Add design docs: glossary, ADR 0001, and project scaffolding

Records the outcome of the design review that reworked the original
handoff notes:

- CONTEXT.md — ubiquitous language for the slicer (slice, cut, discard,
  content rectangle, trim, song scale, bundle, project, jump source)
- docs/adr/0001 — the slicer owns all image processing and the bundle is
  the only channel to noteman; noteman's sharp pipeline goes away
- pyproject renamed to noteman-slicer
- .gitignore keeps scores and bundles out of the repo (copyright, size)

docs/adr/0001-slicer-owns-image-processing-bundle-is-the-only-channel.md

@@ -0,0 +1,48 @@
+# The slicer owns all image processing; the bundle is the only channel to noteman
+
+The slicer was originally scoped to share the work with noteman: it would cut the
+pages, upload them to `POST /api/admin/songs/:id/slices`, and noteman's sharp
+pipeline would trim, downscale and convert ink to alpha on arrival. We decided
+instead that the slicer produces finished, display-ready artifacts and hands them
+over as a single zip; noteman does no image manipulation beyond what plain HTML
+and CSS do at runtime, and there is no API between the two tools.
+
+## Considered options
+
+**Split the work (rejected).** Slicer cuts, noteman normalises. This was the
+existing shape and needed no changes on either side. It failed on inspection:
+noteman's `normalizeSlice` calls sharp's `removeAlpha()` before its luminance
+pass, and `removeAlpha()` discards the alpha channel rather than compositing it
+over white — so any transparency the slicer produced would flatten to black ink.
+A polygon-masked slice would have arrived as a solid black rectangle. More
+fundamentally, the split put trim, downscale and encoding rules in two repos and
+two languages, where they must agree exactly and nothing checks that they do.
+
+**Slicer produces final pixels, noteman stores them (chosen).**
+
+**Direct POST with a bypass flag (rejected).** Keeps the API but adds a
+content-sniff or flag to skip normalisation. Preserves the two-language
+duplication it was meant to remove, and needs an admin token in a local script.
+
+## Consequences
+
+Several rules that read as arbitrary in the slicer's code are only explicable
+from here:
+
+- The slicer owns **trim, staff-height normalisation, the 1920px ceiling, right
+  padding, levels and the ink→alpha conversion.** Nothing downstream can rescue
+  a faint scan or rehabilitate a bad crop, so the slicer's output is final.
+- **Scale is normalised on staff height, not width**, and derived once per song.
+  A per-slice factor would change note size line to line down a continuous
+  scroll; a width-based factor breaks the moment a slice comes from a different
+  source (a rescanned page, or a re-engraved replacement system).
+- **noteman must delete its sharp pipeline** and accept a bundle instead. Until
+  that lands, the slicer produces artifacts nothing consumes.
+- **The bundle is a versioned contract**, not an internal format. Two separately
+  maintained repos now agree on: the marker enum, slice ordering, and the shape
+  of `song.json`. That coupling is deliberately kept as small as possible —
+  which is why every jump source stores its target slice explicitly rather than
+  relying on an unwritten "exactly one Coda per song" resolution rule.
+- **The split is reversible only in one direction.** Re-adding server-side
+  processing means re-deriving all of the above; going further the other way
+  (more processing in the slicer) is free.