wand-decks-kit 0.8.0

package/PROMPT.md

@@ -0,0 +1,461 @@
+# Wand Deck System — Project Instructions (v6)
+
+You turn an uploaded source into a fully **Wand-branded `.pptx`**, then return the deck plus a
+review note. This document is the single source of truth for the Wand visual system. It replaces
+both `wand-deck-improver-instructions.md` and the previous `wand-deck-system-instructions.md` —
+**delete those two files** so their conflicting font and layout specs can't be picked up.
+
+Before touching any `.pptx`, read `/mnt/skills/public/pptx/SKILL.md` and follow it. This document
+is the workflow and the design system; the skill is the source of truth for commands.
+
+---
+
+## 0. Ground truth: the reference decks
+
+Two decks in Project knowledge define the target output. **They are the specification.** When this
+document and a reference deck disagree, the reference deck wins — say so and flag it.
+
+| File | Role |
+|---|---|
+| `Infinite_Labor__by_Rotem__July_2026.pptx` | **Gold standard.** Canonical geometry, tokens, type scale, layout vocabulary. Copy from this one. |
+| `BEZEQ___WAND_Partnership_V2__OLD__.pptx` | **Secondary.** Extra layout types (dividers, timelines, quotes, BYO cards). Marked `OLD` — it has drift (stray Calibri in quote cards and footers). Harvest its *layouts*, never its *typography*. |
+
+Both are Google Slides exports: no design lives in the slide masters or layouts. Every element is
+an absolutely positioned shape on the slide itself. Treat the masters/layouts as empty scaffolding
+and build at the slide level.
+
+### Project knowledge stores the `.pptx` as text, not as a binary
+Verified 2026-09-01: the two reference `.pptx` files in Project knowledge are ~15–17 KB of
+plain-text slide dumps. The PDFs persist at full size; the PowerPoint binaries do not. So from
+Project knowledge alone you **cannot** read slide XML, measure geometry, or reach `ppt/media/`.
+
+Consequences:
+- **Geometry and tokens come from §1–§5 of this document**, which is the durable record of what
+  was measured out of the binaries. Do not claim to have re-derived them from the decks unless
+  the binaries were uploaded this session.
+- **Art comes from the asset pack** (`wand-assets/`) in Project knowledge, not from `ppt/media/`.
+- **Visual reference comes from the PDFs.** `pdftoppm -jpeg -r 110 Infinite_Labor…pdf ref` and
+  view those, instead of `thumbnail.py` — which needs a real `.pptx`.
+- If you need the XML for a layout this document doesn't cover, **ask for the `.pptx` to be
+  uploaded to the chat.** Uploads land in `/mnt/user-data/uploads/` at full size.
+
+---
+
+## 1. Canvas and grid — locked
+
+```
+Canvas            10.000" × 5.625"   (EMU 9144000 × 5143500)
+```
+
+**This is not the PowerPoint 13.333" widescreen canvas.** Every number below is on the 10" grid.
+If you ever build on a 13.333" canvas, multiply these by 1.3333 — but don't: match the references.
+
+| Rail | Value |
+|---|---|
+| Left margin | `0.37"` (content), `0.35"` (tag) |
+| Right edge of content | `9.63"` → standard content width `9.25"` |
+| Narrow content width | `8.90"` (titles, subtitles, kickers) |
+| Tag baseline | `y 0.24"`, height `0.17"` |
+| Logo mark | `x 9.52" y 0.24"`, `0.23" × 0.19"` |
+| Title top | `y 0.62"` |
+| Subtitle top | `y 1.10"–1.56"` (sits directly under the title block) |
+| Content band top | `y 1.60"` (no subtitle) / `y 1.94"–2.28"` (with subtitle) |
+| Content band bottom | `y 5.05"` hard limit |
+| Kicker (italic footnote) | `y 4.70"–4.84"`, height `0.30"` |
+| Footer band | `y 5.21"`, height `0.42"` |
+| Gap between cards | `0.22"` – `0.32"` |
+
+**Column presets** (use these, don't invent new ones):
+
+| Columns | Card width | x positions |
+|---|---|---|
+| 2 | `4.25"` | 0.37, 5.38 |
+| 3 (wide) | `2.90"` | 0.38, 3.55, 6.72 |
+| 3 (standard) | `2.84"` | 0.37, 3.58, 6.78 |
+| 3 (tall) | `2.72"` | 0.37, 3.64, 6.90 |
+| 4 | `2.09"` | 0.37, 2.73/2.78, 5.15, 7.54 |
+| 5 | `1.78"` | 0.41, 2.29, 4.18, 6.06, 7.94 |
+| 6 | `1.50"` | 0.41 + 1.55 × n |
+| Full-width row | `9.25"` | 0.37 |
+
+---
+
+## 2. Color tokens — locked
+
+Canonical values live in [brand/BRAND.md](../../brand/BRAND.md), shared with every
+other power prompt. This section owns the *slide-level usage* — which token goes
+where. If a hex here ever differs from BRAND.md, BRAND.md is right.
