fix ralph codex srt sandbox

[almogdepaz]

summary

• make ralph agent handling shared/typed across server and worker
• replace stdout <subtasks> parsing with a structured .ralph-response.json control channel
• run codex with native --output-last-message / --output-schema and --disable apps
• keep runner-owned .ralph-* and progress.txt files out of agent commits via git excludes
• widen srt sandbox only where needed for codex state/network and git metadata writes from normal + linked worktrees

review

• differential review completed locally: no blocking findings

verification

• bun test tests/unit/ralph-agent-command.test.ts tests/unit/ralph-response.test.ts tests/unit/ralph-prompt.test.ts tests/unit/ralph-git-exclude.test.ts tests/unit/ralph-sandbox.test.ts tests/unit/plan-parsing.test.ts — 116 pass
• bun test — 1500 pass
• bun run typecheck — passed
• prior real codex+srt smoke: sandbox: srt, all_tasks_done: true, no codex session permission failure, no bogus subtasks appended

notes

• pushing over ssh failed locally because the ssh agent had no identities; branch was pushed over https using gh credentials. repo code unaffected.