Run each lens per model variant, scale to a granted time limit

[joahg]

Lets one review run each lens with more than one model, and with different providers, and makes gpt-6.1-sol usable through OpenAI's API.

Six-panelist panels. Config.PanelVariants runs every panelist lens once per named variant. With sol,opus, the review runs six panelists, for example security_contracts.sol and security_contracts.opus. Each panelist reports a name with its variant, such as "Security & Trust Boundaries (opus)". The coordinator prompt now lists the actual roster instead of hard-coding three panelists, and when variants are set it expects panelists that share a lens to report overlapping candidates. Without variants, the panel is the same three lenses as before.

Per-pass providers. Config.RoleProviders joins RoleModels and RoleEfforts, so the coordinator, the judge, and each panelist can use a different Goose provider. All three settings resolve from most to least specific: the pass role, then its variant, then its lens. For example, opus=anthropic covers all three Opus panelists, and security_contracts.opus=… overrides just one of them. The CLI adds --panel-variants and --role-providers, and the ReviewBench adapter adds RB_CONFIG_PANEL_VARIANTS and RB_CONFIG_ROLE_PROVIDERS.

Responses API for OpenAI. Goose 1.43 sends models it doesn't recognize, gpt-6.1-sol among them, to Chat Completions, which doesn't support tool calls for those models. When any pass uses the openai provider, the ReviewBench adapter sets OPENAI_BASE_PATH=v1/responses unless the caller set one. Against a local server, Goose then sent gpt-6.1-sol with its tools to /v1/responses instead of /v1/chat/completions.

Time limit. RB_CONFIG_TIME_LIMIT, in seconds and at least 900, names the per-PR limit granted in the ReviewBench manifest. Every stage scales from it. Without it, the budget is the 15-minute default as before.

Shorter coordinator wrap-up. A coordinator reconciling six panelists sometimes ran out of time, then ran out again while writing its long drop list in the wrap-up turn, which failed the pull request. On a time-up wrap-up, the coordinator now returns an empty dropped_candidates array.

Verification.

• A new test runs a six-panelist review with mixed providers. It checks that every lens runs once per variant, that overrides resolve in specificity order, and that the coordinator prompt lists all six panelists.
• go vet and go test pass. The default three-panelist coordinator prompt renders the same roster as before.
• Running review reviewbench end to end against a fake goose produced six panelists: three on openai/gpt-6.1-sol and three on anthropic/claude-opus-5-5. Every call had the Responses base path set.
• A new subtest checks that a coordinator's time-up wrap-up asks for an empty drop list. RB_CONFIG_TIME_LIMIT was checked by hand: unset gives 15m, 1800 gives 30m, and 600 or abc are rejected.