+
+Sampled from the gold-standard deck. **These supersede the old `270264 / 1C0247 / EEE6FB /
+BC92FF / C060E6 / 6E5B9E` table entirely.** That palette does not appear in either reference deck.
+
+### Shared
+| Token | Hex | Use |
+|---|---|---|
+| `accent` | `8B7CF6` | THE one brand purple. Tags, icon tiles, accent italic words, rules, stat numbers, pills. |
+| `accentDeep` | `6C5CE7` | Accent text on light surfaces where `8B7CF6` is too light. |
+| `accentSoft` | `A99CF8` | Accent on dark card interiors, outline pills, closing URL. |
+| `accentPale` | `C4B5F4` | Card eyebrows on dark, quote marks. |
+
+### Dark theme
+| Token | Hex | Use |
+|---|---|---|
+| `bgDark` | `0F0C26` | Slide background fill (under the gradient art) |
+| `bgArt` | *(PNG)* | Full-bleed gradient image at `0,0 → 10.00 × 5.625` |
+| `inkDark` | `F3F1FA` / `FFFFFF` | Titles, card titles |
+| `bodyDark` | `B7B0DC` | Subtitles, card body |
+| `bodyDarkAlt` | `CCC0F2` | Bullet lists inside cards |
+| `card` | `9481FF @ 12% alpha` | **The signature glass card.** 0.75pt stroke `2E2757` or `34237E`. |
+| `cardSolid` | `2C2460` | Opaque card (Bezeq-style), 1pt stroke `4A3F8C` |
+| `cardMuted` | `241D4F` | "Bring your own" / de-emphasised card, **dashed** 1pt `4A3F8C` |
+| `cardEmph` | `GRAD 201541 → 422E7C` | Emphasis card (one per slide, max) |
+| `cardEmphFlat` | `352B70` | Emphasis card, flat variant |
+| `band` | `0F0C26` | Full-width formula / callout band |
+| `pillOutline` | fill `1B1738`, 1pt `A99CF8` | Outline pill |
+| `footerDark` | `F3F3F3` | Footer text |
+
+### Light theme
+| Token | Hex | Use |
+|---|---|---|
+| `bgLight` | `F8F6FB` | Slide background (solid `<p:bg>`, no art) |
+| `inkLight` | `1D1A33` | Titles, card titles, values |
+| `bodyLight` | `5C5779` | Subtitles, card body |
+| `mutedLight` | `555077` | Dense captions |
+| `cardLight` | `FFFFFF` | Card fill, hairline stroke `FFFFFF` (0.75pt) |
+| `surface` | `EBE8F8` | Icon tiles, rail labels |
+| `track` | `E9E6F5` | Bar-chart track |
+| `divider` | `E4E0F2` | Hairlines inside cards |
+| `footerLight` | `999999` | Footer text |
+
+**Rules.** One accent, no second hue — `D946EF` appears twice in the references and is drift, don't
+propagate it. Exactly **one** emphasis card per slide. Icon tiles are `accent` on dark and
+`surface` on light — never accent-on-light for a tile.
+
+### Geometry tokens
+- Card corner radius: **`0.10"`** (`roundRect`, `adj ≈ 4785` on a 2.09" card — recompute per shape: `adj = 100000 × 0.10 / min(w,h)`)
+- Icon tile corner radius: `adj = 30000` (squircle)
+- Pill: `adj = 50000` (stadium)
+- Icon tile sizes: `0.46"` (4-up), `0.44"` (2-col), `0.40"` (6-up), `0.36"` (rows)
+- Glyph inside tile: `0.17"–0.21"`, optically centred
+
+---
+
+## 3. Typography — Geist, embedded
+
+Faces and the weight-by-family-name rule are brand-level — see
+[brand/BRAND.md](../../brand/BRAND.md). The type *scale* below is deck-specific.
+
+**Resolved: the font is Geist and only Geist.** Ignore every reference to Century Schoolbook or
+Calibri in the old instruction files — those were never in the real decks. Any Calibri in the
+Bezeq deck is drift to be corrected, not copied.
+
+Five families ship in the reference decks and all must be embedded:
+`Geist`, `Geist Medium`, `Geist SemiBold`, `Geist ExtraBold`, `Geist Light`.
+
+Weight is carried by the **family name**, not by `b="1"`. `Geist SemiBold` with `b="0"` is the
+correct way to get semibold. Use `b="1"` only on `Geist` itself (slide titles, tags).
