adobe__skills/plugins/stardust/notes/new-use-cases/ROUND-1-REPORT.md

Round-1 report — three new use cases: validated approaches + proposed skills

2026-07-03. Companion to DESIGN-OPTIONS.md (research synthesis + pre-registered hypotheses). All three validation experiments ran to completion; verdicts below. Experiment folders: /Users/paolo/stardust/uc1-replica-test/, uc2-reskin-test/, uc3-genesis-test/ (each has a VALIDATION.md with full evidence). Round-1 constraint held: no existing skill was modified; every proposal ships as a new skill that consumes existing skills/scripts in place.

Verdict summary

Hypothesis	Verdict	Headline evidence
H1 clean re-authored replica reaches near-pixel fidelity ≤3 iterations	CONFIRMED	a typographic retail home page: pixel-diff 8.31% → 2.93% → 1.31%, height Δ 0px, content-diff "findings: none" (198/198 nodes, 59/59 CTAs)
H2 source-fidelity gate buildable from existing diff scripts + small pixel add-on	CONFIRMED	generic profile worked; 4 adaptations needed (see UC1 notes)
H3 direction step can be mechanical (current → target)	CONFIRMED	zero creative decisions; only measurement-policy calls
H4 faithful-content reskin in one archetype pass	CONFIRMED	a healthcare site × a payments-company donor: text byte-identical 2281/2281, 7/7 images, 47/47 slots, 13/13 metadata, 17/17 donor-token probe
H5 cross-origin mapping brief decidable (≥80% mapped)	CONFIRMED	91% mapped to named donor modules, 1 explicit new-module, 0 silent improvisation
H6 dom-equality adoptable as content gate	CONFIRMED	adapted with structure checks informational; shared normalization module capture↔gate
H7 variant fan-out + axis pinning → distinct non-slop variants	CONFIRMED	pairwise axis distance 5/5/5; pins byte-verified; 1 variant partially collapsed via anchor↔signature incoherence (→ new gate)
H8 heterogeneous inputs measurably steer direction	CONFIRMED	9/9 pool picks changed baseline→informed, all cited; photos falsified the text brief's vibe
H9 refero grounding without convergence	PARTIAL	screens grounded structure cleanly; styles search leaked into personality picks → firewall needed

Discarded in round 1: UC1-C snowflake overlay as UC1 mainline (byte-fidelity without re-implementation value — kept as escape hatch); UC1-B CSS-portation as default (kept as a scoped reserve, see below); UC3-B thin nebula wrapper (fails multi-input + pinning brief).

---

UC1 — PROMOTED: stardust:replica (same-design migration to AEM)

Flow (new skill orchestrating existing ones, no core changes):

1. extract --prep unchanged → inventory, page types, module candidates, screenshots, fonts, media.
2. Mechanical direction promotion (inside replica): stardust/current/{PRODUCT,DESIGN}.{md,json} promoted verbatim as target spec. Permitted deltas only from an inconsistency register (audit design findings and/or user-supplied list), brand_adjacent_refinements[]-shaped.
3. Recreation loop per archetype: clean semantic HTML/CSS authored from captured content + values lifted from the source site's own CSS + screenshot ground truth. Gated by the source-fidelity gate (below), ≤3 iterations per breakpoint pass.
4. migrate (sibling tier) → deploy (bias toward template-slotted decode) → rollout (block dedup already implements "same blocks across the whole site").

Source-fidelity gate (new scripts owned by replica; consumes diff scripts in place):

• content-diff.mjs + visual-diff.mjs --profile generic, live URL vs served prototype.
• Pixel probe: scroll-and-stitch capturer (Chromium fullPage:true renders lazy-decoded images as placeholders — stitching is mandatory) + pixelmatch; navigate by the per-500px band breakdown, not the overall %. Pass bar used: ≤~10% pixels, 0 structural 🔴, all visual flags justified.
• Hardening required (H2 evidence): domcontentloaded instead of networkidle (live sites never idle); real-Chrome UA (default UA gets a Cloudflare challenge page that the diff then silently measures as the source — false-measurement trap); symmetric --main scoping.

