Kami

kami · 紙

紙 · かみ means paper in Japanese.

Good content deserves good paper. Use Kami’s templates and layout rules to create documents and landing pages with serif fonts, warm backgrounds, and ink-blue accents.

Part of Kaku · Waza · Kami - Kaku writes code, Waza drills habits, Kami delivers documents.

Update check (non-blocking). At the start of a task, run bash scripts/check-update.sh. On the first run each calendar day, it writes a marker under the local XDG cache directory and then resolves Kami's latest published GitHub Release. It uploads no user document or task content. When a newer Kami is available it prints one line; relay that line to the user, then continue. It fails silently when offline, sandboxed, or without curl, and must never block the work.

Step 0 · Load brand profile (if exists)

Check ~/.config/kami/brand.md (preferred) or ~/.kami/brand.md (legacy fallback). If found, read references/brand-profile.md for the full four-layer application spec (placeholder substitution, session defaults, visual customization, habit notes) and its six guardrails. If no profile exists, continue without interruption.

Key rule: explicit prompt > editorial judgment > habit notes > frontmatter defaults > built-in defaults. Profile fills gaps silently; it never overrides the current conversation.

Step 0.5 · User project style scan (opt-in)

Run this only when the user explicitly references a sibling project as a visual reference: "like my site", "match the style of ", "use the look from ". Skip silently when no such reference exists.

When triggered, before generating: locate the referenced project's declared design tokens (CSS variables, tailwind/theme config, token files) and extract dominant colors, font stack, spacing scale, and border-radius scale; prefer declared tokens over inline literals. Merge into the in-session brand profile as Layer C (visual customization), not Layer B (session defaults); never override an explicit --brand flag or values the user typed this turn. Report back in one line before continuing: "scanned , extracted N colors / M fonts; using as visual reference."

Skip and fall back to the brand profile defaults if the referenced path does not exist, no CSS-like files are found, or the extraction would conflict with the user's explicit values in the current message.

Step 1 · Decide the language

Match the user's language. Chinese -> *.html / slides-weasy.html. English -> *-en.html / slides-weasy-en.html. Japanese -> CJK path (.html / slides-weasy.html / slides-marp.md) as best-effort, JP Mincho first, visual QA before shipping. Korean -> dedicated *-ko.html / slides-weasy-ko.html family, with slides-marp.md as the best-effort Markdown path, visual QA before shipping. Reference docs are shared English specs.

Default to the WeasyPrint HTML path; fall back to PPTX (slides*.py) only when the user explicitly needs an editable deck.

Design, writing, production, and diagram guidance live in CHEATSHEET.md and references/*.md. Step 3's tier table decides how much of it to open; read the lowest tier that covers the task, not all of it.

Code blocks with class="language-*" are highlighted only when optional Pygments is installed in the build environment. Without it, PDFs still render and code blocks stay monochrome.

Mathematics is strict LaTeX, never pseudo-math. When a document contains mathematics, author formulas only as standard LaTeX delimiters in rendered body text: inline \( ... \) and display \[ ... \]. Delimiters inside title metadata, form options, literal/code, script, style, SVG, template, and MathML regions are treated as text. Do not replace formulas with Unicode approximations, ASCII fractions, screenshots, or raw TeX printed on the page. The renderer requires Node.js 20 or Node.js 22 and newer. Before shipping HTML/PDF, run bash scripts/ensure_mathjax.sh, then python3 scripts/math_render.py --in-place filled.html; this converts every formula into MathJax SVG. render_pdf also performs this conversion in memory as a hard fallback and fails if MathJax or the TeX is invalid. Finish with python3 scripts/math_render.py --check filled.html; a completed HTML must contain no raw TeX delimiters in rendered body text.

Step 1.5 · Intent extraction (silent)

Before picking a template, silently confirm purpose, audience, hard constraints, and what outcome counts as success. Skip any dimension the conversation or the document type already answers (a resume's purpose is always "get an interview").

Question budget, global for this skill: infer first, from the decision tree and context. Ask at most once, in one compact question (max 2 sub-questions), and only when 2+ intent dimensions are genuinely unresolvable or two templates genuinely both fit. Never present the dimensions as a checklist.

Execution contract

Before creating or modifying an output, lock the contract: language, template, output format, page or length target, visual acceptance check, and verification command. Infer from the user's request when clear; ask only when missing fields materially change the deliverable.

