fix(tests): follow-ups from the Codex bounded-time tests review (#2169) #2174

Follow-up to #2169 (notes from the merge pass's review; none blocked).

• Length-only quadratic cost cancels out (omo MEDIUM, server/codex.test.ts): the reference and the subject now have the same length, so a mutant quadratic in length alone passes at 1.02 (the fix(tests): measure what the code controls in three tests that failed on busy hosts #2150 version caught it at 77.8 vs 40). The doc comment states this limit; consider a second, length-scaled reference.
• The large-answer case has ~2x headroom under heavy load (9.63 vs 20), with allocation and GC in its ratio.
• On a coarse clock a regression takes minutes to fail (the regex mutant ran >5 min); the retry loop has no upper bound.
• Two comments quote lower "as written" ratios than measured (~1.3x and ~4x vs 1.3-3.1 and 2.8-9.6); changelog line 7 is 115 columns.