The Slide Stenographer — One Document Per Slide, and the Archive Renamed Around It

new(slides): slide_stenographer — per-slide transcription into markdown
new(slides): slide_schemas — open slide-type vocabulary with typed content shapes
new(slides): date_resolution — cascade that types dates before dating a deck
new(slides): deck_lineage — version vs fork vs sibling, and animation-reveal detection
new(slides): slide_cdn — ImageKit mirror via the prep-images-for-embed skill
new(slides): archive_rename — convention renaming with a reversible manifest
new(tests): regression cover for every failure mode found against real decks

`deck_analyst.py` is untouched, including its JSON, its screenshots, and its
section drafts. The MemoPop workflow depends on it and nothing here changes
that. This is a second mode beside it, invoked deliberately.

The two agents read the same decks at different granularity. The deck analyst
batches five slides per API call and merges the results into one memo-shaped
object — efficient for producing section drafts, and it destroys slide identity
in the process. There is no point at which that pipeline knows what slide 7
said. For a corpus you can search, decks you want to rebuild, and memos that
cite a slide, the slide has to survive as a unit.

So: one render, one call, one document. Text layer first and authoritative for
wording — a PDF's own text is a better record of a heading than any
transcription of a picture of that heading — with Vision reading what the text
layer cannot carry: layout, charts, which people have photographs, what a
diagram depicts. Where they disagree the text layer wins and the disagreement
is recorded.

Four subsystems fell out of running this against real archives.

**Dates need typing, not just sourcing.** A date printed on a slide is not
necessarily the deck's date. One deck's closing slide reads "Closing Sept 30" —
a future deadline, four days after the file was last modified. Another's first
slide reads "BoD Meeting 06-13-2025", which *is* the deck's date; that same file
is named for 09-24 and carries a PDF title naming yet a third date. Same rung of
the cascade, opposite meanings. Dates are now classified by kind
(`deck_date`, `meeting_date`, `closing_deadline`, `as_of`, `mentioned`) and only
some kinds may date a deck. Every resolution records which rung produced it and
at what confidence, and says so out loud when the winner contradicts the
filename.

**Lineage needs three distinctions, and fuzzy matching to make any of them.**
Multiple versions of one deck per round and multiple rounds per company mean
time-series analysis only works if the same logical slide is recognizable across
versions — and if things that are not versions stay out of the series. A deck a
firm assembles from a company's slides is a *fork*, not v2: attributing its edits
to the company misreads the series. Two decks in one round sharing no slides are
*siblings*. A PDF export of a PPTX is a *duplicate*. Authorship is decisive over
overlap. The fingerprinting underneath has to be lenient: an early exact-hash
version reported that a deck and its own export shared 8 of 17 slides when they
share all 17, the differences being line-break placement between two text
extractors. Exact matching fabricates version churn.

**Animation reveals are not slides.** A slide animated so content appears in
stages, exported to PDF, writes each stage as its own page — one slide, five
pages. Untreated, one logical slide becomes five documents and pollutes both the
corpus and the series. The industry calls these "builds"; that word is avoided
throughout because in a codebase it means compilation.

**A portfolio archive is not a dataroom, and its filenames show it.** Files
arrive named for whoever exported them. `archive_rename.py` puts them on the
house convention — `<YYYYMMDD>_<Company>_<Collection>--<Round>/` with each file
carrying its own date — planning and applying as separate calls, every plan
written to a manifest before a file moves, and the manifest sufficient to
reverse the whole thing.

**What the taxonomy is for, and what it is not.** The slide-type vocabulary
exists so a corpus query for "every team slide" returns every team slide. That
requires most slides to land on shared names, not every slide. An earlier
version enforced it: validation rejected an unrecognized type and told the agent
to use `other` instead. Run against a long-form reading deck, that filed eight
slides of organizational design as `other` — a taxonomy gap reported as noise,
with the observation discarded. The vocabulary is now open. An agent that meets
something genuinely new coins a `lower_snake_case` type and explains why;
proposals accrete in `_proposed-slide-types.yaml` with an occurrence count and
the slides that produced them, for promotion into the canon. One sighting is an
oddity; the same coined type across three decks is a type the canon is missing.

**Also included:** `prep-images.mjs` in the `prep-images-for-embed` skill now
kebab-cases each path segment instead of flattening slashes, so `--repo` and
`--slug` can each describe more than one CDN folder level. Backward compatible —
a value with no slash behaves exactly as before.