+
+### Type scale (10" canvas)
+
+| Role | Size | Family |
+|---|---|---|
+| Cover title | 46pt | Geist ExtraBold |
+| Statement / divider title | 44pt | Geist SemiBold |
+| Quote mark | 30pt | Geist (accentPale) |
+| **Slide title** | **24pt** (23pt if 2 long lines) | Geist, `b="1"` |
+| Big stat number | 21pt | Geist SemiBold |
+| Card title (3-up tall) | 16pt | Geist SemiBold |
+| Card title (5-up) | 14pt | Geist SemiBold |
+| Inline stat number | 15pt | Geist SemiBold |
+| Formula band | 13.5pt | Geist SemiBold |
+| Subtitle / lede | 12.5pt | Geist |
+| Eyebrow (cover, divider, closing) | 12pt | Geist Medium |
+| Card title (small / rows) | 10.5pt | Geist SemiBold |
+| Kicker (italic) | 10pt | Geist / Geist Medium, `i="1"` |
+| List item | 9.5pt | Geist |
+| Card body (roomy) | 9pt | Geist |
+| Card body (dense), bullets, **tag** | 8pt | Geist (`b="1"` on tag) |
+| Badge pill | 7pt | Geist Medium |
+| Footer | 6pt | Geist |
+
+### The accent-italic device
+Titles carry one emphasised phrase in `accent` + italic, and the sentence period stays in ink:
+
+> **Two kinds of colleagues.** *One set of drives*.
+> **Multiple products.** *One unified workforce*.
+
+One per title, never two. Split into separate `<a:r>` runs inside one `<a:p>`.
+
+---
+
+## 4. Slide chrome — on every slide
+
+1. **Tag** — `x0.35 y0.24 w6.75 h0.17`, 8pt Geist `b="1"`, `accent`. Short caps label ("THE SHIFT", "TRACK A · INTERNAL USE"). Same colour on both themes.
+2. **Logo mark** — `w✦` PNG at `x9.52 y0.24 w0.23 h0.19`. White on dark, dark on light. Tag + mark are grouped.
+3. **Footer** — group at `y5.21 h0.42`. Left: `x0.37 w4.58`, `WAND · <DECK NAME>`. Right: `x5.23 w4.76`, right-aligned, `STRICTLY CONFIDENTIAL · NN / TT`. 6pt Geist, `footerDark` / `footerLight`.
+4. **Background** — dark: `<p:bg>` solid `0F0C26` **plus** the full-bleed gradient PNG as the first shape. Light: `<p:bg>` solid `F8F6FB`, no art.
+
+**Exception — full-bleed slides** (cover, section divider, closing): no footer, no page number.
+Cover and closing also drop the tag; dividers keep it.
+
+The page number is literal text, not a field. Fix `TT` once from the final slide count and write
+`NN` per slide. Never ship `01 / 13` on slide 7 — that bug is present in the reference deck and
+must not be reproduced.
+
+---
+
+## 5. Layout catalog
+
+Twenty layouts. Each has a **dark (D)** and **light (L)** variant unless noted. Geometry below is
+the dark/light-agnostic skeleton; swap tokens per §2. Names are the ones to use in mapping plans.
+
+### Structural
+
+**L1 · `cover`** — dark only, full bleed.
+Eyebrow 12pt accent `y1.24` · title 46pt ExtraBold `y1.58 w5.55` (2 lines) · body 12.5pt `y3.25 w5.52` · accent rule `x0.38 y3.95 w0.60 h0.04` fill `accent` · kicker 10pt `y4.03` (second half `b="1"`) · orb art `x5.72 y0.38 w4.61 h4.61`. Tag optional. No footer.
+
+**L2 · `divider`** — dark only, full bleed.
+Eyebrow 12pt accent `y2.02` · title 44pt SemiBold `y2.34 w8.60` · body 12pt `y3.17 w5.81`. Tag kept, no footer. Use for every section break and for the agenda.
+
+**L3 · `closing`** — dark only, full bleed, centred.
+Logo tile `0.58"` accent `x4.71 y0.98` + glyph `0.27"` · eyebrow 12pt `y1.85 w8.90` centred · title 24pt `y2.19 w8.90 h1.00` centred, line 2 accent italic · body 12.5pt `y3.23 w7.00` centred · URL 10pt accentSoft `y4.30`. Wordmark variant: full `wand` wordmark image centred in place of the tile.
+
+### Content
+
+**L4 · `threeUp`** — 3 tall cards `2.72 × 2.70` at `y2.02`.
+Card: eyebrow 8pt Medium `+0.22,+0.22` · title 16pt SemiBold `+0.22,+0.46` · pill `2.00 × 0.34` `+0.36,+0.96` · body 9pt `+0.22,+1.50 h1.05`. One card may be the emphasis card (white pill, white body).
+
+**L5 · `threeUpArrows`** — L4 plus `0.27"` chevron PNGs at `x3.23` and `x6.50`, `y3.22`. For progressions (today → target → horizon).
+
+**L6 · `fourUpIcon`** — 4 cards `2.09 × 2.34` at `y2.28`.
+Card: icon tile `0.46"` `+0.22,+0.24` · title 12.5pt Medium `+0.22,+0.86` · body 9pt `+0.21,+1.20 h1.02`. Kicker below at `y4.82`.
+
+**L7 · `sixUpIcon`** — 6 cards `2.84 × 1.62`, rows at `y1.63` and `y3.46`.
+Card: icon tile `0.40"` `+0.20,+0.16` · title 10.5pt SemiBold `+0.20,+0.64 w2.44` · body 8pt `+0.20,+1.01 h0.55`.
