Fix edge cases in glyph rasterization

[mourner]

Fixes two bugs in glyph rasterization, both rooted in misinterpreting TextMetrics:

Fix #63: Inner EDT misses outside-ink seeds in the buffer region

The inner Euclidean distance transform was restricted to [buffer, buffer, glyphWidth, glyphHeight], which excluded the surrounding buffer pixels as seeds. When ink touches the bbox edge (common — Math.ceil(actualBoundingBoxAscent) puts the topmost ink row exactly at the bbox top), the nearest "outside" pixel sits one row outside the bbox and is invisible to the EDT, collapsing the inner distance and producing a hard cutoff (see #63's video).

Fix: pad the inner EDT region by 1 px on each side. The buffer region is uniformly seeded as outside-ink, so the closest external seed for any interior pixel is always exactly one pixel outside the bbox — 1 px of padding is provably sufficient. Padding is clamped to min(buffer, 1) to stay in-bounds when buffer = 0.

Fix #65: actualBoundingBoxLeft sign convention

Per the HTML spec, actualBoundingBoxLeft is positive when ink extends to the left of the origin. The old code treated it as positive=right, so glyphs whose ink sits to the right of origin (typical for letters like "P") were placed at the wrong position within the SDF texture, with extra phantom padding on one side.

Fix: negate when computing glyphLeft = Math.floor(-actualBoundingBoxLeft). This preserves glyphLeft's existing public semantics (canvas-x of ink left edge, positive=right), so all the downstream formulas (glyphWidth, fillText placement) keep working unchanged.

Downstream impact

The standalone SDF output changes — glyphs are now correctly placed within their texture without phantom padding. Consumers that position quads via glyphLeft (e.g. mapbox-gl-js's metrics.left in symbol/quads.ts) render pixel-identically: the old bug was self-cancelling because the wrong ink offset within the texture was compensated by the equally wrong glyphLeft value used for placement. Glyph atlas entries become slightly smaller for glyphs with non-zero actualBoundingBoxLeft (memory win).

Tests

• Existing 材 fixtures regenerated — old fixtures encoded the buggy ink placement.
• does not return negative-width glyphs mock updated: the original metrics described a legitimate (if unusual) glyph under the corrected sign convention; flipping ABBLeft's sign restores the test's pathological-input intent.

Closes #63, closes #65. Supersedes #64 and #66.