Use the nearest existing template and verification path. Do not add a new template, shared CSS layer, dependency, script flag, or optional mode unless the current request cannot be satisfied without it.

A change to SKILL.md, templates, scripts, or references reaches npx skills add and plugin installs on the next push to main; Claude Desktop users get it only when the next release rebuilds kami.zip. Name the channel that carries the change before calling it shipped.

Work mode

Route by the artifact's current state before loading more guidance. This is an internal branch, not a new user-facing command.

Step 2 · Pick the document type

Changelog vs. release notes: The changelog template above is for styled document output. GitHub release notes are a separate deliverable; use /write with Release Note Template Mode.

Landing Page: Screen-first interactive template. No PDF output. Includes gallery carousel with auto-rotate, hero entrance animation, responsive breakpoints (880px / 480px), and prefers-reduced-motion support. Deploy as static HTML to Vercel / Netlify / any host. The agent fills {{PLACEHOLDER}} values and HTML comment blocks, then saves as a ready-to-serve .html file.

Landing Page companion files: For a production multilingual deploy, copy the five landing-page-*.example files alongside the main HTML and drop the .example suffix. {{SITE_ORIGIN}} is the scheme + host of your {{CANONICAL_URL}} (e.g. https://example.com). What each file covers, and the opt-in Accept-Language redirect: references/design.md Section 11 «Companion assets».

Production product site mode: docs, help, releases, changelog, roadmap, legal pages, or more than two locales means a site system, not a page. Lock the five items below from the source material before filling any template, asking only where a missing item would change the deliverable. Then follow references/design.md Section 11 «Product site system», and «Documentation site» in the same section once the page grows a docs or help shell.

Keep project-specific release artifacts, payment providers, appcast rules, and private local paths out of Kami.

Decision tree (use before asking)

Ask only when two cells genuinely both fit.

Ambiguity examples that justify a one-liner:

• "1.5 page career story with heavy visuals" -> ask "resume or portfolio?"
• "2 page exec summary with metric tiles" -> ask "one-pager or equity-report?"
• "5 page argument with several charts" -> ask "long-doc or portfolio?"

Diagrams (primitives, not a separate template type)

When the user asks for a diagram inside a long-doc / portfolio / slide (not a standalone document), route to assets/diagrams/ rather than a template. Eighteen types ship there; references/diagrams.md section 1 «Selection» maps each to what it shows, and section 7 lists the AI-slop patterns to avoid. Read it before drawing. Extract the <svg> block from the chosen file and drop it into a <figure> inside the document.

Three routes inside that file, by trigger:

Auto-select charts from data. When content carries numbers, pick the chart type yourself and embed it without asking. Two house calls that differ from the common default: a ~100% share renders as a donut only up to 6 items, and 7 or more becomes a horizontal bar; a single time series whose absolute count changes dominate (not rate) renders as bars, not a line. When several types fit, prefer the one that shows variance most clearly. Always embed inside a <figure> whose caption states the insight, not the data range.

Illustrations (host image model, not inline SVG)

Inline diagrams above are vector SVGs you assemble by hand. For a standalone raster illustration, or a redraw of a figure, photo, or screenshot in the Kami look, delegate the drawing to the host's own image generation. Never call an external image API or require a key; rendering is the host's job.

• If the running host exposes image generation, apply the brief below and render the image directly.
• If image generation is unavailable, output the same complete brief as text. Route from observed capability, not a remembered list of host names.

Brief: first state the claim the image must communicate, its destination and smallest display size, the accepted reference it should sit beside, and what must not appear. Then apply the Kami visual system: warm parchment (#f5f4ed) background, never pure white; one accent only, ink blue (#1B365D); all else warm gray with a yellow-brown undertone, no other colors; thin single-line geometric strokes and simple flat icons; no gradients, drop shadows, or 3D; serif labels; generous whitespace, composed like a figure in a well-typeset report. Full brief skeleton, icon rules, and QC checklist: references/diagrams.md «Illustration briefs».

Switch to this path (instead of enlarging a hand-assembled SVG) when a figure needs more detail than SVG assembly holds at the target display width, typically web-article figures at teaching depth (see references/diagrams.md density tiers). When one deliverable needs several generated images, drive them through a single handoff file: one line per image (slot, aspect ratio, shared style anchor, prompt, status), generate in batches of at most 5, update the status column after each batch, and check existing generated output before regenerating. The style anchor is shared by the whole batch; per-image style drift is the failure mode.