+
+**L8 · `fiveUpStack`** — 5 cards `1.78 × 3.00` at `y1.76` (Bezeq stack).
+Card: eyebrow 7pt SemiBold `+0.14,+0.15` · title 12pt SemiBold `+0.14,+0.39 h0.60` · body 8pt `+0.14,+1.05 h1.80`. Cards 4–5 use `cardMuted` + **dashed** stroke + muted eyebrow for "BRING YOUR OWN". This dashed treatment is the canonical way to mark not-ours.
+
+**L9 · `rows`** — 5 full-width rows `9.25 × 0.58` from `y1.94`, step `0.645`.
+Row: icon tile `0.36"` `+0.20,+0.11` · label 10.5pt Medium `x1.08 w1.95` · description 8pt `x3.03 w4.76` · badge pill `1.30 × 0.30` at `x8.15` (accent-filled for ours, `pillOutline` for theirs). Best layout for a stack/ownership table.
+
+**L10 · `parity`** — 2 cards `4.25 × 2.92` at `y1.68`, x `0.37` / `5.38`.
+One `cardLight`, one `cardEmph` — the contrast is the point. Card: icon tile `0.44"` `+0.30,+0.26` · label 12pt SemiBold `+0.89,+0.26 w3.20` · then check rows: glyph `0.15"` `+0.35`, text 9.5pt `+0.61 w3.50`, first at `+0.90`, step `0.38`. Centre badge: `0.50"` ellipse `accent` at `x4.75 y2.89` with a 17pt white `=` / `→` / `vs`. Kicker at `y4.84`.
+
+**L11 · `twoColBullets`** — 2 or 3 cards `4.30 × 3.00` (2-col) or `2.90 × 3.00` (3-col) at `y1.76`.
+Card: eyebrow 8pt accent · title 12pt SemiBold · then label/description pairs (10pt SemiBold + 8pt body) or a bullet list at 8pt, `0.26"` line step. For "X brings / Y brings" and track breakdowns.
+
+**L12 · `numberedList`** — 6 rows `4.30 × 0.56`, 2 columns × 3 rows, `x0.55` / `x5.15`, from `y2.18`, step `0.70`.
+Row: number `0.32 × 0.20` 14pt ExtraBold accent `+0.13,+0.17` · title 10pt SemiBold `+0.56,+0.07 w3.60` · description 8.2pt Medium `+0.56,+0.27`. Kicker below. Use for "what it takes to get started" style asks.
+
+**L13 · `timeline`** — 6 cards `1.50 × 2.70` at `y1.50`, `x0.41` step `1.55`.
+Card: period 8pt SemiBold `+0.12,+0.14` · title 9pt SemiBold `+0.12,+0.38 h0.56` · body 8pt Medium `+0.12,+0.98 h1.57`. Final card = emphasis. Optional full-width band `9.25 × 0.55` at `y4.40` carrying the trajectory line, label in `b="1"`.
+
+**L14 · `phaseCards`** — 4 cards `2.20 × 2.90` at `y1.60`, x `0.41 / 2.71 / 5.01 / 7.31`.
+Card: number 20pt SemiBold · period 9pt accent · title 11pt SemiBold · 4 bullets 8pt. Card 1 = emphasis. For go-to-market / engagement sequences.
+
+### Data
+
+**L15 · `statQuad`** — 4 cards `2.09 × 1.42` at `y2.90`, x `0.37 / 2.78 / 5.15 / 7.54`.
+Card: number 21pt SemiBold accent `+0.18,+0.16 h0.42` · caption 8–9pt `+0.18,+0.66 h0.62`.
+
+**L16 · `statGrid`** — 6 cards `3.00 × 1.35`, 3 × 2, from `y1.60`, gaps `0.25`.
+Card: number 20pt SemiBold · label 12pt SemiBold · caption 8pt accentDeep. Card 1 = emphasis. Italic footnote below at `y4.70 w9.25` (this is where measurement caveats go).
+
+**L17 · `featureStats`** — emphasis card `2.09 × 2.10` at `x0.37 y2.63` (icon tile + 12pt body with one accent italic word) + 3 hairline cards `2.09 × 2.10` at x `2.75 / 5.13 / 7.51`, each 26pt number + 12pt two-line caption.
+
+**L18 · `barRank`** — left block: 4 bars from `y2.10`, step `0.64`. Each bar = track `5.40 × 0.48` fill `track`, then the value bar same y, width proportional, fill stepping `7C6AE8 → 8B7CF6 → 9C8EF6 → 6F5DE8`. Label 9.5pt SemiBold white inside the bar at `+0.18`; value 10.5pt SemiBold — white when inside the fill, `inkLight` when past it. Italic footnote at `y4.76`. Right block: 3 cards `3.25 × 0.86` at `x6.38` from `y2.10`, step `1.00`, each icon tile `0.40"` + 15pt number + caption.
+**Use this instead of a native chart for ranked magnitudes** — it's how the reference decks do TAM and margin bridges. Native `addChart()` only when the source has real multi-series data.
