Harden agent execution, recovery, and provider compatibility

[shiftynick]

Agent supervisors could crash on an early provider exit, leave descendants running after cancellation, and lose recovery metadata when a run failed. This change fixes those cases and keeps results usable across provider and package upgrades.

• Handle stdin errors, forward CLI stop signals, and finish process-tree cleanup after the provider leader exits.
• Preserve session/model metadata and provider diagnostics on incomplete runs; normalize Claude schema output and generate unique Claude worktree names.
• Support Codex native forks and persistent isolated worktrees, reject unsupported forks, and use the request environment for bounded rollout attribution.
• Bound stdout/stderr retention with configurable limits, stop explicitly on overflow, and reuse incrementally decoded events.
• Probe installed CLI help for optional capabilities and add cached, opt-out package update notices with structured results.
• Add Linux/Windows Node 20/22/24 CI, native process regression tests, versioned protocol fixtures, and documentation.

Validation: TypeScript, 134 unit tests, 18 Node package/process tests, npm package dry run, production dependency audit, and diff whitespace checks passed locally. Authenticated live provider calls remain opt-in. All six CI jobs passed on Linux and Windows with Node 20, 22, and 24. Windows also executes the environment-casing and .cmd regressions; POSIX-only signal tests are skipped there.

Compatibility notes: stdout/stderr limits now default to 16 MiB/1 MiB; overflow produces a failed partial result. CLI update checks can be disabled with AGENT_HEADLESS_NO_UPDATE_CHECK=1; library checks are opt-in. Codex isolated runs require persistence so checkout metadata can be recovered. This PR does not publish an npm release.