Step 2.1 · Source and material pass

Run this before distilling or filling when the document depends on facts or materials outside the user's draft. Skip it for personal drafts where the user already supplied everything.

Source check fires when the document names a company, product, person, release date, version, funding round, metric, market fact, or technical spec. Work from primary sources, note the source and date for facts that drive the document, and ask the user when sources conflict rather than choosing silently. references/writing.md «5. Sources before phrasing» owns the detail, including which current-sounding claims are banned until checked ("latest", "recent", "new", version numbers, launch dates, financial figures).

Material check fires when the subject is a company, product, project, venue, or personal brand. Confirm what makes it recognizable before layout:

A missing item gets a compact gap table and one question, never a generic stand-in, an approximated logo, or an invented value. references/writing.md «6. Materials serve recognition» carries the writing-side rules.

Logo fallback: when the request names no logo but the brand profile has a logo path, fill the commented .brand-logo slot in one-pager / portfolio / slides-weasy per references/brand-profile.md Layer C. Expand ~ to an absolute path; if the file is missing or the template has no slot, leave it commented and render without a logo, never a broken image. An explicit logo in the current request always wins.

Materials status block

After the material check, output a structured status block before continuing. This is a one-shot transparency display, not a question:

Materials status:
- Logo: OK /absolute/path/to/client-logo.svg
- Brand colors: OK #1B365D mapped to --brand
- Product screenshot: MISSING (using a pure-text layout; no placeholder image)
- UI screenshot: not required for this doc type

Use OK, MISSING, or not required. If a required item is missing and no user input arrived, ask once with the gap table; otherwise continue silently.

Step 2.5 · Distill raw content (if applicable)

Auto-detect whether to distill. Do not ask the user; judge from the input. Skip distill only when the content already maps one-to-one onto the template's sections with quantified metrics in place, or the user said "use as-is" / "直接用这个". When in doubt, run distill.

Distill acceptance: every factual claim, number, date, name, source, and material reference from the raw input must land either in a template section (per-doc structure in references/writing.md) or in a gap table of what the template needs but the content lacks. Share the gap table once; never guess to fill a gap.

Example gap-check:

Then proceed to Step 2.6 (slides) or the layout note (all other doc types) with structured, distilled content.

Persist the distilled content as a content IR (new documents)

When building a new document (not a text tweak on an existing one), write the distilled result and the locked execution contract to a content.json next to your working HTML before filling:

{
  "type": "resume",
  "lang": "cn",
  "brief": {
    "audience": "Hiring manager for a senior product role",
    "job": "Earn an interview by proving scope and outcomes",
    "template": "resume",
    "formats": ["html", "pdf"],
    "page_target": 2,
    "narrative": "Scope first, then evidence, then fit",
    "required_facts": ["team size", "measured outcomes"],
    "required_assets": [],
    "acceptance_checks": ["two pages", "all atomic facts survive"],
    "preserve": [],
    "explicit_deviations": []
  },
  "content": { ... }
}

type is one of the schema names in references/schemas/ (one-pager, letter, resume, long-doc, portfolio, slides, equity-report, changelog, landing-page). brief records why this artifact exists and how it will be judged; it is not audience copy. Only brief.required_assets joins the content-to-HTML coverage gate. For a visual repair, add target, evidence, and preserve so the fix cannot silently grow beyond the reported surface. Older IR files without brief remain valid, but every new document should write it. Read the matching content schema before writing: its $comment notes carry the per-field quality bar. Then validate before any layout work:

The top-level envelope is strict: it contains only type, lang, brief, and content. Use a language tag such as cn, en, ko, or zh-TW; misspelled tags and extra top-level fields fail before template filling begins.

python3 scripts/build.py --check-content content.json

Fix schema errors by fixing the content (or asking the user for the missing fact), not by loosening structure. After filling the template, re-run with the filled HTML to confirm no fact was dropped on the way in:

python3 scripts/build.py --check-content content.json filled.html

Values longer than 80 characters are treated as prose you may rephrase; short atomic values (names, metrics, dates, bullets) must survive verbatim.

Step 2.6 · Deck pre-flight (slides only)

Slides only. Every other doc type skips to Step 2.7.

