fix: skip load for pristine sessions

[lynnzc]

Summary

• skip session/load for pristine saved sessions that have no turns yet and only need a fresh ACP session
• keep existing resume behavior for sessions with history, explicit mode state, or config replay needs
• add a regression test covering the pristine-session fast path

Why

Issue #165 reports a large delay before the first streamed text event on codex prompt --session compared with codex exec.

For freshly ensured sessions with no prior turns, acpx was still attempting session/load before the first prompt. On Codex, that extra resume path appears to be the wrong fit for a pristine session and can delay the first streamed output. This change sends those pristine sessions straight to session/new instead.

Fixes #165.

Validation

• pnpm run build
• pnpm run build:test
• node --test dist-test/test/connect-load.test.js
• node --test dist-test/test/prompt-runner.test.js

[osolmaz]

Triage result

• Solves the right problem: 🛑 Localized, bad, or unclear fix
• Close PR: 🛑 Yes
• Recommendation: 🏁 close PR

Quick read

This PR targets the right symptom from #165, but the current solution is not safe enough to continue. The new "pristine session" heuristic relies only on locally persisted state, so it can misclassify a legitimately resumed ACP session as disposable and replace it with a fresh one.

Intent

The intent here is to avoid an unnecessary session/load for freshly ensured, still-unused sessions so the first streamed output is not delayed by the slow resume path.

Why

The problem is that the heuristic is too local.

• src/session-runtime/connect-load.ts now bypasses session/load whenever the saved record has no local messages, no lastPromptAt, no current_mode_id, and no config_options.
• src/session-runtime.ts creates records for sessions new --resume-session and sessions ensure --resume-session with exactly that same empty local shape.
• That means a real remote ACP session resumed via --resume-session can later be treated as "pristine" after the original process exits, even though the remote ACP session may still contain real state that should be loaded.
• The added tests cover the new fast path and warmed resumed sessions, but there is no regression test for prompting a session created via --resume-session after reconnect.

Codex review

Not run in this step. This was a read-only triage judgment.

CI/CD

Not run in this step. Validation and CI are intentionally deferred by the flow.

Recommendation

Close this PR. A follow-up should distinguish "fresh local wrapper around a brand-new ACP session" from "empty local record that points at a resumed remote ACP session", and it should add a regression test covering prompt/reconnect behavior for sessions created via --resume-session.

[osolmaz]

Closed by automated triage.