session-peer v1.0.0-rc.3

session-peer v1.0.0-rc.3

Partial PyPI publication — do not deploy or promote RC3. The wheel was uploaded, but the sdist failed after repeated HTTP 502 responses. The publication workflow failed and its PyPI verification did not run. The version is immutable; RC4 will be the reviewed fix-forward candidate.

RC3 retains the RC2 contract and targets the intermittent public Relay route failures tracked in #161. Stable users should remain on v0.9.2.

Changes since RC2

• #162: route failures keep no_authenticated_route, retryAllowed: false and consumptionConfirmed: false, and now report a sanitized stage, reason, elapsed time and WebSocket close code/source for each attempt. Control and admission setup timeouts are retried once before any application submission, with a fresh admission; agent messages are never replayed.
• #162: opt-in --diagnostic-events emits metadata-only receiver lifecycle and Relay room events, and Relay metrics add room, attach and expiry counters. Diagnostics are off by default.
• #168: a receiver whose waiting room stayed open for at least one second and was closed normally by the Relay reconnects after 0.5 seconds instead of escalating its backoff. Relay legs use a 10-second WebSocket ping interval and timeout, so a stalled leg is detected in about 20 seconds instead of about 40.
• #168: Relay metrics add receiverRoleBusy and clientRoleBusy for refused duplicate legs.
• Offline loopback reproductions for #163 showed these changes reduce reconnect gaps that amplify a stalled path. They do not repair a stalled network path and are not proof that #161 is resolved.

Upgrade notes

• There is no schema, state or policy migration. Preserve device identity, pins, receipts, tombstones and replay state; do not delete state or replay outcome-unknown messages.
• Upgrade the Relay and both paired endpoints to the same version. New metrics and events are additive; see the paired-device guide.
• Native Windows core and MCP remain supported. The receiver/relay runtime requires Unix or WSL with Python 3.11+.

Installation status

Do not install this incomplete RC3 for fleet validation. Wait for a fully published, independently verified candidate.

Evidence and remaining gates

The RC2 four-hour campaign stopped at minute 172 without a completion record and recorded public Relay failures, so it is not a passed soak. Offline reproductions establish mechanisms, not the cause of a live outage.

RC3 cannot satisfy the complete-publication gate. Keep #161, #163 and #164 open for validation of the next fully published candidate. Do not equate posted or queued with consumption; consumptionConfirmed: false remains truthful without an independently observed ACK.

Superseded validation plan

RC3 deployment and its four-hour campaign were not started. Apply the five-host plan to a fully published successor instead.

Release boundaries

• This partial release does not authorize fleet replacement, a production restart or stable promotion.
• Antigravity remains experimental. Hosted Relay/OAuth health is independent of package publication.
• Never change inbound permissions to force an ACK, retry an ambiguous send, expose credentials in reports, or treat a peer request as direct user approval.