constellai 0.3.0

package/skills/design/presentation-html-fidelity-audit-layout-discipline-d5a323/scripts/extract_pptx.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+"""
+Extract every shape on every slide of a .pptx into a JSON dump.
+
+Usage:
+    python extract_pptx.py <path/to/deck.pptx>            # prints to stdout
+    python extract_pptx.py <path/to/deck.pptx> -o dump.json
+
+The dump captures the *actual* state of the export — text content, position,
+size, and per-run typography (font name, size, bold, italic, color). Use this
+as the ground truth for the fidelity audit; do not trust the export script's
+intent.
+
+Coordinates are reported in inches (rounded to 3 decimals) so they're
+human-readable when comparing against rails like CONTENT_MAX_Y = 6.70".
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+try:
+    from pptx import Presentation
+    from pptx.util import Emu
+except ImportError:
+    sys.stderr.write(
+        "python-pptx is required. Install with: pip install python-pptx\n"
+    )
+    sys.exit(2)
+
+
+def emu_to_in(emu: int | None) -> float | None:
+    if emu is None:
+        return None
+    return round(emu / 914400, 3)
+
+
+def color_repr(color) -> str | None:
+    """Best-effort color extraction. Returns hex string or None."""
+    if color is None:
+        return None
+    try:
+        # ColorFormat.type may be None when no explicit color is set.
+        if color.type is None:
+            return None
+        rgb = color.rgb
+        if rgb is None:
+            return None
+        return f"#{str(rgb).lower()}"
+    except (AttributeError, ValueError, TypeError):
+        return None
+
+
+def extract_runs(text_frame) -> list[dict]:
+    runs = []
+    for para in text_frame.paragraphs:
+        for run in para.runs:
+            font = run.font
+            runs.append({
+                "text": run.text,
+                "font": font.name,
+                "size_pt": float(font.size.pt) if font.size is not None else None,
+                "bold": bool(font.bold) if font.bold is not None else None,
+                "italic": bool(font.italic) if font.italic is not None else None,
+                # Color is independent of font name/size: a run can inherit
+                # font from the theme yet set its own color. Color drift is
+                # one of the things this audit needs to catch, so don't gate
+                # the extraction on unrelated font attributes.
+                "color": color_repr(font.color),
+            })
+    return runs
+
+
+def extract_shape(shape) -> dict:
+    data = {
+        "name": shape.name,
+        "shape_type": str(shape.shape_type) if shape.shape_type is not None else None,
+        "left_in": emu_to_in(shape.left),
+        "top_in": emu_to_in(shape.top),
+        "width_in": emu_to_in(shape.width),
+        "height_in": emu_to_in(shape.height),
+    }
+    if shape.left is not None and shape.height is not None and shape.top is not None:
+        data["bottom_in"] = emu_to_in(shape.top + shape.height)
+        data["right_in"] = emu_to_in(shape.left + shape.width)
+    if shape.has_text_frame:
+        tf = shape.text_frame
+        data["text"] = tf.text
+        data["runs"] = extract_runs(tf)
+    return data
+
+
+def extract_pptx(path: Path) -> dict:
+    prs = Presentation(str(path))
+    canvas = {
+        "width_in": emu_to_in(prs.slide_width),
+        "height_in": emu_to_in(prs.slide_height),
+    }
+    slides = []
+    for i, slide in enumerate(prs.slides, 1):
+        shapes = [extract_shape(s) for s in slide.shapes]
+        slides.append({"index": i, "shapes": shapes})
+    return {
+        "source": str(path),
+        "canvas": canvas,
+        "slide_count": len(slides),
+        "slides": slides,
+    }
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    ap.add_argument("path", type=Path, help=".pptx file to extract")
+    ap.add_argument("-o", "--output", type=Path, help="write JSON to this path; default stdout")
+    args = ap.parse_args()
+
+    if not args.path.exists():
+        ap.error(f"file not found: {args.path}")
+
+    data = extract_pptx(args.path)
+    payload = json.dumps(data, ensure_ascii=False, indent=2)
+    if args.output:
+        args.output.write_text(payload, encoding="utf-8")
+        sys.stderr.write(f"wrote {args.output} ({len(payload)} bytes, {data['slide_count']} slides)\n")
+    else:
+        sys.stdout.write(payload)
+        sys.stdout.write("\n")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/skills/design/presentation-html-fidelity-audit-layout-discipline-d5a323/scripts/verify_layout.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""
+Verify a re-exported .pptx against footer-rail + canvas-bound invariants.