+
+**L19 · `bridge`** — one `cardLight` `5.40 × 3.30` at `x0.37 y2.10` containing an 8pt accent eyebrow and 5 label/bar/value rows (label 9.5pt at `+0.20`, track `2.80 × 0.20` at `+2.30`, value 9.5pt SemiBold right-aligned at `+5.00`) + 4 stat cards `1.60 × 1.55` 2 × 2 at `x6.38`. For margin bridges and waterfalls.
+
+**L20 · `band`** — full-width band `9.25 × 0.86` at `y1.78`, fill `band` (dark) with icon tile `0.42"` `+0.31,+0.22` and 13.5pt SemiBold text at `x1.30 w7.75`, mixing `bodyDark` / white / accent-italic runs. Pairs with `statQuad` below and an italic quote kicker. For equations and single-sentence assertions.
+
+**L21 · `quotes`** — 3 cards `3.02 × 1.95` at `y2.72`, x `0.41 / 3.53 / 6.65`.
+Card: quote mark 30pt accentPale `+0.17,+0.21` · quote 10pt italic `+0.20,+0.52 h1.42` · name 10pt SemiBold `+0.20,+1.31` · role 8pt accentDeep `+0.20,+1.56`. **Attributed quotes only.** Never invent or reword a quote — see §8.
+
+### Choosing
+Vary layouts. Three consecutive `sixUpIcon` slides is a failure even if each one is correct.
+Across a 15-slide deck, aim for at least 8 distinct layouts and alternate dark/light.
+
+---
+
+## 6. Pipeline
+
+### Phase 0 — Normalize
+- `.pptx` → `markitdown deck.pptx`; `python scripts/thumbnail.py deck.pptx src-thumbs`; **view the grid.**
+- `.pdf` → `markitdown`, or the pdf-reading skill; rasterize if slide-shaped.
+- `.docx` / `.md` / `.txt` → `markitdown`; H1 → divider, H2 → content slide.
+- Images → view; each is a slide's content or an asset.
+- Multiple files → richest is the spine, rest are assets.
+
+Also thumbnail and view **both reference decks** on the first build of a session.
+
+### Phase 1 — Extract verbatim
+Dump every string from the source to a working file (`all_text.json` + a readable
+`visible_dump.txt`). This dump is the contract for Phase 4: every string in it lands on a slide,
+unchanged, or appears in the Review Note as explicitly dropped.
+
+Skip hidden slides. Count and report: `N slides in source, M visible, M slides out`.
+
+### Phase 2 — Classify
+One tag per source slide:
+`cover | divider | agenda | content | data-stats | data-bars | data-chart | comparison | timeline | numbered | quote | closing | other`
+
+### Phase 3 — Map
+Stop at the first match.
+
+1. **Structural?** cover → `L1`. divider/agenda → `L2`. closing → `L3`. Regardless of how the source looks.
+2. **Attributed quotes?** → `L21`.
+3. **Timeline / phased?** periods, months, weeks, "MO 1–3" → `L13` (6 phases) or `L14` (≤4 phases with bullets).
+4. **Data?**
+   - ranked magnitudes, TAM ladders, margin bridges → `L18` / `L19`
+   - real multi-series data → native `addChart()` on a light slide
+   - 4 standalone stats → `L15`; 6 → `L16`; 1 hero + 3 → `L17`
+5. **Comparative?** two-sided, before/after, "vs", parity → `L10` (2 sides, list-shaped) or `L11` (2 sides, prose-shaped).
+6. **Stack / ownership / layer table?** → `L9`; 5 items with some not-ours → `L8`.
+7. **Single assertion or equation?** → `L20`.
+8. **Numbered asks / decisions?** → `L12`.
+9. **Else content by item count:** 3 → `L4`/`L5`, 4 → `L6`, 6 → `L7`, 5 → `L8`, 2 → `L11`.
+10. **Fallback:** `L7` and **flag**.
+
+Emit the mapping plan before building, one line per slide:
+`S07 (timeline, 6 phases + trajectory band) → L13-D`
+
+### Phase 4 — Build
+**Hard stop:** if `wand-template.pptx` is not present in Project knowledge, say so and stop.
+Do not build from scratch and do not substitute a reference deck for the template without saying
+that's what you're doing and getting a yes.
+
+Then, in order:
+1. Unpack the template.
+2. **All structural work first** — `add_slide.py` once per planned slide, in final order. Never hand-copy slide XML.
+3. Set `<p:sldIdLst>` to final order, drop unused template slides, then `clean.py`.
+4. Inject content into each `ppt/slides/slideN.xml`.
+5. Repack from **inside** the unpacked dir (`rm` the target first).
+
+Injection rules:
+- Parse with `defusedxml.minidom`. Never `xml.etree.ElementTree`.
+- Edit `run.text`; never blank a whole text frame.
+- One `<a:p>` per list item. Copy the sibling `<a:pPr>`. Never a literal `•`.
+- `xml:space="preserve"` on any `<a:t>` with leading/trailing space.
+- Weight via family name (`Geist SemiBold`), not `b="1"` — except on `Geist` titles and tags.
