Performance: contour PNG export misses the 10× Matplotlib target (8.50× in release audit)

Start with benchmarks/bench_pyplot_vs_matplotlib.py and spec/benchmarks/methodology.md at the audited commit; first correct the XY save call to export at 200 DPI, then confirm the contour gap on a controlled runner and record raw samples, dimensions, and tier decisions. Profile build, rasterization, and PNG encoding separately before choosing an optimization. Done means meeting the contour 10× standard target, or documenting reproducible evidence for revising it, while preserving output correctness and adding benchmark coverage.

The matched-output release benchmark found a contour export gap against the existing per-family 10× Matplotlib/Agg goal in the benchmark methodology.

Both outputs are nonblank 1800×840 PNGs. The metric is figure construction through a completed compressed PNG, excluding imports, data generation and warm-up. XY renders this row directly. Preserve contour topology, level placement and labeling behavior.

These are diagnostic results from three repetitions on a shared workstation, with one warm-up, recorded during the 2026-11-05 release audit. Confirm the gap on a controlled runner before selecting an optimization. This does not establish a regression from v0.0.8.

Reproduce

Audited source: 5e58b12c0a1b89a58c4555af370204ef1ed0515e plus local release-validation fixes. Environment: Linux x86_64, 12 logical CPUs, CPython 3.14.6, NumPy 2.5.1, Matplotlib 3.11.0, native XY core, Rust 1.96.0.

Harness prerequisite: the audited upstream bench_pyplot_vs_matplotlib.py omits export DPI in _xy_render. Set its save call to the following before measuring, matching the Matplotlib adapter:

fig.savefig(out, format="png", dpi=EXPORT_DPI)

Both adapters build at 100 DPI and export at 200 DPI. Without that correction, the harness fails its dimension check before producing a valid comparison. The correction and output-size regression test are prepared locally on farhan/release-validation, but have not landed upstream.

cargo build --release
uv sync --extra matplotlib --group dev
uv pip install 'matplotlib==3.11.0'
MPLBACKEND=Agg uv run --no-sync python benchmarks/bench_pyplot_vs_matplotlib.py \
  --profile smoke --reps 3 --warmups 1 \
  --json /tmp/pyplot-smoke.json

The harness runs all six families; inspect the contour row. For confirmation and the existing standard target gate:

MPLBACKEND=Agg uv run --no-sync python benchmarks/bench_pyplot_vs_matplotlib.py \
  --profile standard --reps 21 --warmups 3 \
  --target-speedup 10 --require-target \
  --json /tmp/pyplot-standard.json

--require-target gates every family, so a failure in another family does not by itself establish this issue's outcome.

Raw observations

[
  {
    "library": "xy.pyplot",
    "total_samples_ms": [
      10.909707,
      10.98602,
      11.471173
    ],
    "build_median_ms": 0.357385,
    "render_median_ms": 10.587838,
    "total_median_ms": 10.98602,
    "total_p95_ms": 11.471173,
    "png_bytes": 314653
  },
  {
    "library": "matplotlib",
    "total_samples_ms": [
      93.381723,
      92.888934,
      95.391998
    ],
    "build_median_ms": 5.579867,
    "render_median_ms": 87.794813,
    "total_median_ms": 93.38172300000001,
    "total_p95_ms": 95.391998,
    "png_bytes": 239043
  }
]

Acceptance

• Reproduce on a controlled runner and record environment, exact commit, raw samples, output dimensions and tier decisions.
• Profile build, rasterization and PNG encoding separately; optimize the measured bottleneck.
• Meet the existing standard-profile 10× total-time target for contour, or explicitly revise the specification with reproducible evidence explaining why the target is not met.
• Retain image correctness, labels and dimensions; do not achieve the target by silently reducing visual quality or dropping data.
• Record smoke, standard and huge profile results and output sizes, and add a meaningful regression benchmark.
• Update benchmark evidence and specifications with the resulting behavior.

The local audit report and complete raw comparison are prepared under spec/process/release-validation/2026-11-05/; they are currently uncommitted. The observations above are included here so this issue is reproducible without access to that worktree.