Fix CI: bounded hexbin payloads + pyplot per-build overhead

[Alek99]

Fixes the two tests that have kept CI red on main since #37 (and made every open PR — e.g. #42 — fail checks it didn't cause).

Hexbin payload (CodSpeed / test_first_payload_hexbin_core_2d)

_emit_hexbin expanded every cell into six fan triangles across seven f32 columns — 168 B/cell, 3.17 MB at gridsize=128 vs the 1.6 MB raw input (#45 had loosened the benchmark bound to accept this). Every hexagon shares the same geometry, so the payload now ships centers + one color value per cell (12 B/cell, 227 KB, 14× smaller) with hex_dx/hex_dy in the trace style, and each renderer expands locally:

• WebGL client: new _buildHexbinMark expands to instanced triangles in the centers' encoded space (deltas scale by meta.scale), reuses the mesh draw path unchanged.
• SVG: one <polygon> hexagon per cell (18,890 elements instead of 113k triangles).
• Raster: vectorized fan expansion into the existing cmd.triangles op.

Verified pixel-identical WebGL readback (headless Chromium, same lit-pixel hash old vs new), plus SVG/PNG render checks. The benchmark bound is restored to the grid-bounded centers contract and re-asserts payload < raw input.

Pyplot perf guardrail (test_pyplot_build_tracks_declarative[10000-60-1.6])

The shim's fixed per-build cost had drifted to ~2× the declarative baseline (locally measured). The failures were exactly the "per-build revalidation / O(n) copy" class the guardrail exists to catch:

• styles.compile_axis_style / compile_mark_style / _validate.style_mapping now memoize per exact input dict — the same handful of rc-derived chrome dicts were re-run through the native CSS grammar on every chart build.
• The rc chrome snapshot (_load_rc_chrome) is cached per (RcParams.version, dpi); RcParams now versions all mutations. Per-axes mutable members (_theme_tokens, _hidden_spines) are copied from the snapshot.
• plot() gap/path detection: replaced an astype copy + np.diff + compare with the native is_sorted kernel, and clean (all-finite) data skips the isfinite mask via a sum probe — a non-finite value anywhere poisons the sum, so the exact scan only runs when there might be gaps. Gap/parametric routing semantics verified unchanged (NaN, inf, unsorted-x, sorted-clean).

Local best-of-N ratios at 10k: 2.03× → ~1.6×; at 100k: 1.26× (ceiling 1.5). Guardrail passes repeatedly locally.

Validation

• uv run pytest: 1770 passed, 1 skipped
• cargo test, ruff check, ruff format --check, ty check: clean
• scripts/render_smoke_nonumpy.py: OK
• hexbin WebGL pixel probe: identical lit-pixel count before/after

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will improve performance by 70.32%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 1 improved benchmark
✅ 79 untouched benchmarks
⏩ 1 skipped benchmark¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_first_payload_hexbin_core_2d	34.2 ms	20.1 ms	+70.32%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing fix/ci-hexbin-payload-and-pyplot-overhead (6f57cb2) with main (c36e908)}

Footnotes

1. 1 benchmark was skipped, so the baseline result was used instead. If it was deleted from the codebase, click here and archive it to remove it from the performance reports. ↩