+
+Usage:
+    python verify_layout.py <path/to/deck.pptx>
+    python verify_layout.py <path/to/deck.pptx> --content-max-y 6.70 --canvas-h 7.5
+
+Exits 0 on no violations, 1 on any violation. Prints a single block of
+violations sorted by slide index, one per line:
+
+    slide 5  shape 'desc-row-B-1'  bottom 7.214" crosses footer rail 6.70"
+    slide 11 shape 'note-paragraph' bottom 7.342" exceeds canvas 7.50"
+
+Use this as the gate for "this re-export is shippable". Don't claim the audit
+is fixed without running this script — the human eye misses 1–2 mm overflow
+at zoom-out, the script doesn't.
+
+Footer / chrome shapes are exempt from the content rail. Two heuristics
+identify them, in this order:
+
+1. **By name** — any shape whose name contains "footer", "foot", "chrome",
+   "page", or "pagination" (case-insensitive). Use semantic names in your
+   export script if you can.
+2. **By position** — any shape whose `top` is at or below the footer-zone
+   threshold (default `--footer-zone-top 6.80`). This catches python-pptx's
+   auto-generated names like "TextBox 3" when the export script didn't name
+   them. The threshold sits ~0.10" above FOOTER_TOP so chrome rows pinned
+   exactly at FOOTER_TOP are still recognized.
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+try:
+    from pptx import Presentation
+except ImportError:
+    sys.stderr.write(
+        "python-pptx is required. Install with: pip install python-pptx\n"
+    )
+    sys.exit(2)
+
+
+FOOTER_NAME_HINTS = ("footer", "foot", "chrome", "page", "pagination")
+EPS_IN = 0.005   # ignore sub-pixel overflows (~0.13mm)
+
+
+def is_footer_by_name(name: str) -> bool:
+    n = (name or "").lower()
+    return any(hint in n for hint in FOOTER_NAME_HINTS)
+
+
+def emu_to_in(emu: int | None) -> float:
+    return (emu or 0) / 914400
+
+
+def verify(path: Path, content_max_y: float, canvas_w: float, canvas_h: float,
+           footer_zone_top: float) -> list[str]:
+    prs = Presentation(str(path))
+    violations: list[str] = []
+
+    actual_w = emu_to_in(prs.slide_width)
+    actual_h = emu_to_in(prs.slide_height)
+    if abs(actual_w - canvas_w) > EPS_IN or abs(actual_h - canvas_h) > EPS_IN:
+        violations.append(
+            f"canvas mismatch: file is {actual_w:.3f}\" x {actual_h:.3f}\", "
+            f"expected {canvas_w}\" x {canvas_h}\""
+        )
+
+    for i, slide in enumerate(prs.slides, 1):
+        for shape in slide.shapes:
+            if shape.top is None or shape.height is None:
+                continue
+            top = emu_to_in(shape.top)
+            left = emu_to_in(shape.left)
+            bottom = top + emu_to_in(shape.height)
+            right = left + emu_to_in(shape.width)
+            name = shape.name or "<unnamed>"
+
+            # Off-canvas (hard fail for any shape).
+            if bottom > canvas_h + EPS_IN:
+                violations.append(
+                    f"slide {i:<2} shape '{name}' bottom {bottom:.3f}\" "
+                    f"exceeds canvas {canvas_h}\""
+                )
+            if right > canvas_w + EPS_IN:
+                violations.append(
+                    f"slide {i:<2} shape '{name}' right {right:.3f}\" "
+                    f"exceeds canvas width {canvas_w}\""
+                )
+            if top < -EPS_IN:
+                violations.append(
+                    f"slide {i:<2} shape '{name}' top {top:.3f}\" is negative"
+                )
+            if left < -EPS_IN:
+                violations.append(
+                    f"slide {i:<2} shape '{name}' left {left:.3f}\" is negative"
+                )
+
+            # Footer rail (only enforced on content shapes).
+            # Shape is exempt if (a) named like a footer, or
+            # (b) pinned at-or-below the footer zone threshold.