Load references/deck-preflight.md before drafting for path selection (WeasyPrint HTML by default), page size, unresolved delivery choices, and slide content rules. Its intake follows the shared question budget.

Step 2.7 · Layout note (transparent, non-blocking)

Before loading specs and filling the template, write a short editor-style note stating the layout intent: template choice, length target, narrative arc, embedded diagrams, material status, and output formats. Match the document's language. Keep it under 80 words, written as prose, not a status panel. Continue immediately after; do not wait.

Example (CN):

使用中文财报模板，制作 2 页 A4，输出 HTML 和 PDF。内容依次为投资判断、目标价、估值、催化剂与风险，加入营收趋势和 FY26 收入构成图，使用已有 Logo，缺少产品图的页眉采用文字。

Example (EN):

Layout intent: Equity Report (EN), two pages A4. Open with thesis and price target, run through valuation (DCF and comparables), close on catalysts and risks. A revenue line chart and an FY26 waterfall sit mid-doc. Logo is in hand; product image is absent, so the header stays text-only. Output: HTML and PDF.

If the user pushes back, adjust; otherwise proceed to Step 3.

Step 3 · Load the right amount of spec

Pick the tier that matches the task. Default to the lowest tier that covers the work.

Step 4 · Fill content into the template

• Copy the template into your working directory; don't write HTML from scratch
• CSS stays untouched during content fill. Layout adjustments go through the Layout-tweak tier (Step 3) and stay within spec; any real style change syncs references/design.md and the sibling templates, never a single file
• Content follows writing.md: data over adjectives, distinctive phrasing over industry clichés
• Avoid patterns listed in references/anti-patterns.md: emptiness, fabrication, mimicry, excess, source gaps, tone contamination, landing page, image slots, slides
• Before filling, read the quality bar for your document type in writing.md section "Quality bars by document type". Structure is necessary but not sufficient: a resume bullet needs Action + Scope + Result + Business Outcome; an equity report needs variant perception + quantified catalysts; slides need assertion-evidence titles. Meeting the quality bar is as important as filling every placeholder.

Do not generate

These are the most common AI document failures. Cross-reference references/anti-patterns.md for the full list; --check-placeholders catches leftover template tokens mechanically.

• Do not invent metrics, financial data, or statistics; mark gaps with [DATA NEEDED: description]
• Do not use stock-image descriptions as image placeholders ("A diverse team collaborating in a modern office")
• Do not pad content to fill template slots (a resume with 3 real projects does not need 5 fabricated ones)

Fill PDF metadata (WeasyPrint reads these into the PDF)

Every template has meta placeholders in <head>. Fill all four before building:

<meta name="generator" content="Kami"> is already fixed in the template; do not change it.

Author inference: build.py automatically sets PDF /Author metadata from:

1. git config user.name (primary)
2. KAMI_AUTHOR environment variable (fallback)
3. "Kami" (final fallback)

For personal documents (resume/letter/portfolio), the HTML <meta name="author"> should match the person's name in the content. For non-personal documents (one-pager/long-doc), leave the placeholder as-is and let the build script infer it.

Step 4.1 · Per-page density target (multi-page templates only)

适用：slides-weasy / long-doc / portfolio / equity-report / changelog。不适用 resume / one-pager / letter（这些有独立的长度合约）。

正文页填充率目标 60-80%。封面 / 目录 / 末尾署名页豁免。这条规则解决的是 AI 生成多页文档时最常见的 draft 缺陷：把内容拆得太散，结果几页都填不满。

Items-per-page contract

Sparse-page merge rule

Before finalizing, scan the draft. Any body page that would render under 50% full → apply one of, in order:

1. Merge upward into the previous section.
2. Merge downward into the next section.
3. Remove the page and fold its one useful assertion or proof into a neighbor.
4. Convert existing information to a small diagram or table only when that form is clearer even without the density problem.

Forbidden ways to "fill" a sparse page: padding with filler prose, repeating the heading as a sentence, inventing statistics, restating the prior page in different words, or adding a callout, chart, icon, or image whose only job is to occupy space. If the merge options don't apply, the page itself shouldn't exist.

Last-page exemption

The last body page is allowed to run 40-60% fill. Forcing balance on the last page usually means padding. The colophon / closing slide may have any fill level.