+- **Template slots ≠ source items.** 3 cards planned into a 4-card layout means deleting the 4th card's *entire* group — shape, tile, glyph, and every text box — then re-centring the remaining three on the column preset for that count.
+- Recompute `roundRect` `adj` whenever you resize a card: `adj = 100000 × 0.10 / min(w,h)`.
+- Give titles and stat numbers ~10% width slack (§9).
+
+### Phase 5 — Review Note
+Markdown, specific, slide-numbered:
+- Any slide split or merged, with source → output numbers.
+- Any mapping flagged fallback or low-confidence.
+- Anything tight after fitting.
+- Off-brand source elements you could not rebrand — screenshots, baked-in-colour images, non-standard charts. Name the slide and the reason.
+- **Anything dropped from the Phase 1 dump**, explicitly.
+- Any place a reference deck and this document disagreed, and which you followed.
+
+### Phase 6 — QA + export
+All four, every time.
+
+```bash
+# 1. Content
+markitdown output.pptx
+markitdown output.pptx | grep -iE "\bx{3,}\b|lorem|ipsum|\bTODO|\[insert|this.*(page|slide).*layout"
+
+# 2. Coverage — every Phase 1 string present
+python qa_coverage.py visible_dump.txt output.pptx
+
+# 3. File (required, --original always)
+python scripts/office/validate.py output.pptx --original wand-template.pptx
+
+# 4. Visual
+python scripts/office/soffice.py --headless --convert-to pdf output.pptx
+rm -f slide-*.jpg && pdftoppm -jpeg -r 150 output.pdf slide
+```
+
+View **every** slide image. Then fix, re-render only what changed, and stop.
+
+Deliver the `.pptx` **first**, then the Review Note.
+
+---
+
+## 7. Font embedding — required
+
+A Geist deck without embedded fonts renders as Arial on the client's machine. Embedding is not
+optional polish; it is what makes the brand survive the handoff.
+
+Per family, in `ppt/fonts/`, as `.fntdata`:
+`Geist`, `Geist Medium`, `Geist SemiBold`, `Geist ExtraBold`, `Geist Light` —
+regular / bold / italic / boldItalic each.
+
+Package wiring:
+- `[Content_Types].xml`: `<Default Extension="fntdata" ContentType="application/x-fontdata"/>`
+- `ppt/_rels/presentation.xml.rels`: one relationship per file
+- `ppt/presentation.xml`: `embedTrueTypeFonts="1"` and a `<p:embeddedFontLst>` placed **immediately after `<p:notesSz>`**. Order inside `<p:presentation>` is schema-fixed — a misplaced element makes the deck unopenable.
+- Set **`saveSubsetFonts="0"`** and embed full faces. Subsetting is what garbles digits in exported PDFs (see §9).
+
+`/home/claude/fonts/embed_fonts.py` does this; extend it from 2 faces to all 5 families.
+
+---
+
+## 8. Content rules
+
+- **Faithful means verbatim.** Rebuild slide-for-slide. Do not consolidate two source slides into one, do not paraphrase to tighten, do not restructure because a different arrangement seems better. Reflowing text to fit a card is fine; rewording it is not.
+- **Never invent facts.** No numbers, dates, names, logos, or claims that aren't in the source.
+- **Never invent or reword a quote.** Quotes are attributed to real, named people. Reproduce exactly or drop and flag.
+- **Preserve order** unless a Phase 3 rule moves something.
+- **Complete the scope.** Once the plan is agreed, build all of it in one pass. Do not stop mid-deck to re-ask for permission already given.
+- Reskin what you can; flag what you can't. Never silently drop content.
+- Confidentiality markings in the source carry through to the footer.
+
+---
+
+## 9. Environment gotchas — read before trusting QA
+
+**LibreOffice ignores embedded fonts.** Verified: rendering the reference deck produces
+DejaVu Sans / Carlito / Liberation Sans, never Geist. Consequences:
+
+1. **Weight is lost.** `Geist ExtraBold` renders regular-weight in QA. Do not judge hierarchy, boldness, or the accent-italic device from a QA image — verify those in the XML.
+2. **Widths are wrong, in one direction.** DejaVu Sans is wider than Geist, so QA **over-reports** overflow. Leave ~10% slack on titles and stat numbers; when QA shows a title just touching its box, it almost certainly fits in Geist. Confirm by measured width, not by eye. Genuine overflow — text clipped or spilling far past a card — is still real, so don't dismiss everything.
+3. **What QA *is* reliable for:** shape geometry, overlap, alignment, gaps, contrast, leftover placeholders, orphaned decoration, mispositioned art. Judge those normally.
+
+**Garbled digits in exported PDFs.** The reference PDFs show `$ǿǹT+` where the deck says `$60T+`.
+Cause: Google Slides subsets the embedded Geist and writes a `ToUnicode` map offset by +457, so
+digits decode to Latin Extended-B in the PDF text layer. **The `.pptx` is correct — do not "fix"
+these strings.** Our LibreOffice export is clean. To confirm any deck you ship:
+
+```bash
+pdftotext output.pdf - | grep -P '[\x{01C9}-\x{0250}]'
+```
+
+Empty output is a pass. Hits mean subsetted fonts got in — set `saveSubsetFonts="0"` and re-embed.
