v0.7.1

• A run ends needs_login only when the page itself shows a sign-in to act on: a credential field, or a control that names one. A model that called a page with no such control a sign-in wall, or a page with nothing left to read, no longer makes the run claim the caller is signed out; it ends stuck instead.

• A verifier keeps the screenshot on a page whose state is a picture. An observation now reports a visible chart, image, drawing or video even where the page offers no control for it, so a text-rich page is no longer judged from its text alone. A control-less bot-check interstitial is reloaded once per address rather than once per document, so a challenge that survives its reload is reported as blocked instead of looping. And the run's own verifier caps its verdict at 400 output tokens, the bound the standalone llm_verify already applied.

• A benchmark row now carries the input and output token counts behind its model throughput, alongside the LLM and Jev call counts. They are what lets a comparison show speed and cost on a fixed, published basis, so a slow page or a provider's margin cannot move it. A row from a build that did not record them omits them and reads as not recorded, never as the measured wall clock or billed charge they replace.