Adversarial suite: gate cases for a design flag with no reason, an eval case lost at the same score, and deletes-data words (follow-up to #957) #1018

Follow-up from #957 (adversarial suite cases, part of #707): its "Follow-ups (listed, not done)" and the second review issuecomment-6333173487 (round 3, head fd9893c).

The reviewer wrote one-line mutants of the built engine, applied in memory, and the adversarial suite stayed green under each (measured by the reviewer). These are gate cases (the suite's non-flow decide()) that would kill them.

• Q1, hold.doubt: an evaluator design flag with no reason. Add a gate case where the evaluator flags a design change with no design_reason (measured survivor, issuecomment-6333173487).
• Q2, Q3, hold.eval: an eval outcome dropped at the same score, visible and held-out. Every suite case with inputs.evals: dropped also lowers the score, so no case separates "a case went pass to fail" from "the score went down" (measured, issuecomment-6333173487). On main, dropped is also set when a case is lost at the same score (inferred from the reviewer's reading of skilleval.ts).
• Q11, hold.high-risk (deletes-data): the word list. Add a case with a TRUNCATE line (and DROP COLUMN / DROP SCHEMA) (measured survivor for TRUNCATE; the other two are the reviewer's suggestion, not run).

Fix by design (issuecomment-6333173487):

• Q2/Q3: let the suite's dropped input keep the score, so the case drives the dropped-outcome path alone.
• Q11: either add the cases, or record the deletes-data word list as a named limit, as is done for weakening.

After these land, #957's coverage count (at most 8 of 49 guarantees, with these named as open limits) can be recounted (inferred).