Verify after build

python3 scripts/build.py --check-density   # flags >25% (WARN) / >50% (SPARSE) trailing whitespace

If a body page (not cover, not last page) gets a SPARSE warning, treat it as a draft defect and re-author with the merge rule.

Step 4.2 · Resume recruiter pass (resume only)

Mechanical checks (--check-placeholders, --check-resume-balance, --check-density) validate structure and layout, not prose. A resume can pass all of them and still read broken. After filling and before building, reread every project card the way a recruiter would, against the row definitions in references/resume-writing.md ("What goes in each row"): Role carries your position in the project, not background alone; Actions are verb-led, one concrete approach per sentence; Impact reads as an outcome, not a restatement of the process. One cross-row check on top: no row repeats another row's information.

Fix a failing row by rewriting from the source material. If the source cannot support a row (for example, no outcome fact exists), ask the user for the missing fact. Do not pad, and do not fall back to generic claims ("保障稳定运行", "improved efficiency").

This pass is internal: run it silently; surface it only when a row cannot be fixed without new information from the user.

Step 4.3 · Table editorial pass (any table)

Run this whenever the artifact contains a table. A Kami table should separate rows through alignment and breathing room first; rules are only quiet guides.

• One chromatic system: keep table text in the neutral ink hierarchy and use var(--border) for header, body, and total-row rules. Do not color values by category, use --brand borders, add per-column hues, tint the header, or draw vertical rules. Use weight, signs, and labels for meaning instead.
• Hairline hierarchy: header and total rules are 0.6pt; body row rules are 0.25pt. If the line is noticed before the values, it is too heavy. Total rows gain weight from typography plus the slightly stronger neutral rule, not from a blue line.
• Vertical padding floor: normal tables target at least 6pt on headers and 5pt on cells. One-page and resume templates may step down once to 5pt / 4pt after a real render proves the page contract needs it. .compact still keeps at least 3pt on headers and 2.5pt on cells.
• Compact is earned: use .compact for five or more columns, eight or more body rows, or a verified page-fit constraint. Do not use it by reflex, and do not reduce padding again before merging or re-authoring nearby content.
• Striping is exceptional: start without .striped. Add its neutral fill only when a table has at least eight body rows and the normal-size render still makes row tracking difficult. Do not use striping merely to decorate the table.

Verify the rendered page at normal viewing size. Each row must read as a separate unit without the rules forming a grid, the first and last lines of text must not crowd a rule, and changing the padding must re-pass neighboring pages plus the page-count contract.

Step 4.4 · Subtractive visual pass

Before rendering, remove every visual primitive that does not encode data, state, grouping, or a relationship. Kami hierarchy should come from type, whitespace, labels, alignment, and restrained fills before it comes from a line.

• Do not add decorative eyebrow ticks, short cover or contact rules, side-border accents on headings, quotations, or callouts, or fake dash bullets.
• A filled callout or analyst box needs only its fill, padding, and typography. A quotation needs only indentation, olive text, and reading space. A section title needs only type scale and margin.
• Keep rules that do real work: table hairlines, chart axes, diagram connectors, input boundaries, full-width separators between content regions, and current-state indicators.
• Use the deletion test: hide the line. If meaning, state, grouping, and navigation remain clear, delete it. Restore spacing rather than replacing it with another ornament.

Render every affected page after the pass. Confirm that the removed primitive did not collapse the intended pause, alter the page-count contract, or leave neighboring elements visually unanchored.

Step 4.5 · Auto-select output format

Do not ask the user which format to export. Decide from context:

PDF is the default for document templates; an explicit format request takes precedence. Landing pages ship as a ready-to-serve static HTML file. Add PPTX only when the user explicitly needs an editable deck. PNG follows sharing context.

