Shape Indic scripts, with the shaped cluster as the spacing unit

[ikashapov]

Shape Indic scripts, with the shaped cluster as the spacing unit

Closes #153.

What

Two halves, in the order the issue names them:

The predicate. needs_shaping gains a second property: the Brahmic
scripts whose OpenType shaping reorders glyphs relative to characters — a
Devanagari prebase matra is stored after its consonant and drawn before it,
which Joining_Type cannot see because those letters never join. The list
(Devanagari, Bengali, Gurmukhi, Gujarati, Oriya, Tamil, Telugu, Kannada,
Malayalam, Sinhala, Khmer, Myanmar) is drawn from the OpenType shaping models
themselves; Thai and Lao are deliberately excluded — their vowels are stored
in visual order, so the cmap walk is already right and their measured widths
must not move. New dependency: unicode-script.

The unit. §17.3.2.35 w:spacing and §17.3.1.13 distribute on a shaped
run now operate on the shaped cluster — HarfBuzz cluster values collected
from Skia's shaper and folded to UAX #29 grapheme starts, so an unreordered
combining mark can never become a spacing unit of its own (Skia reports
character-level clusters; unfolded, distribute opened a gap between ह and
its own matra). One HarfBuzz pass per shaped fragment measures both the
advance and the unit count; the count rides on the fragment so distribution
can count units without a measurer, and the painter groups its own shaped
glyphs by the same values, shifting each visual rank by one spacing step in
the same visual-gap model distribution reserves by.

Standing defects repaired along the way

• A w:spacing run in Arabic was measured and painted with no spacing at
  all (the documented pre-Complex-script shaping: make the spacing unit the shaped cluster so Indic reordering works #153 state).
• distribute stretched a shaped run's underline, borders and hyperlink rect
  while the painter left its glyphs unspread.
• §17.3.2.45 w:w scaled the reserved box but not the painted glyph
  positions of a shaped run — glyphs overlapped under expansion and the
  underline overran them.
• The emergency splitter measured pieces of a shaped word by cmap while the
  painter drew them shaped; for an Indic conjunct the cmap sum runs ~1.8× the
  ligature width. Pieces are now measured through the shaper (SplitMeasure
  trait, implemented by TextMeasurer).

Verification

• The 28-document corpus renders pixel-identical — the
  Latin/Cyrillic/CJK/Thai fast path is untouched.
• New fixture test-files/devanagari.docx (built by
  scripts/make_devanagari_fixture.py, OPC-verified) pins: the matra
  reorder, conjunct-safe w:spacing, distribute over Devanagari and Arabic
  lines, and a Latin w:spacing control that must stay off the shaping path.
  Exercised by tests/devanagari_shaping.rs with structural assertions (no
  host font names) and graceful skips for hosts without Devanagari coverage
  (ubuntu CI).
• The cluster-fold, the shift geometry, the predicate, the spacing formula,
  the shaped split measure and the RTL tie-break each carry a
  mutation-checked test — every deliberate break was caught by exactly the
  tests written for it.
• Rendering verified visually against LibreOffice; cargo fmt /
  clippy -D warnings / test --all / doc / --no-default-features all
  green.

README coverage table updated (76 ✅ / 12 ⚠️ / 10 ❌; the Indic row flips to
✅, the FAQ gap list drops it), AGENTS.md corpus table and entry-point notes
updated.