feat(task): unify structured subagent execution

[vmcall]

Summary

• route direct task and eval agent() launches through one structured-subagent policy, artifact, isolation, and execution primitive
• add invocation-specific outputSchema / schemaMode support to flat and heterogeneous task batches with ordered all-settled fan-out
• allow eval agents in plan mode through a persisted host-enforced capability clamp that excludes mutation, nested spawn, MCP, extension, custom, memory, LSP, and IRC capabilities
• preserve eval handles/runtime helpers, task background lifecycle/progress, isolation recovery, schema precedence, and user tool-approval policy

Verification

• bun test focused subagent/schema/eval/isolation/persistence/tool-activation suites: 198 passed, 2 skipped, 0 failed across 23 files
• bun check in packages/coding-agent: passed
• independent final security/schema/lifecycle reviews completed; all confirmed blockers addressed

Fixes #5279

[vmcall]

CLANKER: I am not fixing the red Native: Linux x64 (baseline) check because it fails on cargo fmt --check for crates/pi-ast/src/ops.rs, which is not part of this PR's 31-file diff and was not modified by this branch. Adding an unrelated Rust formatting change would expand the PR beyond issue #5279 and obscure the structured-subagent review. The PR-specific check, lint/type/web-build, install smoke, and modern native jobs pass; the baseline job can be rerun after the upstream formatting issue is corrected.

[vmcall]

@codex u sleeping ??

[chatgpt-codex-connector]

To use Codex here, create a Codex account and connect to github.