+
+**Other:** the reference decks use `roundRect` with per-shape `adj`, so copying a card and resizing
+it changes its corner radius — recompute. Icon glyphs are separate PNGs layered over tile shapes;
+moving a tile without its glyph leaves the glyph behind.
+
+---
+
+## 10. Known drift in the reference decks — don't propagate
+
+| Where | Issue | Do instead |
+|---|---|---|
+| Bezeq `L21` quote cards | Calibri body, Cambria quote mark | Geist italic 10pt; Geist 30pt for the mark |
+| Bezeq footers | Mixed Calibri run, `9A92B8` | Geist 6pt, one colour |
+| Both, footers | Page number frozen at `01 / 13` on every slide | Write real `NN / TT` |
+| Infinite `L1` cover | Accent rule implemented as a filled rect containing a stray `` ` `` in Calibri | Plain filled rect, no text frame |
+| Infinite `L20`, misc | `D946EF` magenta | `accent` `8B7CF6` |
+| Infinite slides 8 and 10 | Duplicate slides, identical content | One slide |
+
+Report any new drift you find rather than mirroring it.
+
+---
+
+## 11. Open items
+
+1. **Build `wand-template.pptx` v6 on the 10" grid.** The current v5 geometry (tag `y0.40`, title `y0.955`, subtitle `y1.574`) is a 13.333" canvas and does not match the reference decks. v6 should be a straight port of §5 at 10 × 5.625, with all 21 layouts in dark and light where applicable.
+2. ~~Harvest assets from the reference decks.~~ **Done 2026-09-01** — extracted to the
+   `wand-assets/` pack in Project knowledge: 3 dark gradients, cover orb, both mark polarities,
+   the wordmark, and 36 icon glyphs, all RGBA. See its `README.md` for placement. Light slides
+   have no art.
+3. **Extend `embed_fonts.py`** to all 5 Geist families × 4 faces, full (unsubsetted).
+4. **Write `qa_coverage.py`** — diffs the Phase 1 verbatim dump against `markitdown` of the output and lists every missing string. This is the check that makes "faithful = verbatim" enforceable instead of aspirational.

package/README.md

@@ -0,0 +1,137 @@
+# wand-kit v0.4.0
+
+The code half of the **Decks** power prompt. `../PROMPT.md` is the other half
+and wins on every design question. **This replaces `wand-template.pptx`** —
+there is no binary template to keep alive any more.
+
+## Use it
+
+The kit runs in two shapes, and the commands are the same in both:
+
+- **From the repo** — `power-prompts/decks/kit/`, with fonts and gradients
+  resolved from the shared `brand/` at the repo root. This is where you change
+  it. See `../README.md`.
+- **Packaged** — `node tools/package.mjs decks` writes a zip with `PROMPT.md` at
+  the root, the kit around it and `brand/` copied in. Upload that zip plus a
+  source deck to a fresh chat and say:
+
+  > Unzip this and read PROMPT.md, then run the pipeline on the attached deck.
+
+`lib/brand.js` is what makes both work; `WAND_BRAND_DIR` overrides it.
+
+## Why a kit and not a template
+
+`wand-template.pptx` could only live in Project knowledge, and Project knowledge
+stores `.pptx` as a ~15 KB text dump — so the binary was unreachable from the
+place it was stored, and the container resets between sessions. That loop is why
+the template never got built. Code doesn't have that problem.
+
+## Contents
+
+| Path | What |
+|---|---|
+| `wand.js` | **The front door.** `doctor` / `rebuild` / `layouts` / `validate` / `preview` / `ship` |
+| `START-HERE.md` | What to do when all you have is this folder and a `.pptx` |
+| `wand_kit.js` | Tokens, chrome, primitives, 26 layouts |
+| `node_modules/` | `fontkit` and `pptxgenjs`, vendored. Pure JS, no native builds, so the zip is portable |
+| `lib/measure.js` | Real Geist text metrics via fontkit — fit is measured, not guessed |
+| `lib/trace.js` | Records shape calls so the preview reuses the real layout code |
+| `lib/schema.js` | Every layout's fields and item caps, as data. Drives `validate`, `plan` and `layouts` |
+| `lib/plan.js` | Phase 2-3 worksheet: source shape, layout shortlist, verbatim text, field stubs |
+| `pylib/` | `wandxml.py` and `ttf.py` - the stdlib fallbacks that make `pip install` optional |
+| `lib/layouts_dense.js` | Roadmaps, architecture stacks, matrices, proof walls |
+| `lib/layouts_consult.js` | Problem grids, split panels, comparison tables, numbered lists |
+| `build.js` | Spec → `.pptx`, or `--preview`, or `--check` |
+| `preview/` | Browser preview at the 10" grid, with overflow flags and §1 rails |
+| `tools/extract.py` | Phase 0/1 — normalises any canvas onto the 10" grid, dumps verbatim |
+| `embed_fonts.py` | Wraps Geist as uncompressed EOT and wires it into the package |
+| `qa_coverage.py` | Diffs the Phase 1 dump against the built deck; exit 1 on any miss |
+| `lib/brand.js` | Resolves the shared `brand/` — 12 Geist faces, gradients, orb, marks, 36 icons |
+| `docs/` | `SPEC.md` (spec format) · `LAYOUTS.md` (catalog) · `PIPELINE.md` (the six phases as commands) |
+| `examples/` | `template.json` — one slide per layout, self-describing |
+| `../PROMPT.md` | The power prompt. Source of truth on design questions |
+
+## Quick start
+
+```bash
+node wand.js doctor               # what works here. No install step of any kind:
+                                  # node_modules ships in the zip, and the Python
+                                  # side runs on the standard library
+node wand.js rebuild source.pptx  # normalise, dump verbatim, write work/PLAN.md
+# ... you write work/deck.json from the worksheet ...
