Add pre-session billing and cost notice - #436

An interview runs on Gemini Live, which bills the Google AI project the interview runs on for the whole context of the session: retained audio and images are paid for again on every later turn (docs/provider-cost-and-degradation.md, "What an interview spends"). The final report, the quiet-pause interim reviews and the phase judge run on the same report model and keys, so they are billed to the same project. Neither the operator entering a key nor the candidate pressing Start was told so. This adds the pre-session notice #161 asks for; its session length cap and text-only mode stay open there.

Which project pays is the interview's, not always the server's: agent_config_for swaps in a LiveKit provider's own Google keys when it has any, so its rooms bill its project (docs/providers.md, "Gemini credential failover"). The wording holds for both. The Setup page says, in its Google AI section, that Gemini Live is billed to the project of the key an interview runs on, this key unless the interview's LiveKit provider sets Google keys of its own, and that the reports, interim reviews and phase judge run on the same keys and are billed to the same project.

The lobby shows a notice above Start, linked through aria-describedby, that says Gemini Live usage is billed on the whole context of the session to the Google AI project this interview runs on, and names what the chosen mode sends besides the text of the interview: voice and video for a coding interview with candidate video enabled, voice without it, and voice and board images at a whiteboard, where the camera is off but every board reaches Gemini as a realtime image on the same stream ("What a whiteboard interview adds"). In a coding interview with example drawing checked it also names the drawing, sent whenever Jim reads it: each read_board call sends the current example JPEG, which then stays in the billed context. Switching mode or toggling example drawing updates it.

The notice lists media against "the text of the interview" rather than claiming to be everything sent, because which text reaches the model depends on the mode, the loop and the candidate's consent: the code in a coding interview, which reaches the model as text and is resent on a reconnect, and, outside a coding-only session, any resume or job description snippets the candidate selected and agreed to send, built into the instructions of every generation. The doc paragraph in "What the candidate sees" spells that text out.

/api/session gains candidateVideo, answered by the room dispatcher: LocalDispatcher reports its agent config's setting, and any other dispatcher, or none, reports null. A web-only host cannot see the camera setting of an agent that joins from elsewhere, so it promises nothing, and the lobby keeps the notice hidden while the value is null, before the response arrives, and in offline practice when the request fails or a later refresh does.

The longest notice wraps to seven lines in the Start column, which put the 1280x800 lobby at 804px with taller font metrics, so the laptop media query sets the notice smaller and tighter. A lobby test of its own measures the page with that notice showing under the same three fonts and long title as the existing one-screen test, which is unchanged and runs with the notice hidden; the test fails at 804px without the rule.

Browser tests cover each mode, the switch between them, toggling and restoring example drawing, a null or missing value, a failed request and a failed refresh, and wait for the account status that /api/session sets rather than for the initial "Checking account..." text; the Setup page test checks both billing sentences. tests/web/accounts.rs checks candidateVideo as true, false and null in both the signed-out and the signed-in response, with null both from a dispatcher that does not know and from no dispatcher at all; a unit test covers LocalDispatcher reporting its config.

UI Verification

Setup page (Google AI section):

Lobby — Coding mode (candidate video disabled / default):

Lobby — Coding mode (candidate video enabled):

Lobby — Coding mode (candidate video enabled + example drawing checked):

Lobby — Whiteboard mode: