Design criticism wanted — argue with the contract

[DivyamTalwar]

This thread is for arguing with the design, not for bug reports.

The contract in skills/fablewright/SKILL.md makes a series of choices that are opinionated and could reasonably be made differently. Disagreement is more useful to me than agreement, so here are the ones I think are most attackable — take any of them apart.

solo is the default. Every other orchestrator sells parallelism; this one makes you justify delegating at all. The delegation test is: if specifying the work requires solving it, delegation has already spent the expensive part. Is that too conservative for real teams?

Independence is defined by vendor family, not by capability. Claude / GPT / DeepSeek / GLM is a table, which makes independence decidable rather than a judgement call — but two models from different vendors trained on overlapping data are not obviously independent in the way the word implies. Is family the right proxy?

The reader is read-only by tool set. It is structurally incapable of fixing what it finds, so its verdict cannot be contaminated by its own patch. The cost is a round trip. Is that trade right?

Routing proof is post-hoc. The pin is checked against the session record after the lane runs, which is why a refused run can still have modified your tree. Should it fail earlier and accept a weaker guarantee, or is refusing late but honestly the better trade?

ensemble requires a single-family cast. One reader cannot be cross-family with two families at once, so mixed-family parallel work becomes two sequential full routes. That is a real capability cost paid for an honesty guarantee.

Exit codes are the interface. 0 verified, 64 usage, 69 unproven routing, 70 worker failed, 75 waived, 127 missing dependency. Refusing with a distinct code rather than warning is deliberate — is it too strict to live with in CI?

If you think any of these is wrong, say so plainly and say what you would do instead. Evidence beats assertion here too.