novelty analyze --estimate-only prices below the measured bill #631

Problem 2 of #624, left out of the fix for problem 1. The issue's run: 777 calls cost $6.00 under --max-usd 6, while --estimate-only printed expected_usd: 5.93 for expected_calls: 815, so a cap set to the estimate stops the run. The extractor role (335 of 815 expected calls) is the suspect: its output is long (a triage run billed 4 extractor calls at 16.4k output tokens, $0.38 of $0.95).

It needs measured runs per tier and role, which cannot be had offline, and estimate_usd and stops_at_cap are being moved to novelty/cost.py by the open PR that does that refactor, so editing them now would conflict. Fix options: refit the token profile per role against measured runs per tier and correct it, or report which role the estimate is least sure of next to expected_usd.

Found while fixing #624. The fix's PR is #633.