0.33.0

• feat(animation): derive a GIF from an existing video and fix the clip palette (#317)

• The shared GIF palette was chosen by pixel population, so on a clip with a
  large textured area the background claimed nearly the whole table and small
  saturated marks collapsed to the nearest muddy neighbour. It is now chosen
  for colour coverage over the set of colours the clip contains, collected at
  full 8-bit precision, so a mark survives however few pixels it covers.

  • Add gif_from_video, deriving a GIF from a video already on disk so the
    frames are rendered once and every other format read back off that file
  • Build the palette from a colour census rather than a spatial downsample,
    which blended one-pixel marks away before the quantiser could see them
  • Share build_clip_palette and quantize_to_palette between the rendered and
    derived paths so both quantise identically
  • Stream the video in two passes instead of buffering it, keeping the
    decoded RGB frames from ever being resident together
  • Warn when the source is chroma-subsampled, since that loss precedes the
    palette and no quantiser can undo it
  • Add quantize_method for clips better served by a population-weighted split
  • Validate writer inputs, close the decoder, resolve ffmpeg the same way for
    reading as for writing, and classify pixel formats by family
  • Document the real memory cost: Pillow accumulates the quantised frames, so
    peak stays proportional to the clip's length
  • Record in SCOPE.md why re-encoding cleopatra's own animation output is in
    scope, and what still is not

  Closes #315, #308

• feat(styling): add stamp_mark figure watermark / brand-mark helper (#314)

• Stamp a logo/watermark image onto a matplotlib Figure with one call,
  sized as a fraction of the figure so it stays proportional across the
  dpis a figure is exported at (MP4 master, web copy, GIF).

• • Draw on a frameless inset axes in figure-fraction coordinates (the
    dpi-independent counterpart of Figure.figimage). frac sizes the
    mark's longer side, so it is never distorted and always fits, in any
    of the four corners; margin is a scalar or an (x, y) pair.

• Optionally composite a centred, gaussian-blurred halo behind the mark
  (alpha-over into a single axes) so it separates from a busy or dark
  canvas; blur is a fraction of the mark's unpadded width, and the
  axes is grown so the mark's own painted extent still equals frac.

• Accept a file path (via Pillow) or an in-memory RGB/RGBA array (uint8
  0-255 or float 0-1); reject out-of-contract inputs -- bad shape, a
  non-uint8 non-float dtype, a float outside [0, 1] or with NaN/inf, a
  zero-size image, or a margin that pushes the mark off-canvas -- with
  clear ValueErrors.

• No new dependency (Pillow is already a base dependency; no SciPy).

• Add a docs reference page and a SCOPE.md note that reading a
  presentation asset (a logo, not user data) is an allowed exception.

• Closes #312

• feat(glyphs): add TexturedGlobeGlyph for 3-D textured globes (#316)

• Add cleopatra's first 3-D glyph: wrap an equirectangular (lon/lat)
  RGB(A) texture onto a tilted, spinnable sphere on a matplotlib Axes3D.

  • Take an (H, W, 3)/(H, W, 4) equirectangular array (the north-up
    layout of basemap.reference.relief()) and return (fig, Axes3D); a
    standalone class like HistogramGlyph, not a Glyph subclass.
  • draw(spin=...) rotates the globe about its fixed tilted polar axis by
    rotating only the once-sampled mesh; animate() returns a FuncAnimation
    of a full rotation.
  • Normalize textures by dtype: integer by dtype max, float by their own
    peak only when a channel exceeds 1 (RGB only, so alpha is preserved);
    NaN and negative cells render black; unknown render options raise
    ValueError.
  • Add no dependency (mpl_toolkits.mplot3d ships with matplotlib);
    default mesh 180x90, with the quadratic render cost documented.
  • Add a full test suite (100% line + branch) and a reference doc page.

  Closes #311

• feat(basemap): add world_texture() and mercator_to_equirectangular (#313)

• Port the two generic basemap helpers from the earthlens satellite
  showcase notebook's inline basemap_texture into cleopatra.basemap.tiles.

• • mercator_to_equirectangular(mosaic, bounds, n_lon, n_lat): a
    pure-NumPy area-averaging resample of a Web Mercator (EPSG:3857) tile
    mosaic onto an equirectangular lon/lat grid via np.add.reduceat. It
    reads the mosaic's own 3857 bounds and sizes each cell's divisor from
    the actual reduceat block widths so the poles do not seam. Returns
    float32 in the input value scale; the output always spans the globe,
    clamping out-of-coverage edges. No network, Pillow, or pyproj needed.

• world_texture(provider, *, zoom, n_lon, n_lat, cache, ...): the XYZ
  analogue of reference.relief -- fetches the whole 2**zoom world tile
  grid (zoom capped at 6), stitches, reprojects, and returns an
  (n_lat, n_lon, 3) float32 texture in [0, 1]. Accepts a provider name
  or a resolved TileProvider. Caches the texture under
  Config.get_cache_dir() with a guarded read that rebuilds a corrupt
  file and an atomic mkstemp write. Requires the [tiles] extra.

• The earthlens-specific two-tone / ocean-land recolour stays in the
  notebook.

• Closes #309, #310