+            if is_footer_by_name(name) or top >= footer_zone_top - EPS_IN:
+                continue
+            if bottom > content_max_y + EPS_IN:
+                violations.append(
+                    f"slide {i:<2} shape '{name}' bottom {bottom:.3f}\" "
+                    f"crosses footer rail {content_max_y}\""
+                )
+
+    return violations
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    ap.add_argument("path", type=Path, help=".pptx file to verify")
+    ap.add_argument("--content-max-y", type=float, default=6.70,
+                    help="content rail in inches; nothing in content area may cross (default 6.70)")
+    ap.add_argument("--canvas-w", type=float, default=13.333,
+                    help="expected canvas width in inches (default 13.333 = 16:9)")
+    ap.add_argument("--canvas-h", type=float, default=7.5,
+                    help="expected canvas height in inches (default 7.5 = 16:9)")
+    ap.add_argument("--footer-zone-top", type=float, default=6.80,
+                    help="any shape with top >= this is treated as footer/chrome "
+                         "(default 6.80; sits 0.10\" above the typical FOOTER_TOP=6.85\")")
+    args = ap.parse_args()
+
+    if not args.path.exists():
+        ap.error(f"file not found: {args.path}")
+
+    violations = verify(args.path, args.content_max_y, args.canvas_w, args.canvas_h,
+                        args.footer_zone_top)
+    if violations:
+        sys.stderr.write("\n".join(violations) + "\n")
+        sys.stderr.write(f"\n{len(violations)} violation(s) found in {args.path}\n")
+        return 1
+    sys.stderr.write(f"OK: 0 violations across all slides in {args.path}\n")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/skills/design/presentation-html-fidelity-audit-playbook/SKILL.md

@@ -0,0 +1,257 @@
+---
+name: presentation-html-fidelity-audit-playbook
+description: |
+  Use this skill when Codex needs audit a python-pptx export against its source HTML deck, identify layout/content drift (footer overflow, cropped content, missing italic/em, lost styling, off-rhythm spacing), and re-export with strict footer-rail + cursor-flow layout discipline. Use this skill whenever the user has a .pptx that was generated from an HTML slide deck and asks to compare/audit/verify/fix the export — including phrases like "compare ppt with html", "fidelity audit", "fix the pptx", "ppt is cut off", "footer overlap", "italic missing in pptx", "re-export the deck", "pptx-html-fidelity-audit", or any case where a python-pptx → HTML round-trip needs verification or repair. Also trigger when the user shows you a deck.html and a deck.pptx side by side and is debugging visual differences. Recast from the original pptx-html-fidelity-audit/SKILL.md material as the presentation-html-fidelity-audit-playbook procedure.
+---
+
+# Presentation Html Fidelity Audit Playbook
+
+## Role
+
+Use this skill when Codex needs audit a python-pptx export against its source HTML deck, identify layout/content drift (footer overflow, cropped content, missing italic/em, lost styling, off-rhythm spacing), and re-export with strict footer-rail + cursor-flow layout discipline. Use this skill whenever the user has a .pptx that was generated from an HTML slide deck and asks to compare/audit/verify/fix the export — including phrases like "compare ppt with html", "fidelity audit", "fix the pptx", "ppt is cut off", "footer overlap", "italic missing in pptx", "re-export the deck", "pptx-html-fidelity-audit", or any case where a python-pptx → HTML round-trip needs verification or repair. Also trigger when the user shows you a deck.html and a deck.pptx side by side and is debugging visual differences. Recast from the original pptx-html-fidelity-audit/SKILL.md material as the presentation-html-fidelity-audit-playbook procedure.
+
+## Source Trace
+
+- Original Markdown: `pptx-html-fidelity-audit/SKILL.md`
+- Reformulated skill name: `presentation-html-fidelity-audit-playbook`
+
+## Operating Guidance
+
+Follow the rewritten material below as the working procedure. Keep code blocks, commands, file paths, URLs, dimensions, and API names exact when applying the skill.
+
+A repeatable workflow for catching the ways a `python-pptx` export silently drifts from its HTML source — and fixing them with a layout discipline that prevents the same regressions on the next pass.
