fix: fall back to system font when a font lacks a glyph for a character

[chubes4]

What

Fixes the glyph-fallback half of Extra-Chill/data-machine-events#853 (the geometry/clipping half is a separate PR against data-machine-events, since it's a template-layout concern, not a renderer concern).

GDRenderer::draw_text() called imagettftext() with a single font and no fallback. When a codepoint is missing from that font's cmap table, FreeType/GD draws the font's own .notdef placeholder glyph — which, for the Extra Chill display font (Wilco Loft Sans), happens to be a literal NO GLYPH label baked into the glyph slot by the foundry. Production symptom: event OG cards for international venues rendered BAGGELYCKE G[NO GLYPH]RD instead of BAGGELYCKE GÅRD.

Why this layer owns it

GDRenderer is the shared rendering primitive under every GD-based image template in this plugin (event OG cards, quote cards, charts, flow diagrams — grep found 4 template consumers across data-machine, data-machine-events, and data-machine-socials). The bug is in how draw_text() resolves glyphs, not in any one template's layout, so the fix belongs here. Every template benefits without any per-template changes.

The fix

Added a minimal TrueType/OpenType cmap parser (parse_font_cmap() + parse_cmap_format_4() / parse_cmap_format_12()) that reads a font file's actual glyph-coverage table. has_glyph_coverage() checks a codepoint against the parsed table (cached per font path — parsed once, reused for the life of the request). draw_text() now splits text into runs by coverage (split_by_glyph_coverage()) and renders uncovered runs with the existing system fallback font (DejaVu Sans, already used elsewhere in this class when a font file can't be resolved at all) instead of the primary font's placeholder glyph.

No character list, no Unicode-range heuristic. Coverage is read from the font's own table. Proven necessary, not just convenient — the real Wilco Loft Sans font covers uppercase Å (U+00C5) but not lowercase å (U+00E5), which a case-insensitive or "is it Latin-1 Supplement" heuristic would have gotten wrong in one direction or the other. See test_display_font_lacks_accented_latin_but_not_ascii.

Supports both cmap format 4 (BMP segments, the common case) and format 12 (full-Unicode groups), which between them cover effectively every TTF/OTF font. Works for both TrueType and OpenType/CFF outlines since cmap lives in the same sfnt wrapper regardless of outline format.

Angled text (draw_text(..., $angle) with non-zero angle) skips the fallback path — run-splitting assumes a horizontal baseline for advancing $x between runs, which doesn't hold once rotated. No current caller passes non-ASCII text through an angled draw.

Verification

Unit tests (tests/Unit/Abilities/Media/GDRendererGlyphFallbackTest.php) — pure PHPUnit\Framework\TestCase, no WP bootstrap, run against real fonts:

PHPUnit 9.6.34 by Sebastian Bergmann and contributors.

GDRenderer Glyph Fallback (DataMachine\Tests\Unit\Abilities\Media\GDRendererGlyphFallback)
 ✔ System fallback covers basic latin
 ✔ System fallback covers accented latin
 ✔ System fallback lacks cjk
 ✔ Display font lacks accented latin but not ascii
 ✔ Unreadable font assumes coverage
 ✔ Run splitting isolates uncovered characters
 ✔ Run splitting returns single run for fully covered text

OK (7 tests, 19 assertions)

Tests run against the actual system DejaVu Sans and the actual production Wilco Loft Sans font file (skip gracefully if either isn't present in the environment, e.g. a bare CI checkout without the theme).

Rendered-pixel verification — instantiated the real EventOgCardTemplate + GDRenderer + the real Extra Chill brand-token bridge (WilcoLoftSans heading font, Helvetica body font) and rendered actual PNGs. BAGGELYCKE GÅRD (event 415181's exact venue name from the issue) renders å correctly — no NO GLYPH box. Screenshots and full render harness are in the linked data-machine-events PR, which is what actually exercises this code end-to-end.

Lint — phpcs (homeboy WordPress ruleset) clean on both changed files.

Not changed

No brand tokens, colors, proportions, or the geometry/layout logic in any template. This PR touches only glyph resolution inside GDRenderer::draw_text().

Refs Extra-Chill/data-machine-events#853

[chubes4]

Merging: the shard-4 failure is pre-existing and unrelated

ScaffoldAbilitiesTest::test_layer_scaffolding_skips_machine_managed_files and
::test_layer_scaffolding_aggregates_per_file_failures fail by shard
composition, not because of anything in this PR.

Root cause is tracked in #3539: MemoryFileRegistry::reset() also resets
WP_Agent_Memory_Registry, whose entries are registered at bootstrap and are
not restored by re-firing the datamachine_memory_files action. One test
calling reset() therefore leaves the registry permanently short for every test
after it in the process. ScaffoldAbilitiesTest has no set_up() and depends
entirely on ambient state, so its result is decided by which shard it lands in.

Evidence it is not this change

PR	Change	Shards
#3536	GDRenderer cmap glyph fallback	❌ same two assertions
#3538	BrandTokens logo resolution	❌ same two assertions
#3535, #3510, #3484	unrelated work	✅ shards pass

Two PRs with nothing in common fail identically, while unrelated PRs pass the
same shards. The only thing #3536 and #3538 share is adding a test file, which
reshuffles shard routing.

Holding correct fixes behind another component's test-isolation bug is the wrong
trade, particularly while NO GLYPH boxes render on production OG cards for
accented venue names.

#3539 stays open and owns the fix.