Step 5 · Build & verify

python3 scripts/build.py --verify           # build all templates + page count + font check + slides
python3 scripts/build.py --verify resume-en # single target full verification
python3 scripts/build.py landing-page        # screen-first static HTML template check
python3 scripts/build.py --verify slides    # single slide deck verification
python3 scripts/build.py --check-placeholders path/to/filled.html
python3 scripts/build.py --check-markdown path/to/filled.pdf
bash scripts/ensure_mathjax.sh                            # strict TeX renderer dependency
python3 scripts/math_render.py --in-place filled.html     # TeX -> MathJax SVG in delivered HTML
python3 scripts/math_render.py --check filled.html        # fail on raw/unrendered TeX
python3 scripts/build.py --check-content content.json path/to/filled.html
python3 scripts/build.py --check-visual path/to/filled.pdf
python3 scripts/build.py --check-fonts path/to/filled.pdf   # which family actually drew the CJK text
python3 scripts/build.py --check-style path/to/filled.html  # template rules against the produced document
python3 scripts/build.py --check-resume-balance path/to/resume.pdf
python3 scripts/build.py --check-density path/to/filled.pdf  # page whitespace (one-page docs included)
python3 scripts/build.py --check-density              # repo sweep (skips cover and template skeletons)
python3 scripts/build.py --check-rhythm slides slides-en   # warn on monotonous slide sequences
python3 scripts/build.py --doctor         # installed render/check/font capability report
python3 scripts/build.py --check            # lint + token/theme + public-site fact checks

Strict LaTeX mathematics: Use only \( inline \) / \[ display \] as formula source. The delivered HTML/PDF must contain MathJax SVG, not Unicode pseudo-formulas, raw TeX, or formula screenshots. Run ensure_mathjax.sh, math_render.py --in-place, and math_render.py --check before render/hand-off.

Screen verify: --check-density is a print gate. For ANY browser-delivered surface (landing page, docs page, dashboard, testimonial wall, article index), screenshotting the rendered page across the responsive matrix in every locale is a hard step before declaring done, not an on-request extra: scan for line widows, sparse blocks, and single-line-surface wraps, and report the result. Do not wait for the user to ask "does it work on mobile". See references/design.md Section 12 «Responsive screenshot verification».

Perceptual verify (PDF deliverables): geometry checks cannot see a fallback glyph or an arrow crossing a label. Before shipping a filled PDF, run python3 scripts/build.py --check-visual path/to/filled.pdf, then view every exported page image against the printed checklist. One hit means a whole-document sweep for that class of issue. If your host cannot view images, send the image paths and checklist to the user instead of skipping the pass. --check-visual runs the font gate for you and prints its verdict above the checklist.

Font verify (CJK deliverables): a missing CJK serif produces no fallback boxes. It silently substitutes a sans that still reads, so the page passes an eyeball pass while carrying typography the parchment metrics were never tuned for, and the result reads heavy and flat without anything looking obviously broken. --check-fonts settles it from the rendered PDF's own span table: it names the family that drew the body ideographs and fails on a sans substitution or on text split across two families. Never report a CJK document as visually verified without it, and never assume a sandbox has the fonts: the commercial TsangerJinKai02 files never ship inside the skill package.

Fresh review: after the mechanical and perceptual checks pass, review once from the artifact contract rather than from the builder's rationale. Read brief, the content-coverage result, and the rendered evidence; check every acceptance item and every preserve boundary; report P0/P1 findings with the page, viewport, or element that proves them. Use an isolated reviewer when the host supports one. Otherwise reload those three evidence surfaces and do a distinct second pass. Fix P0/P1 findings before handoff; do not let the same pass that made the artifact approve its own intentions.

Source templates intentionally keep {{...}} fields. Run placeholder checks on completed documents, not on the template library.

For Markdown-sourced long documents, also run --check-markdown on the rendered PDF. It catches visible raw ---, **bold**, and inline-code backticks that should have been converted or removed before delivery.

Visual anomalies (tag double rectangle, font fallback, page break issues) -> production.md Part 4.

Definition of done

A task is done when the user receives, in the closing message:

1. The path of every deliverable, in every promised format (Step 4.5).
2. Which checks ran and their results, including the page-count contract.
3. Every remaining [DATA NEEDED] gap, listed explicitly. Never declare done with an unreported gap.
4. The visual verdict, stated honestly by surface: for PDFs the --check-visual pass status (which includes the font gate); for screen surfaces the responsive-matrix screenshot result; when rendering could not be inspected, say "build verified, visuals unconfirmed", not "done".
5. For documents containing mathematics, the strict LaTeX result: MathJax SVG rendering and scripts/math_render.py --check both passed.

Fonts

When --check-fonts reports a sans substitution or a split family, or a render fails on a missing font: run bash scripts/ensure-fonts.sh, then read references/production.md Part 1 «Fonts» for the stacks, the fallback chains, and where the recovered files land. Waiting for

…(truncated)