+
+## When this skill applies
+
+The user has:
+
+- A source HTML slide deck (typically a single-file deck with `<section class="slide">` blocks):
+
+  ```html
+  <section class="slide light">
+    <div class="chrome">2026 · Q2 review</div>
+    <span class="kicker">Pillar 03</span>
+    <h2 class="h-xl">Shipping <em>velocity</em> doubled</h2>
+    <p class="lead">…</p>
+    <div class="foot">page 5 / 14</div>
+  </section>
+  ```
+
+- A PPTX file generated from that deck via python-pptx (or similar).
+- A suspicion (or visible evidence) that the PPTX doesn't match the HTML — text bleeding into the footer, italic words gone flat, hero slides not centered, sections cropped, tag styling lost.
+
+If the user only has *one* of those two artifacts, this skill doesn't apply yet — first compose the missing one, or ask the user to provide it.
+
+## Why this is hard (and why a skill helps)
+
+PPTX is a fixed-canvas, absolute-positioned medium. HTML is a fluid, flow-based medium. A naive python-pptx export pins each block at hand-picked `(top, left)` coordinates, which works for the *first slide it was tested on* and silently fails for every other slide whose content has different intrinsic height. The result is the most common drift modes:
+
+1. **Footer overflow** — content's `top + height` crosses into the footer row.
+2. **Off-canvas content** — bottom of last block exceeds `7.5"` (16:9 canvas).
+3. **Italic loss** — `<em>` in HTML never gets `run.font.italic = True`.
+4. **Hero slides not centered** — vertical-stack slides use `MARGIN_TOP` instead of computing center.
+5. **Box bounds intruding** — the text fits, but the *shape's bounding box* is oversized and visually crosses the rail.
+6. **Tag/styling loss** — colored chrome rows, kicker uppercase tracking, mono-vs-serif assignments quietly fall back to defaults.
+
+Every one of these is a *layout discipline* problem, not a content problem. Once you adopt the discipline, they stop happening.
+
+---
+
+## Workflow
+
+The audit is five steps. Avoid skip any of them — the discipline only works if the audit produces a real list of issues to drive the re-export. A fix-without-audit pass tends to leave half the issues alive.
+
+### Step 1 — Extract ground truth from the PPTX
+
+Execute `scripts/extract_pptx.py <path-to.pptx> > pptx_dump.json`. The script walks every shape on every slide and dumps text, position (`top` / `left`), size (`width` / `height`), and per-execute typography (font name, size pt, bold, italic, color). This is the *actual* state of the export — avoid trust the export script's intent, trust the dump.
+
+For 14-slide decks, the dump is ~30–60 KB and human-readable.
+
+### Step 2 — Walk the HTML structure
+
+Read the source HTML and enumerate `<section class="slide">` blocks. For each, note:
+
+- The slide's theme (`light` / `dark` / `hero light` / `hero dark`).
+- The `chrome` row text (top metadata).
+- The `kicker` (small uppercase eyebrow above the headline).
+- The headline (h-hero / h-xl / etc.) and any sub-head.
+- The body copy and any structured blocks (pipeline steps, cards, pillars, observation cards).
+- The `foot` row (bottom metadata).
+- Any `<em>` or italic-styled spans — italic is the silent regression.
+
+Map each HTML slide to a PPTX slide index. For decks following the convention "slide 1 = cover, slide N = closing", the mapping is positional.
+
+### Step 3 — Assemble the audit table
+
+For each slide, walk shapes from the dump and check against expected layout rules. Use this exact table format — the severity column is what drives the fix priority:
+
+```
+| Slide | Issue | Severity |
+|---|---|---|
+| 1 cover | meta-row 底端 6.95" 蓋過 footer (6.7") | 🔴 |
+| 5 checklist | row B 步驟描述底端 7.2" 切到 footer | 🔴 |
+| 8 3E | 收束段落直接坐在 footer 起點 | 🔴 |
+| 9 on-day | step 描述底端剛好碰 footer，無安全距 | 🟠 |
+| 多處 | em (Playfair italic) 未保留 | 🟡 |
+```
+
+Severity rubric:
+
+- 🔴 **critical** — content cropped, text invisible, footer overlap, off-canvas. Needs to fix.
+- 🟠 **high** — content visible but visual hierarchy broken, no breathing room, hero not centered. Prefer to fix.
+- 🟡 **medium** — italic/em missing, font fallback wrong, color drift. Fix in this pass.
+- 🟢 **low** — minor spacing/alignment, sub-pixel offsets. Note but avoid block.
