feat(evaluator): run sequential judges in the shared case runtime

[zpzjzj]

Multi-judge evaluations previously required filesystem snapshot/fork support. Run the evaluated Agent once, then execute named judges sequentially in its existing runtime, following shared-verifier semantics. Each Agent judge starts a fresh session; verdicts, timeouts and archived materials remain separate by judge ID.

Filesystem changes, installed tools, services, volumes, environment and network policy remain shared. A completed judge FAIL or ERROR does not prevent later judges from running; case cancellation skips pending members. Remove the obsolete judge snapshot capability and workspace-copy implementation, retaining the normal case cleanup policy.

Stacked on #314 to preserve independent judge engine validation and provisioning. Part of #308; filesystem-isolation-specific acceptance remains outside this shared-runtime change, so #308 stays open.

Validation:

• make fmt, make verify, make test with race detection, make build, and make e2e.
• Regression coverage for one Agent run, fresh sessions without cross-member resume, shared mutations, independent engines, gate/timeout/cancellation outcomes and namespaced artifact readback.
• Real Docker shared-runtime tests: all members PASS, and script FAIL / both Agent judges PASS / aggregate FAIL.
• Live OpenSandbox shared-runtime tests in stacked test(runtime): verify sequential judges in a shared OpenSandbox #316 exercise the same runtime path. The shared-environment fixture is deterministic; an additional test(runtime): verify sequential judges in a shared OpenSandbox #316 test runs the real OpenCode CLI and makes live model-provider requests.

The stack includes #314's independent OpenCode provider initialization fix and its endpoint/API-key/MCP preservation regressions.

Default-endpoint follow-up:

• Independent OpenCode judges reset only their selected provider's inherited inline configuration, including endpoint and API-key/SDK/model settings; case MCP, unrelated providers, plugins and permissions remain intact. Normal runner and legacy judge initialization retain their existing behavior.
• Regression uses the real resolver and adapter factory. Both OpenAI and Anthropic omitted-base_url cases reproduce on the previous evaluator and pass after the fix; explicit-endpoint controls pass as well.
• Real OpenSandbox/OpenCode 1.14.24 default-configuration installation, preflight and readback passed. The real qwen3.8-max explicit-endpoint control passed with two distinct judge sessions.
• Native OpenAI model calls were not completed: direct api.openai.com connectivity timed out from the test sandbox, and both model invocations hit their deadlines. No native OpenAI model success is claimed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T07:44:55.544167Z	3796e6e	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.