwen3-asr at 1215 MB RSS with 6.0 GB. A harness that only reported RSS would have said both were fine. So per cell it records:

• peak RSS — the kernel's high-water mark, never sampled, so no peak is missed
• peak footprint — macOS phys_footprint; the best single "does this fit on this machine"
• peak RssAnon — Linux only, sampled; the share an OOM killer charges, which is the whole argument behind mmap model weights instead of copying them into anonymous memory #195
• scheduler scratch — summed over backends from the wave-0 instrumentation (backend: log per-graph scheduler scratch at debug level #217), with the CPU=x;MTL0=y split in notes; the portable number where RSS under-reports GPU scratch
• load time, wall, exit code, transcript hash (so two runs can be compared without keeping the text), and flags such as gate_up=aliased / repetition-guard

Anything the host cannot measure stays empty rather than being faked. --repeat 2 gives a warm number; --report min exists for noisy hosts and is never mixed with last in one table.

Scope

Script only: no library code, no build change, no ABI. It is already what produced the 16-variant re-baseline in reports/memory/ and the numbers in the moonshine PR.