refactor(stommel): use create_space_time_figure + add_magnitude_labels in build_desert_farm.py

[madsCodeBuddy]

Summary

Conservative refactor of docs/build_desert_farm.py to call package functions for the parts that were duplicating package code, while preserving the parts that exist for UX-specific reasons.

Replaced with package calls

Figure setup (~15 lines → 9 lines + overrides). figure(...) →
create_space_time_figure(space_on_x=False) plus per-attribute overrides for the figure-specific defaults (width 900, height 650, ranges 1e-3..1e13 / 1e-28..1e22, 11pt axis labels, 16pt bold title, #fafafa background, toolbar above, original 4-tool toolbar).

Reference grid (~28 lines → 1 line). The hand-rolled for time_val in TIME_MARKERS: ... for space_val in SPACE_MARKERS: ... loop with manually-positioned dashed Spans and Labels is replaced by add_magnitude_labels(p, font_size=LABEL_FONT_SIZE, space_on_x=False). Bonus: the package version has CustomJS callbacks that move the magnitude labels to follow the visible range edge on pan/zoom — the previous hand-rolled version anchored them at fixed positions.

Kept hand-rolled (with explanatory comment)

Manual ETL. Already calls package functions directly; no further consolidation possible without forfeiting Energy_type-based color assignment or losing the original Name column to create_name's ShortName fallback.

Patches + hover. add_processes produces patches but its data source has only x/y — no metadata for tooltips. Replacing the hand-rolled patches would lose the rich hover (Name, Scale, Energy, formatted ranges).

4-row energy-grouped legend. add_processes(category_col=, category_colors=) produces a 28-row legend (4 category headers + 24 per-process rows). The current 4-row hand-rolled version groups all glyphs of an energy type under one LegendItem so click-to-hide toggles the whole group at once. Materially better UX for a blog figure.

Verification

• Refactored script runs: Built docs/desert_farm_stommel.html (219,746 bytes).
• Spot-checked RuBisCO catalytic cycle — x_coords range matches Time (1e-1..1e0), y_coords range matches Space (1e-27..1e-24). Boyd convention preserved through ETL.
• Figure config: x-axis Time (s) at top, y-axis Space (m³), x_range = 1e-3..1e13, y_range = 1e-28..1e22. Toolbar tools = Pan/WheelZoom/BoxZoom/Reset (matches pre-refactor).

Stats

• 368 → 346 lines (-22, -6%).
• Diff: +37 / -59.

Dependency

Branched off PR #22 because it uses space_on_x=False through transform_process_response_sheet only via the manual ETL — actually this PR doesn't strictly depend on PR #22 (the script doesn't call transform_process_response_sheet at all). PR #22 should still merge first to fix the existing leverage-points HTML, but this refactor would also work on plain main. Will rebase if you'd prefer.

Possible follow-ups (not in this PR)

• Add a hover_tooltips parameter to add_processes so future refactors of this kind can also replace the hand-rolled patches.
• Add an add_compact_legend(category_col=) helper that produces the 4-row energy-grouped legend with click-to-hide.
• Move the energy-type coloring into add_processes via a new color_col argument.

[madsCodeBuddy]

Updated to actually use transform_process_response_sheet. You called out that my justification for keeping the manual ETL was thin, and you were right.

What changed:

1. etl.py: added n_points parameter to transform_process_response_sheet (default 1000 preserves existing behavior). Without this, ETL would have produced ~2 MB of HTML for the 24-process desert farm figure (10× the current 217 KB), since it hardcoded n_points=1000.
2. docs/build_desert_farm.py: load_processes now calls transform_process_response_sheet(... space_on_x=False, n_points=100) and drops the imports of process_magnitude_column, classify_process_geometry, create_ellipse_data, set_fill_alpha. Two pre-ETL hops remain: pre-computing Color from Energy_type (ETL doesn't know domain colors), and renaming Name → FullName so create_name's fallback doesn't clobber the original full name. Two post-ETL hops for label_x / label_y geometric centers.

Updated stats: 368 → 336 lines (-32, -9%). Same rendered output (219,746 B; matches previous refactor commit because the visible behavior is unchanged).

ETL bonus picked up for free: drops rows where Time_min > Time_max or Space_min > Space_max — data-quality guard the manual version skipped.

Wait remaining: ~30 hand-rolled lines for patches+hover and ~12 for the compact legend. Both still need package additions to be replaceable (noted in the PR description's follow-ups section).

[madsCodeBuddy]

Closing — branch is diverged (ahead=10, behind=10) and per branch hygiene rule the right move is recreate from current main rather than rebase via API. Replacement PR coming on a fresh branch.