feat(html-artifact): saved-template clone mode + named gallery

[notque]

What

Adopts the clone-don't-regenerate pattern from OpenAI's curated-template skills — without importing their skills. Those 20 skills ship editable Office binaries (.pptx/.xlsx/.docx, 27 MB) and are host-locked to an OpenAI Office runtime; they self-abort elsewhere and are thin wrappers (20 copies of one skill). We kept the idea, rejected the files.

A frozen HTML template in templates/saved/ is cloned and slot-filled by a deterministic script. Layout is fixed; only declared content slots change.

Added

• scripts/fill-template.py — clone-and-fill engine. Fails loud (exit 1) on a missing required slot, an undeclared slot name (catches typos), or a leftover {{MARKER}}. Enforces content-vs-layout authority mechanically, not by prompt discipline.
• Gallery (one skill, many template files — not a skill per template): business-review, project-kickoff, system-design, all in the Birchline design system, each with a .slots.json manifest.
• SKILL.md — Phase 0 clone-first check + "Fidelity & Authority" section (clone-don't-regenerate; content controls content, template controls layout; fail loud).
• PHILOSOPHY.md — worked example under External Components Are Research Inputs, Not Imports (practice adopted, files rejected).
• 13 tests, incl. gallery-integrity guards (every declared slot has a marker; every marker is declared).

Blind A/B: clone vs regenerate

Same 3 content briefs, two arms: clone (frozen template + slot fill) vs regenerate (3 independent agents build layout from scratch). 6 outputs anonymized and scored by a blind grader (no arm labels, no key).

Dimension	Clone	Regen	Δ
Visual consistency	5.00	4.33	+0.67
Professional polish	5.00	4.00	+1.00
Slop-free	5.00	3.33	+1.67
Structural completeness	3.00	5.00	−2.00
Overall	4.50	4.17	+0.33

Read: Clone wins decisively on consistency, polish, and slop — its purpose. The blind grader independently clustered the 3 clone outputs as "one identical design system" and the 3 regen outputs as each "comparatively bespoke" (blue gradients, card shadows, differing headers) — i.e. it detected the drift clone mode removes, without knowing the arms.

The completeness gap is a test-harness artifact, not a design flaw: the clone arm was fed thin placeholder slot text; the regen agents wrote fuller prose. Content is the variable there, not layout. With real content, clone mode keeps its consistency edge and fills completely — see the fully-populated Issue #907 kickoff rendered from the same template.

Verification

• 211/211 html-artifact script tests pass (13 new + 198 existing)
• ruff check + ruff format --check clean on new files
• Generated artifacts pass validate-artifact.py (structure, self-containment, slop scan)

Not done (by design)

No Office-binary output. If editable .pptx/.xlsx/.docx is ever wanted, that's a separate python-pptx/openpyxl toolchain decision.

[github-actions]

Skill-eval coverage

Changed skills in this PR: 1 (4651345990f5..9abfb0c6e342).

Report only. CI cannot run evals (the runner needs the claude CLI). Run the ablation locally to get the base->head delta.

Uncovered

no eval coverage for changed skill(s): html-artifact
A gap to consider, not a failure.