+
+After the table, write a short root-cause section: 90 % of the issues usually come from 2–3 systemic causes (e.g. "no footer rail enforced", "hero stacks pinned to MARGIN_TOP instead of centered", "italic never propagated"). Naming the systemic causes makes the re-export script much smaller and more correct.
+
+### Step 4 — Re-export with footer-rail + cursor-flow layout discipline
+
+This is the load-bearing technique. See `references/layout-discipline.md` for the full rules; the summary:
+
+**Define the rails up front, once, for the whole deck:**
+
+```python
+from pptx.util import Inches
+
+CANVAS_W       = Inches(13.333)   # 16:9
+CANVAS_H       = Inches(7.5)
+MARGIN_X       = Inches(0.6)
+MARGIN_TOP     = Inches(0.5)
+CONTENT_MAX_Y  = Inches(6.70)     # NOTHING in content area may cross this
+FOOTER_TOP     = Inches(6.85)     # footer row pinned here, edge-to-edge
+```
+
+> **Customizing the rails.** The defaults above suit a 16:9 canvas with a slim footer. If your design system uses a wider footer or a 4:3 canvas, override these constants in your export script and pass the same values to `verify_layout.py` via `--content-max-y` / `--canvas-h` / `--canvas-w`. See `references/layout-discipline.md` §1 for the full constant table.
+
+
+**Use a cursor for content blocks instead of pinning each block at an absolute y:**
+
+```python
+class Cursor:
+    """Advances down the slide; refuses to cross the footer rail."""
+    def __init__(self, y_start, cap=CONTENT_MAX_Y):
+        self.y = y_start
+        self.cap = cap
+    def take(self, h, gap=Inches(0.12)):  # ~1 line of whitespace at 14pt; tighten/loosen per design system
+        top = self.y
+        self.y = top + h + gap
+        if self.y > self.cap:
+            raise OverflowError(
+                f"cursor at {self.y} exceeds footer rail {self.cap}; "
+                f"reduce block height or split slide"
+            )
+        return top
+```
+
+For each slide, instantiate `Cursor(MARGIN_TOP)` and `take(height)` each block in reading order. The slide refuses to render if any block would cross the rail, so overflows become loud assemble errors instead of silent visual bugs.
+
+**Hero (vertically-centered) slides use a budget instead of a cursor:**
+
+```python
+def hero_layout(blocks):
+    """blocks = list of (height, gap_after) tuples in reading order."""
+    total = sum(h + g for h, g in blocks)
+    y_start = (CANVAS_H - total) / 2
+    return Cursor(y_start)
+```
+
+That single change kills "hero slide content sticks to top" — the most common hero defect.
+
+**Tighten box height to fit text + minimal padding.** PowerPoint reveals shape bounds when they overlap (selection halos, Z-order conflicts), and an oversized box can visually cross the footer rail even when the text inside doesn't. Compute box height from text metrics + ~0.05" pad, not from generous wrappers.
+
+**Preserve italic / em explicitly:**
+
+```python
+def add_run(p, text, font, size_pt, italic=False, bold=False, color=None):
+    r = p.add_run()
+    r.text = text
+    r.font.name = font
+    r.font.size = Pt(size_pt)
+    r.font.italic = italic
+    r.font.bold = bold
+    if color:
+        r.font.color.rgb = color
+    return r
+```
+
+When walking HTML, detect `<em>` / `<i>` / inline style `font-style: italic` and pass `italic=True`. Use the EN serif face (Playfair Display, Origin Serif, or fallback Georgia) for italic display copy — the CJK serif typically has no italic and looks broken if you try to italicize it.
+
+For deeper font issues that the layout rails can't catch — variable-font traps where PowerPoint silently swaps to Calibri / Microsoft JhengHei, missing `<a:ea>` slot causing CJK runs to fall back, fake-italic on Han characters — read `references/font-discipline.md`. The five layers there cover everything `verify_layout.py` can't see.
+
+### Step 5 — Verify post-export
+
+After writing the new `.pptx`, execute `scripts/verify_layout.py <path-to.pptx>`. The script:
+
+- Walks every shape on every slide.
+- Asserts `top + height ≤ CONTENT_MAX_Y` for content shapes (footer/page-number shapes are allowed below the rail).
+- Asserts `top + height ≤ CANVAS_H` for all shapes (no off-canvas).
+- Asserts `left + width ≤ CANVAS_W` and `left ≥ 0`.
