slide-maker 6.1.0 — brand kits, deck versions, numbers traced to your files

Brand kits, slide-by-slide versions, and every number held to the user's own files.
A company or lab deck can now be built from the user's logo, brand guide or old deck and checked for off-brand
colours and fonts; every rebuild is kept as a version that can be compared slide by slide and rolled back; and a
deck built from the user's material has every number on its slides traced to those files, with the files checked
against themselves first. Two end-to-end runs (one as Claude, one as a docs-only non-Claude agent) drove a round of
fixes to the gates, their scaffolds and the isolation of a user's own profile.

• Brand kit (scripts/brand_kit.py). extract reads a logo (SVG or raster), a brand-guide PDF, an old
  .pptx/.potx and explicit hex codes and font names into one draft spec, every value with its source; review,
  confirm (roles, installed faces, CJK glyph coverage, contrast), style (the palette and faces the build
  imports), check (off-brand colours, faces and logo use in the built deck, calibrated so tints of a brand colour
  pass) and save / list / use for reuse. A brand gate section on both runtimes.
• Deck versions (scripts/deck_versions.py). Each completed build is snapshotted into .versions/; diff
  aligns slides (an inserted, deleted or moved slide reads as one event), lists every changed number first and
  writes a side-by-side page; restore rolls back the whole deck, or single slides through the build script's own
  source. The hand-off can say how many slides changed since the version the user last saw.
• Numbers traced to the user's files (scripts/source_check.py). Every number on the built slides must be
  found in the files listed in content.sources (pdf, docx, pptx, xlsx including percent formats, csv, markdown,
  text, html); a rounded, re-scaled or computed figure is recorded with its source value or its expression and
  re-checked, a figure read off a picture is counted as not machine-checked, and the report names the closest file
  figure for a mistyped one. material checks the files against themselves (totals that are not the sum of their
  parts, shares off 100%, the same table in two files disagreeing) and each finding needs a recorded resolution.
  Scientific notation, abbreviated scales, a percent against a stored fraction, percent-formatted charts,
  year-month dates and GB18030 csv files are read as written. Both gates run it.
• The registry override is exclusive. With SLIDE_MAKER_REGISTRY set it is now the only registry root for
  reads and writes, and the taste ledger lives in it; before, reads fell through to the user's own profile, so an
  isolated run or a test still scored against it. Every test now runs isolated, and a guard keeps it so.
• Gates and scaffolds agree. The record scaffolds carry the fields the gates read (arc candidates, a one-row
  "n/a" image list, a note that a material-only deck needs no provenance section); the Codex template shows every
  delegated-pick axis; a self-review records that it had no fresh context instead of being made to claim one; an
  unattested icon or hotspot names the ids and the row that attests it; deck_gates.py picks and merge record
  picks and plan fragments (and refuse the Codex evidence's section names).
• Rules stated once. Review tiers, the fan-out threshold, the auto-mode review tier, the purpose keywords, the
  imagery vocabulary and the strict build-time lint in every scaffold now say the same thing everywhere, with a test
  holding them together; a confirmed provenance row may cite the user's own file.
• Layout. A two-line CJK title's accent rule sits below its last line; fonts are measured in the face the
  renderer draws (including lining figures of old-style faces), and CJK line breaks, the spaces at a break and
  mixed Chinese and Latin runs follow one rule set for the layout, the delivery lint and the measure.
• Tooling. CI runs as four parallel jobs; two lint passes reuse their measurements (identical output on 628
  decks); scripts/build_skillhub.py builds the SkillHub package (198 files, no images, every module importing)
  and CI refuses one that is over the limit.
• Docs. The READMEs lead with npx skills add, describe first use for every assistant, and show the visual
  styles as one template library with full-size previews.

Upgrading: a .deck-gates.json written before 6.1.0 needs content.sources ([] when the deck has no user
material), and a setup that relied on SLIDE_MAKER_REGISTRY falling back to ~/.claude/slide-templates must point
it at the folder it means.

Full changelog: https://github.com/addsumtech/slides_maker/blob/main/CHANGELOG.md