feat(runtime): support native self-hosted daemons across platforms

[SaladDay]

Behavior

Self-hosted Linux, macOS and Windows now run the same daemon preparation and execution flow. Core does not branch on the operating system or Environment source; managed Providers remain Linux-only. Harness differences stay in their adapters.

The daemon runs as its launching account without an inner sandbox or elevation. Managed isolation belongs to the outer Environment. Runtime installs user-directory dependencies; system dependencies must be preinstalled in the image/template or host. Session and Template packages.system input now fails explicitly, including null and empty values. Responses keep the pinned official SDK's required system: [] field.

This includes native installation/configuration, process and descendant cleanup, Files and artifacts, shared capability snapshots, reconnect/recovery, and current architecture/usage documentation. Retired Files helper binaries, configuration and build chains are removed. Native completion timestamps use one clock-skew rule for every Environment source. Windows supports Codex and Claude Code; the current MiniMax adapter rejects Windows. Historical version upgrades and cross-platform isolation are out of scope.

Validation

• Native GitHub Actions build and automated tests passed on Linux, macOS and Windows, including Codex/Claude startup without model calls: https://github.com/MiniMax-AI/parsar-core/actions/runs/36471973164
• Local macOS daemon lifecycle, Files and initialization race tests passed; MiniMax completed a real public Skill/MCP Turn and native cancellation tests. Codex native startup and tool effects passed; final post-fix model completion was not repeated successfully because the test provider quota was exhausted.
• Focused Linux daemon, adapter, API and store regressions passed. Fresh Linux Claude and MiniMax completed real Skill/MCP Turns; MiniMax public cancellation settled parent/child processes and recovered the same Session after daemon restart, preserving the capability snapshot.
• Three independent blind reviews completed; in-scope findings were fixed with targeted regression tests, including Codex and MiniMax descendant cleanup, required tool configuration, self-hosted variables, and Windows MCP PATH/npm resolution.
• Official SDK/PostgreSQL package response and input rejection regressions passed; independent PostgreSQL clock-skew/duplicate completion race tests passed.
• Final corrections passed complete Linux CLI/localworkspace race tests and MiniMax native worker cancellation regressions. Final Windows regression tests passed on native CI.
• Required make check passed on the combined installer integration at 19ac80f8, using zju for Core/adapter checks and mx for all 77 Web acceptance tests. The later 781a4dcd corrections passed complete CLI/localworkspace race tests and the final three-platform native CI. Duplicate full GitHub checks are still running; they are not claimed as passed. Evidence: /home/fanjingluo/.oac/acceptance/unified-native-installer-check-20260929/make-check-final.log.

The PR targets the current integration branch. A separate dependent installer task owns interactive Harness selection and native distribution packaging.