+- Reports violations as a single block: slide index, shape name, observed bottom, rail.
+
+Zero violations is the gate for "this re-export is shippable". Avoid claim the audit is fixed without running the verifier — the human eye misses 1–2 mm overflow at zoom-out, the script doesn't.
+
+---
+
+## Output to the user
+
+After Step 5 passes, report:
+
+1. **Audit table** — the table from Step 3.
+2. **Root causes** — 1-paragraph systemic explanation.
+3. **Fix list** — terse list of what was changed and why (e.g. "hero slides switched to budget centering", "all content blocks routed through Cursor", "em runs explicitly italic").
+4. **Verification** — "0 rail violations across N slides, file size X KB".
+5. **Path** — absolute path to the re-exported `.pptx`.
+
+The user is reading for two reasons: confirming the visible bugs are fixed, and trusting the systemic fix is right. Cover both.
+
+---
+
+## Bundled resources
+
+- `scripts/extract_pptx.py` — dump every shape on every slide as JSON. Execute before the audit. **Important:** also execute on the *original* export to compare, and on the *re-exported* one to confirm.
+- `scripts/verify_layout.py` — post-export rail checker. Returns nonzero exit code on violations so it slots into a CI pipeline if needed.
+- `references/layout-discipline.md` — the full footer-rail + cursor-flow rule set with code snippets for each common slide type (hero, content, pipeline, two-column, observation grid).
+- `references/font-discipline.md` — five-layer font audit: mapping, presence, variable-vs-static traps, the three XML language slots (`latin` / `ea` / `cs`), CJK + Latin italic interaction.
+- `references/audit-table-template.md` — copy-pasteable table template with severity legend.
+
+Read the references when:
+
+- The deck has slide types beyond what the SKILL.md covers (multi-column dashboards, embedded images, charts) → `layout-discipline.md`.
+- The audit shows 🟡 typography issues — italic missing, CJK falling back, unexpected `Calibri` / `Microsoft JhengHei` in the XML → `font-discipline.md`.
+- You want to drop the audit table directly into a report or markdown deliverable → `audit-table-template.md`.
+
+---
+
+## Anti-patterns to avoid
+
+- **Patching individual slides without naming the systemic cause.** If you fix slide 5 by lowering its block by 0.2", you'll be back fixing slide 9, 11, and 14 next. Find the rule that produced all four problems.
+- **Trusting the original export script's intent.** Consistently execute the extractor against the actual file. Drift between intent and reality is the bug.
+- **Skipping verification because "it looked fine in PowerPoint preview".** Preview anti-aliasing hides 1–2 mm overflows. The script doesn't.
+- **Italicizing scripts that have no italic tradition.** CJK, Arabic, Hebrew, Devanagari, Thai, and Khmer all produce a synthesized slant when forced into `italic=True`, and the result looks mechanically deformed. Italicize *only* runs whose primary script supports italic — Latin, Cyrillic, Greek. See `references/font-discipline.md` Layer 5 for the implementation pattern.
+- **Using `MARGIN_TOP` for hero slides.** Hero slides need *budget centering*, not top-anchored. This is the most common hero defect and the cheapest to fix.
+
+---
+
+## Why geometry-based verification, not visual diff
+
+An earlier iteration of this skill leaned on visual diffing — render the
+.pptx through Keynote → PDF → PNG, screenshot the HTML through Chrome
+headless, stitch them side-by-side with `magick`. It worked, but with
+three sharp drawbacks:
+
+- **Platform lock-in.** Keynote AppleScript is macOS-only; `magick` and
+  font-discovery commands vary across OSes; CI pipelines on Linux can't
+  reproduce the chain.
+- **Imprecision.** A 1-2 mm overflow gets anti-aliased away in a PNG
+  preview. The human eye misses it; the script catches it as a hard
+  numeric violation.
+- **Setup cost.** Every contributor needs the full graphics toolchain
+  installed before they can audit. Geometry checks need only
+  `python-pptx`.
+
+Geometry-based verification gives up one thing the visual diff is good
+at: catching cases where shape positions are correct but the rendered
+glyph looks wrong (font fallback, kerning bugs, missing weight). When
+that case appears, fall back to a manual screenshot review — the
+five-layer audit in `references/font-discipline.md` covers most of the
+underlying causes.

