…ary family

effectiveSignature was a ';'-joined delimited string of 'logical|weight|style=>physical|reason' entries; a font family is a free ST_String that may contain ';', '|', or '=>', so distinct resolution sets could serialize to the same key and cause wrong cache reuse. Serialize collision-safe JSON tuples instead (empty stays '' for the shared-cache fast path), matching FontResolver.signature.

primaryFamily now strips surrounding quotes like the resolver's normalizeFamilyKey, so a quoted primary ('"Calibri"') and its bare form collapse to one used face / report row / signature entry instead of two.

…correct

Bundled face metadata was registered only when a real document.fonts existed, but hasFace (the face-availability oracle the face-aware resolver consults) reads that metadata. So an editor with no font set (SSR/jsdom, some iframe/embedded timings) saw hasFace=false for every bundled face and disabled all bundled substitutes - even Calibri Regular stopped resolving to Carlito. Install the pack whenever a registry is resolved (metadata only; loading still needs a font set), guarded per registry instance so it installs once.

…-load'

fonts-changed folds the render plan's effectiveSignature into its dedup key, so the set of rendered faces changing from ordinary editing (e.g. the first Bold of a family) emits an event - but it defaulted to source: 'late-load' even though nothing loaded. Consumers filtering on 'late-load' got spurious load signals on typing. The font-config epoch bumps on a real late load and on a config mutation but NOT on editing, so distinguish them: epoch bump => 'late-load', key change with no epoch bump => new 'render-change'.

…+ realistic test

The report keyed 'missing' purely on the per-face load status, and the single-face-substitute test asserted loadStatus 'unloaded' / missing false for the pass-through Bold face. That state cannot occur in production: the planner awaits the pass-through (Georgia 700), and document.fonts.load resolves only registered faces (never system fonts), so the awaited pass-through always settles to 'fallback_used'. Make 'missing' reason-based for fallback_face_absent (the substitute lacks the face, so it is not faithfully rendered - deterministic, not probe-dependent) and rewrite the test to model the real awaited-pass-through path (fallback_used, missing true). getMissingFonts now correctly lists a family whose Bold has no faithful substitute.