Design notes from the experiment:

• Two things computed-style capture missed and only the gate caught: rendered heading face living on an inner span (width-probe font fork caught it), and a hero scrim invisible to DOM/style probes (recovered by per-row luminance fitting). Keep both probes in the gate.
• "0 structural 🔴" required mirroring live DOM quirks (span-in-heading granularity, hidden carousel clones, sr-only labels) — the known #87 JOIN/SPLIT tension; replica needs a documented "granularity parity" policy rather than fighting it per page.
• Mobile is not free: 360px reference diff was 24% — the gate runs per breakpoint.
• "Capture-state" policy: CDN-403'd images and "Loading…" hydration states are replicated as captured, logged, and flagged for the delivery phase.
• CSS-portation reserve (from UC1-B): only for paint-level effects, JS-hydrated commerce components, and video/animated heroes — per-section, never page-level.
• Prior art to fold into the skill text: the airport-site SKILL-IMPROVEMENTS.md (fidelity-by-CSS- extraction, neutralize boilerplate block CSS, preserve-and-move, coverage guards) and the the healthcare site's archetype→sibling scale pattern.

UC2 — PROMOTED: stardust:reskin (existing content × defined new design)

Flow:

1. Donor ingestion: live URL → the existing --design-source capture shape; local static prototypes → same path served on localhost; Figma → adapter (Figma MCP: variables→palette, text styles→type, screenshots→vision refs; new provenance class) — Figma adapter is round-2 scope, the contract is defined now.
2. Content-model capture of the content site (new profile, the round-1 gap): slot-granular text, CTAs with absolute hrefs, ordered visible images, metadata, an explicit content-root scope declaration, and an executable normalization ledger (cookie chrome, carousel de-duplication) shared between capture and gate.
3. Mapping brief (per content slot): donor module id + rationale + status ∈ {mapped, new-module, chrome}; gate: ≥80% mapped, new-modules explicit, zero silent improvisation.
4. Programmatic render from content-model.json — never retype strings; byte fidelity holds by construction. Donor tokens + module patterns supply the surface.
5. Dual gates: adapted dom-equality (text bytes + ordered image set; structure informational)
   • slot coverage + metadata carry-over; donor-token computed-style probe. Then migrate/deploy/ rollout unchanged.

Guards from the experiment: content-root discovery (naive scoping silently dropped 30% of content), carousels break byte determinism (normalize at both ends), donors running multiple design systems concurrently (pin one donor reference page per module).

UC3 — PROMOTED: stardust:genesis (greenfield creative, Claude-Design-as-skill)

Flow:

1. Input dossier: any mix of text brief / site URLs / brand images / Figma / inspiration URLs / refero — each an origin with citations; conflicts recorded and resolved explicitly (photos can outvote the text brief — that happened and was correct).
2. Direction via nebula machinery as a read-only library (stardust-greenfield pools, anchor selection, register, tension, distinctiveness gate). Refero firewall: screens = structure evidence only; styles search = category/avoid evidence; personality cites non-web anchors + register.
3. Variant fan-out: variants.json with 6 pinnable slots (typography, density, substrate, accent, motion, edges — colour split into substrate/accent); pairwise axis-distance floor ≥2; anchor-promotion policy (variant anchors = direction's scored winner + runner-ups — gave distance 5 for free); named tension per variant.
4. Renders + contact sheet as first-class artifact: ≥2 scroll states per variant (hero-blindness finding: re-rolls can be invisible at rest) + axis table + a "what moves" line.
5. Pin + re-roll loop: constrained resample (register bias, signature anti-pairs re-run — V11×S14 collision was caught this way, tech budget); plus one qualitative anchor↔signature coherence gate (the instrument gap that let variant C partially collapse).
6. Optional back half: pinned winner → stardust prototype/migrate/deploy for multi-page delivery.

Open items (round 2): image generation hook; pool content debt surfaced as findings (edges pool has no playful posture; registers/gestures stubs forced substitutions); precedence decision where impeccable's "overused font" hook flags pool-curated picks (genesis must declare which instrument wins); nebula upstream is one unpushed commit ahead of GitHub — sync before depending on it.

---

Round-2 synergy candidates (deliberately NOT done in round 1)

• extract: content-model capture profile; --design-source-dir; Figma provenance class.
• direct: native preserve mode (replaces replica's mechanical promotion).
• diff: absorb the pixel probe + UA/wait hardening + granularity-parity policy (#87).
• nebula pools upstreamed/shared; refero firewall generalized to uplift/direct Mode B.
• Evals: none of the seven existing evals cover these flows — each new skill needs its own.