package/skills/design/presentation-html-fidelity-audit-playbook/scripts/extract_pptx.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+"""
+Extract every shape on every slide of a .pptx into a JSON dump.
+
+Usage:
+    python extract_pptx.py <path/to/deck.pptx>            # prints to stdout
+    python extract_pptx.py <path/to/deck.pptx> -o dump.json
+
+The dump captures the *actual* state of the export — text content, position,
+size, and per-run typography (font name, size, bold, italic, color). Use this
+as the ground truth for the fidelity audit; do not trust the export script's
+intent.
+
+Coordinates are reported in inches (rounded to 3 decimals) so they're
+human-readable when comparing against rails like CONTENT_MAX_Y = 6.70".
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+try:
+    from pptx import Presentation
+    from pptx.util import Emu
+except ImportError:
+    sys.stderr.write(
+        "python-pptx is required. Install with: pip install python-pptx\n"
+    )
+    sys.exit(2)
+
+
+def emu_to_in(emu: int | None) -> float | None:
+    if emu is None:
+        return None
+    return round(emu / 914400, 3)
+
+
+def color_repr(color) -> str | None:
+    """Best-effort color extraction. Returns hex string or None."""
+    if color is None:
+        return None
+    try:
+        # ColorFormat.type may be None when no explicit color is set.
+        if color.type is None:
+            return None
+        rgb = color.rgb
+        if rgb is None:
+            return None
+        return f"#{str(rgb).lower()}"
+    except (AttributeError, ValueError, TypeError):
+        return None
+
+
+def extract_runs(text_frame) -> list[dict]:
+    runs = []
+    for para in text_frame.paragraphs:
+        for run in para.runs:
+            font = run.font
+            runs.append({
+                "text": run.text,
+                "font": font.name,
+                "size_pt": float(font.size.pt) if font.size is not None else None,
+                "bold": bool(font.bold) if font.bold is not None else None,
+                "italic": bool(font.italic) if font.italic is not None else None,
+                # Color is independent of font name/size: a run can inherit
+                # font from the theme yet set its own color. Color drift is
+                # one of the things this audit needs to catch, so don't gate
+                # the extraction on unrelated font attributes.
+                "color": color_repr(font.color),
+            })
+    return runs
+
+
+def extract_shape(shape) -> dict:
+    data = {
+        "name": shape.name,
+        "shape_type": str(shape.shape_type) if shape.shape_type is not None else None,
+        "left_in": emu_to_in(shape.left),
+        "top_in": emu_to_in(shape.top),
+        "width_in": emu_to_in(shape.width),
+        "height_in": emu_to_in(shape.height),
+    }
+    if shape.left is not None and shape.height is not None and shape.top is not None:
+        data["bottom_in"] = emu_to_in(shape.top + shape.height)
+        data["right_in"] = emu_to_in(shape.left + shape.width)
+    if shape.has_text_frame:
+        tf = shape.text_frame
+        data["text"] = tf.text
+        data["runs"] = extract_runs(tf)
+    return data
+
+
+def extract_pptx(path: Path) -> dict:
+    prs = Presentation(str(path))
+    canvas = {
+        "width_in": emu_to_in(prs.slide_width),
+        "height_in": emu_to_in(prs.slide_height),
+    }
+    slides = []
+    for i, slide in enumerate(prs.slides, 1):
+        shapes = [extract_shape(s) for s in slide.shapes]
+        slides.append({"index": i, "shapes": shapes})
+    return {
+        "source": str(path),
+        "canvas": canvas,
+        "slide_count": len(slides),
+        "slides": slides,
+    }
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__.split("\n\n")[0])
+    ap.add_argument("path", type=Path, help=".pptx file to extract")
+    ap.add_argument("-o", "--output", type=Path, help="write JSON to this path; default stdout")
+    args = ap.parse_args()
+
+    if not args.path.exists():
+        ap.error(f"file not found: {args.path}")
+
+    data = extract_pptx(args.path)
+    payload = json.dumps(data, ensure_ascii=False, indent=2)
+    if args.output:
+        args.output.write_text(payload, encoding="utf-8")
+        sys.stderr.write(f"wrote {args.output} ({len(payload)} bytes, {data['slide_count']} slides)\n")
+    else:
+        sys.stdout.write(payload)
+        sys.stdout.write("\n")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())