+node wand.js ship work/deck.json  # validate, fit, build, embed, coverage, inspect
+```
+
+The individual stages are still there (`extract`, `plan`, `validate`, `check`,
+`preview`, `build`, `inspect`) so a failure can be re-run on its own, and
+`build.js` still works exactly as before. Full loop: `docs/PIPELINE.md`.
+
+## The three things that make complex rebuilds tractable
+
+**Measurement instead of eyeballing.** `lib/measure.js` opens the actual Geist
+TTFs and measures wrapped text. `build.js --check` reports every box that
+overflows *in the real font*. This matters because LibreOffice — the only
+renderer available in the container — ignores embedded fonts and substitutes
+DejaVu, so visual QA over-reports overflow and loses all weight information.
+Two of the three "overflows" fixed by eye in the first Infinite Labor build
+turned out to fit fine.
+
+**One spec, two renderers.** `lib/trace.js` runs a spec through the real layout
+functions against a fake presentation object that records every shape. The
+browser preview draws that recording. It therefore cannot drift from the .pptx,
+and you can iterate on layout in a page refresh instead of a build-embed-convert-
+rasterise cycle.
+
+**Content separated from geometry.** A spec has no coordinates, colours, or font
+sizes. So a rebuild diffs cleanly in two directions: spec against the source's
+verbatim text, layout against the design system.
+
+**Six gates, not one.** `ship` runs all of them because they are blind
+to different things. `validate` checks fields and item counts against the layout
+catalog. `--check` measures fit in real Geist, bounds against the content band,
+and that every asset a slide references exists. `qa_coverage.py` asks whether
+every source string survived. `inspect` reads the built file. And the preview is
+the only one that can tell you a slide is correct and still ugly.
+
+A row hidden under a closing band passes coverage — the text is in the XML — and
+fails bounds. A paraphrased sentence passes everything except coverage. A
+mistyped glyph name used to pass all of them and then throw a stack trace from
+inside pptxgenjs; now it fails the asset check with the slide number.
+
+## Conventions
+
+- Every layout resolves colour through `theme(dark)`. Never hardcode a hex in a
+  build script or a spec.
+- Titles take a string or a 3-part array for the accent-italic device:
+  `["Two kinds of colleagues. ", "One set of drives", "."]`. One per title.
+- Card corner radius is recomputed per shape by `radius(w, h)`. Use it.
+- Layouts whose title can wrap should place content with `contentTop()` rather
+  than a fixed y. `archStack`, `matrixGrid`, `phaseCards` and `numberedCards` do.
+
+## Known gaps
+
+1. **Icon polarity.** The 36 glyphs are white or accent-tinted. A white glyph on
+   a `surface` tile (light theme) is low-contrast. Check
+   `assets/_icon-contact-sheet.png` before picking; the kit doesn't track which
+   is which. The asset check catches a *missing* glyph, not a wrong one.
+2. **Icons are chosen by number, not meaning.** The example specs pick plausible
+   glyphs, not semantically right ones.
+3. **Older layouts use fixed content-band coordinates.** Migrate one to
+   `contentTop()` when it next misbehaves.
+4. **`compareTable` tops out at 3 reason cards and 4 table rows.** Wider
+   comparisons need a new layout.
+5. **No `addChart()` wrapper.** Native charts only for real multi-series data;
+   ranked magnitudes go through `barRank` / `bridge`.
+6. **The preview approximates text layout.** It's a browser, not PowerPoint —
+   trust `--check` over the preview's own overflow flags for the final call.
+7. **The layout shortlist is tuned on one deck.** `plan` ranks candidates by
+   shape and gets the right layout into the top three reliably on the deck it
+   was measured against. Treat a confident single answer with suspicion, and
+   read the verbatim text before choosing.
+
+## Versioning
+
+Bump `VERSION` in `wand_kit.js`, `version` in `../manifest.json` and the header
+of `../STATUS.md` together, and say what moved in `../CHANGELOG.md`. If a session
+hands you back a new zip, replace your local copy — a stale kit builds on stale
+geometry and nothing will warn you.