CorePulse v0.2.3

CorePulse v0.2.3 — session holdouts and lab diagnostics preview

Declare collection occasions prospectively with --session-id in baseline, pod-crash and dashboard startup. Declarations survive interrupted/final checkpoints and optional window metadata. Explicit split --group-by session keeps every member run in one training, held-out or purged group; an embargo purges the entire affected training session. Unknown/insufficient groups, lost class support and forged CSV assignments refuse. Schema-2 saved reports display grouping explicitly; legacy run mode remains schema 1 and its numerical feature allowlist is unchanged.

The synthetic no-cluster example exercises both grouping modes and retains wrong predictions/abstentions. Session names are metadata, not authenticated proof of independent network conditions. The original 22-run study predates session declarations; its missing AMF predictions and stronger binary packet-rule result remain visible. No new cross-session accuracy claim is made.

Add corepulse doctor: eight distinct read-only checks for ownership, context, workloads, packets, N2/N4, current-pod metric coverage and stable pod names. It creates no run, writes no telemetry, performs no repair and never releases a lock. Incomplete observations exit 2; a snapshot cannot authorize bypassing collection preflight. A real tagged baseline/AMF lifecycle test retained the declaration, correctly reported N2 loss despite 2/2 packets, and passed all doctor checks after the ordered reset. Single-session splitting correctly refused its separate exported windows.

Native Windows/macOS CI now accompanies Linux minimum-source-version checks. It exposed a macOS loader failure with the old Go 1.22 linker. Releases and the container use Go 1.27.3; macOS archives require macOS 13 Ventura or newer. Actual packaged Linux amd64, Windows amd64 and macOS arm64 binaries must pass version and both offline walkthrough modes before release/container publication. Linux arm64 and Darwin amd64 are cross-built without native execution checks.

Session guide: https://github.com/lazypanic999/CorePulse/blob/v0.2.3/docs/session-holdouts.md
Doctor: https://github.com/lazypanic999/CorePulse/blob/v0.2.3/docs/doctor.md
No-cluster walkthrough: https://github.com/lazypanic999/CorePulse/blob/v0.2.3/examples/offline/README.md
Measured progress and remaining limits: https://github.com/lazypanic999/CorePulse/blob/v0.2.3/docs/progress/2026-11-09.md

This remains an isolated-lab preview. Packet loss/latency/SCTP injection and live model serving are not implemented. Raw lab telemetry remains local. Verify downloaded archives with SHA256SUMS. Container: ghcr.io/lazypanic999/corepulse:v0.2.3.

Post-publication verification (9 November): SHA256SUMS matched the downloaded Linux/Windows amd64 and both Darwin archives. Windows/Linux packaged binaries ran the complete offline walkthrough in both modes; the Linux evaluator reproduced the original CSV and every fitted parameter, prediction and metric. Both Darwin archives contain LC_UUID and the documented macOS 13 minimum; the exact arm64 archive also passed native hosted execution. The network-disabled, read-only-rootfs container ran the session pipeline against invented inputs only and matched native results. Container digest: sha256:ace9bbb7f40cfe9d55aa0867f36b2368c44bb2882b2c1989f9c53b7c7b4a3b99. Govulncheck v1.9.0 reported no known vulnerabilities in the published CorePulse Windows executable. The running lab dashboard uses this release and its live doctor passed all eight gates after the reset.