feat(agentd): introduce AgentdConfig to read env vars once at startup

[dijdzv]

Summary

test_pty_reader_drains_ready_fd fails intermittently (~10%) under cargo test in crates/agentd/. test_request_user_overrides_env_default sets MSB_USER=0:0 via set_var, which leaks into the parallel PTY test's fork() — the child picks it up, attempts a privilege switch, and _exit(1)s.

The root cause is that resolve_requested_user reads ENV_USER as a fallback when ExecRequest.user is None. But the SDK already resolves config.user into ExecRequest.user in build_exec_request(), so this env var fallback never fires in normal SDK/CLI usage — it's effectively dead code.

Changes

• session.rs: Remove ENV_USER fallback from resolve_requested_user, relying solely on ExecRequest.user
• session.rs: Remove test_request_user_overrides_env_default (the code path it tested was removed above)
• session.rs: Reduce PTY test sleep (1s → 0.1s) — original delay was unnecessary for verifying output ordering

Verification

Before: 3 failures in 30 runs
After: 0 failures in 80 runs

[appcypher]

You are right that ENV_USER seems unnecessary because we use SandboxConfig.user anyway. But ENV_USER serves as a defense-in-depth mechanism on the guest side.

ENV_USER is a sandbox-level property on the guest, not an exec-level one, and even if the host forgets to set it in their ExecRequest, the guest should always use their ENV_USER from sandbox creation.

As for the flaky PTY failures, I believe the root cause is reading/setting the same env vars in multiple unit tests, which is thread unsafe. So there is an opportunity for a proper fix here.

Maybe we should introduce AgentdConfig where the env vars are read into when agentd starts so we don't have to go through std::env::var every single time.

[dijdzv]

Replaced the ENV_USER removal with AgentdConfig — reads all MSB_* env vars once at startup. This preserves the ENV_USER defense-in-depth fallback while eliminating the repeated std::env::var calls that caused the flaky test.

[appcypher]

Just one more nit 😅.

It would be nice if all the env parsing operations are moved into this file as well.

So we parse all the env vars once in from_env and fields in AgentdConfig can be the actual types. Something like

pub(crate) struct AgentdConfig<'a> { 
    block_root: Option<BlockRootSpec<'a>>, 
    mounts: Option<VolumeMountSpec<'a>>, 
    ... 
}`