Improve pyplot legend and dash parity

[FarhanAliRaza]

No description provided.

[Alek99]

Reviewed the full diff (Python shim, payload/exporters, JS client) and dug into the CI failures.

CI status — not this PR's fault

Both failing checks (Test test_pyplot_build_tracks_declarative[10000-60-1.6] and CodSpeed test_first_payload_hexbin_core_2d) are pre-existing regressions from #37 — they fail identically on main (5a58834, eb0c3f7, and still on c36e908). Fixes are in #46. I verified locally that this branch merged with #46 passes all 558 pyplot/plot-family tests plus both previously-failing gates. Once #46 lands, update the branch and CI should go green. One thing to watch: python/xy/static/*.js are committed build artifacts and git will happily automerge them — after rebasing, re-run node js/build.mjs (for the current merge I verified the automerge result matches a fresh rebuild exactly).

What I verified works

• _mpl_dash scaling is right: -- at linewidth=2 ships [10.28, 4.44] = (3.7, 1.6) × lw × dpi/72, matching Matplotlib's scale_dashes + point sizing; line width point-scaling is consistent between the trace and the frozen legend item (no double scaling).
• loc parsing (independent h/v anchors with "center" fallback) is consistent across JS, SVG, and raster, and the single-transform fix in _legendBox is correct — the old code's two transform: declarations would indeed clobber each other.
• _best_legend_loc stays bounded (512 samples/entry) and the ±0.02 tie band correctly preserves Matplotlib's candidate preference order.
• Second-legend payload shape (extra_legends with explicit items) renders in all three renderers and survives primary-legend suppression; good test coverage for both.

Minor findings

1. Axes.add_artist (_axes.py): artist._parent = host is reassigned after the Legend already froze _options and _items against the constructing axes in __init__ — the assignment has no effect. If a Legend(ax1, ...) is added to ax2 with a different dpi/rcParams state, the swatch scaling silently comes from ax1. Either drop the dead assignment or recompute options on attach.
2. JS _legendBox: only the primary legend registers as this._legend, so extra legend boxes never get their max-height refreshed by _resize — a window resize can leave a tall second legend overflowing the plot. Consider tracking all boxes.
3. JS line swatch: it.style?.width || 1.5 treats an explicit lw=0 as unset and draws a 1.5px handle; ?? 1.5 (plus letting 0 render nothing) is closer to mpl.
4. Legend.__init__: zip(handles, labels, strict=False) silently truncates a handles/labels length mismatch — Matplotlib warns. A warnings.warn here would match shim policy of not deciding silently. Handles without an _entry (e.g. Text) are also silently dropped.
5. Nit: in _best_legend_loc the final return "upper right" is unreachable — best is the min of scores, so the first candidate within the band always returns.

None of these block merging; 1 and 2 are worth a follow-up. Nice work on keeping the three renderers' legend geometry in lockstep — the shared item shape makes the parity easy to trust.

[codspeed-hq]

Merging this PR will not alter performance

✅ 80 untouched benchmarks
⏩ 1 skipped benchmark¹

---

_{Comparing mpl-shim (d52daf5) with main (b8aaa71)}

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. ↩