feat(docs): architecture-diagram-3d, interactive 3D code-architecture diagrams

[NeuralEmpowerment]

What

A new procedural skill, skills/docs/architecture-diagram-3d, that teaches how to build an interactive, orbitable 3D diagram of a codebase as one HTML page.

Live example it was distilled from (five views of a knowledge vertical): https://claude.ai/artifact/HnENkkWr7qKV23mYB3tnVU

Files

• skills/docs/architecture-diagram-3d/SKILL.md: procedural chassis (when to use, input, workflow, output, outcomes, dated recommendations, references).
• assets/template.html: the full engine (three.js r128 UMD from cdnjs, hand-written orbit with drag / shift-drag pan / wheel / pinch, auto-orbit toggle, reset view, per-frame projected HTML labels scaled by distance, raycast picking, ResizeObserver, prefers-reduced-motion, phone layout, THREE-load fallback) plus one small worked example view. Adding a view is data only.
• references/visual-grammar.md: floors are layers, walls are boundaries, edge colour is the kind of flow, dashed is reference or planned, pulse shape tells bytes from facts, status strip, glass columns for slices, the room view, and the five-view set.
• references/data-schema.md: every field of a DIAGRAMS entry, units, camera, sizing guidance.
• references/build-checklist.md: ground it in the code, check every path with git cat-file -e origin/main:<path>, keep private content out, node --check the extracted script, publish under the host's artifact rules.
• README.md: listed in the Docs collection.
• skills/docs/system-infographic/SKILL.md: one added non-trigger so the static poster and the interactive 3D explorable route apart.

Checks

• node --check on the extracted template script: pass.
• scripts/verify-install.sh: install proof passed, docs module now 4 skills in both harnesses.
• npx skills add ./skills/docs --list discovers the skill.
• No em dashes, no home paths or hostnames (docs/PII-HYGIENE.md grep).

Do not merge yet; opened for review.

[NeuralEmpowerment]

Codex review (codex exec, read-only) + fixes in c2c16db

Codex found:

• BLOCKER: the console path export failed because DIAGRAMS is scoped inside the IIFE. The grep fallback truncated .tsx to .ts and skipped other extensions.
• BLOCKER: the workflow read the working checkout but checked only path existence on main, so unmerged behaviour could be drawn as built.
• BLOCKER: in the worked example, "bytes cross only here" never crossed the wall, while handler to blobs did cross it carrying bytes.
• NITs: planned-reader edges were solid; block labels weren't keyboard-reachable and there was no keyboard orbit; wheel and keyboard zoom left auto-orbit on; the legend hardcoded four edge kinds; "a syntax error shows the fallback" was wrong.

Fixed in c2c16db (all of the above):

• Step 1 reads code at origin/main's SHA (worktree or git show) and records it in the eyebrow. Paths are checked at that SHA.
• The template sets window.DIAGRAMS. The checklist exports resolved where entries from the console and checks them in a loop that takes any extension and directories.
• Example text now says payload crosses twice, both through the blob store (the PUT, and the reader resolving a reference). Planned edges are dashed.
• Labels are role=button with tabindex, opened with Enter/Space. Q/E orbit. Zoom turns auto-orbit off. The legend is built from EDGE_TEXT and STATUS_TEXT. Docs are updated to match.

Left: none.

Checks: node --check on the extracted script passes and scripts/verify-install.sh passes (every module, Claude Code + Codex). Not re-walked in a browser after the engine edits; worth one visual pass before merge. Not merged; the operator merges.