docs(cartography): trim both cartography references, lift family principle, restore safety nets

[anamanvil]

Summary

Trims and reorganises both cartography.md references (one per skill) to address two concrete behaviours we were seeing in agent-generated maps:

1. Sunset reflex — agents reaching for Sunset / SunsetDark on data that wasn't thermal/severity, because the docs recommended them by name in multiple tables.
2. Sequential / diverging palettes on string columns — agents picking a sequential ramp for unordered categories because the family-matching principle was buried under prescriptive tables.

The fix leans on established cartographic doctrine (Brewer / ColorBrewer / MacEachren — match palette family to measure character) rather than naming specific palettes by reflex.

Three commits on this branch

5cdcf85 — carto-render-inline-map/cartography.md trim (426 → 441 lines)

• Hoisted the cartographic principle into §0 (sequential ↔ magnitude, diverging ↔ signed, qualitative ↔ unordered).
• Stripped named palettes from recipes; recipes use placeholders pointing back at §6.
• Expanded §6 with the full CARTO palette registry per family + colorblind-safe subset.
• Added §3.1 "Opacity as a design lever" (three jobs, fit-by-role).
• Reframed drawing-order rule (§2) as coverage/occlusion, not size.

d94a69c — carto-create-builder-maps/cartography.md trim (1049 → 678 lines)

• Lifted the same principle into §4.
• Dropped troubleshooting prose (§4.5a NULL-ratio post-mortem, rooftop-PV failure narrative, ENERGY STAR worked example, why agents over-use quantile meta).
• Demoted Sunset / SunsetDark from the §4.4 dark-basemap default list and §4.2a thermal branches; kept in the §4.1 registry.
• Reframed §1.8 drawing-order rule as coverage/occlusion with a per-row "Why this position" column.
• Collapsed anti-patterns §7.10–7.13 to index pointers; worked recipes (§8) reduced to three archetypes.

Net: ~-35% line, ~-50% word reduction. Lost no Tier-1 capability content.

f874f2c — restore safety nets

Honest review after the first two commits found regression risk for the median-reasoning agent (placeholders without a working example, opacity range without an anchor, string-column rule too nuanced). This commit restores three guardrails without reintroducing the original Sunset reflex:

1. Default-per-family table in §4.2 / §6 — Teal (sequential), Temps/Geyser (diverging), Bold/Safe (qualitative), Prism/Pastel (7–12 cats). Framed explicitly as fallback when reasoning gives no preference, not a reflex.
2. Opacity anchored at 0.7 — the no-context default. The 0.4–0.8 design-lever range remains for agents that have a reason.
3. String-column rule simplified — "string → qualitative palette, period". The rare ordered-strings exception lives in §3.5 as an escape hatch, not surfaced in the anti-pattern.
4. Checklists in both files reference the §4.2 / §6 defaults for the final-gate pass.

Companion PR

CartoDB/cloud-native#24630 updates the view-map MCP prompt to align with the new skill structure — concrete palette examples (Teal / BluYl / Bold) with inline family-matching rationale, names varied across recipes to avoid a new reflex.

Validation

• make validate clean (23 skills, 236 snippets, 31 untagged skipped) after each commit.
• No SKILL.md changes; no frontmatter changes; no plugin manifest changes — pure prose refactor of two reference files.

Review focus

• Is the default-per-family fallback table (§4.2 / §6) framed clearly enough that strong agents don't reach for it as a reflex, while weak agents do reach for it when reasoning runs out?
• Did the recipe trim in §8 (create-builder-maps) and §10 (render-inline-map) cut anything load-bearing?
• Drawing-order reframing in §1.8 / §2 from "size" to "coverage / occlusion" — does it survive contact with the layer matrix?