Migrate visualization layer from Plotly to matplotlib

[OmerShubi]

Summary

Replaces the entire visualization layer (scanpath, heatmaps, comparison, animation, and the research charts) with matplotlib. The biggest practical win: static export (PNG/SVG/PDF) and the GIF/MP4 animation rasterizer are now fully in-process — no Kaleido, no headless Chrome, no chromium system package. The hard-won true-to-scale word labels are preserved and validated at the documented ~0.28 font-to-box ratio.

Net: 33 files, +2893/−3104 — the migration removed more code than it added (Plotly's frame-masking/slider/updatemenus machinery and the Kaleido warm-server logic are gone).

What changed

• New scanpath_studio/mpl_render.py — the rendering foundation: a fixed render DPI, screen-px↔point unit conversions, Plotly-compatible colour / colorscale / dash parsing (so constants.py and saved plot configs are untouched), and exact-pixel figure scaffolding.
• plots.py — every builder returns a matplotlib Figure; make_scanpath_animation returns a new ScanpathAnimation dataclass. The true-to-scale text math is unchanged; saccades are a single LineCollection (preserves the old one-trace perf bound).
• tabs.py render layer — spatial plots embed as a fixed-size inline SVG, uniformly CSS-scaled to the column (the true-to-scale trick, minus the Plotly.js CDN); the animation embeds the to_jshtml interactive player (play/pause/scrub); research charts use st.pyplot.
• animation_export.py — keeps the GIF/MP4 encode half verbatim; only the Kaleido frame rasterizer is swapped for matplotlib savefig.
• export.py / api.save_figure — savefig dispatch for PNG/SVG/PDF; HTML is now a self-contained SVG page; animations export as the to_jshtml player.
• Deps — matplotlib in, plotly + kaleido out; packages.txt (chromium) deleted. Docs / CHANGELOG / AGENTS / CLAUDE.md updated.
• Tests — the figure-coupled suites now assert matplotlib artists (new tests/mpl_helpers.py documents the artist-label contract).

Notable behaviour

• No browser anywhere. Export and animation rasterizing run in-process (MP4 uses the already-bundled imageio[ffmpeg]); animation export is much faster (no ~10s Chrome cold start).
• The on-screen animation player inlines every frame as base64 PNG, so it's heavier than the old vector Plotly animation; the player is capped at 250 frames (total runtime preserved) so long readings stay manageable. GIF/MP4 export keeps its own separate cap.
• Two real rendering bugs surfaced by the new tests were fixed: the animation trail's color array was permanently masked by an all-NaN init, and hollow markers used facecolors="none" (now an explicit transparent RGBA). Saccade-direction arrows now render as proper arrowheads (data-unit quiver width) instead of blobs.

Verification

• Full test suite green (427 passed); ruff format + ruff check clean.
• Verified live in the browser: static scanpath, per-word heatmap, the interactive animation player, and saccade arrows all render with zero console errors.

🤖 Generated with Claude Code

[OmerShubi]

Superseded by #83 — the branch was renamed claude/kind-gagarin-2c5f63 → matplotlib-migration, which closed this PR. Same commit; continue review in #83.