Let agents go back, forward, reload and hard reload a preview tab

[Entrpi]

In the Browser panel, the user has Back, Forward, Refresh and Hard reload. Agents have none of them. preview_navigate takes only url or target, so the only route is preview_evaluate with location.reload() or history.back():

• It returns before the navigation starts, so the next snapshot can show the old page.
• It reports no error when there is no earlier or later page.
• Nothing an agent can call bypasses the cache. After a CSS change behind a long Cache-Control, the agent cannot load the new styles, while the user can with one click on Hard reload.

Agents do use that fallback. In one developer's threads, an agent called preview_evaluate with location.reload() 19 times over six days to refresh its tab. AGENTS.md, under "Hit every surface", says: "A capability a user can trigger is usually one an agent should reach through MCP tools".

Proposal: one optional action on preview_navigate, with the values back, forward, reload and hardReload. hardReload sends the same Page.reload {ignoreCache: true} as the user's Hard reload.

• Exactly one of url, target and action.
• The same ownership, human-control and dialog rules, readiness wait and result as URL navigation.
• A clear error when there is no history entry, for example "This tab has no earlier page to go back to."
• If you want it smaller, hardReload is the value to drop.

Cost: 203 more bytes in the MCP tool list, one Page.getNavigationHistory round trip per action, and no new status fields.

Evals, 3 samples: a new case asks the agent to load changed CSS that is cached for a long time. It went from 0/3 to 3/3 on Claude Sonnet 5.5 and GPT-6.1 Sol, and from 0/3 to 2/3 on Claude Opus 5.5, whose one miss reloaded through preview_evaluate. Nearby preview guard cases pass 3/3 on both sides on Sonnet.

Checked in a desktop dev build: a hard reload loaded the changed CSS, so the cache was really bypassed, and back and forward worked across origins.

Branch: main...Entrpi:t3code:fix/preview-navigate-history

Would you take this as a PR?