Record both redesign lanes, and remove three confusions - #548

Paul opened a from-scratch design exploration ALONGSIDE the extraction, not
instead of it. Without this, a cold session reads the stitch output as approved
direction - or reads ADR-0003 and does not know the exploration exists.

**Extraction governs; discovery has no authority.** ADR-0003 says plainly this
is "extraction and propagation, not a redesign" - the system comes from the
course page and /services/vibe-code-rescue/. Discovery produces candidate
designs and never a PR, and wins only via a superseding ADR carrying evidence,
because 2.1/2.2 already shipped the extracted system to the blog - 77% of the
site's Google traffic. Brand invariants (logo, ruby #cc342d, display face,
covers) stay fixed even in discovery: "from scratch" means layout, structure and
components, not a rebrand nobody asked for.

The discovery brief deliberately never mentions the course page - anchoring it
there reproduces extraction wearing a different hat and proves nothing. First
artifact recorded: Stitch project 4487418121858436329, which produced a system
it named "Artifact Editorial" and, unprompted, rejected the logo wall, the
testimonial carousel and the repeated "Book a call". Also disambiguated from the
older Claude Design project, which renders the EXTRACTED system.

Three confusions found and removed while writing this up:

1. "Held until the read is in" hid a date. The 2.4 gate needs a read due
   ~2026-09-17 - 28 days after the e1fa540 DEPLOY, which was never confirmed.
   It reads as imminent and is not, so anything depending on 2.4 is parked until
   mid-September and the first action on picking it up is confirming that live
   date. Writing the read up early does not unblock it; it pre-writes a
   conclusion the data cannot support.

2. Phase 2.5 has no spec. 2.1 and 2.2 each got one (20.03, 20.04); 2.5 is nine
   words in a table cell. Its first deliverable is 20.05, not code - and the
   binding constraint is canon-sensitive copy across 12 pages rather than the
   code-line cap. Only 2 of 12 service pages carry screenshot baselines, so ten
   money pages would otherwise ship visually blind.

3. The sequencing decision was unrecorded. Next unit is 2.3, not 2.5, decided by
   a three-lens panel that SPLIT: the ICP lens ranked service pages first, the
   risk lens won on facts the ICP lens did not have. Recorded with the reasoning
   because it deviates from "biggest ICP impact first" and a cold session would
   otherwise reverse it.

Gates: bin/hugo-build clean; v0.2 okf_validate .okf exits 0. Concept stamped
with measured UTC. Docs + bundle only.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>