fix(observability): harden OTLP collector delivery

[afourniernv]

Summary

Make the sandbox trace relay tolerate a collector that is unavailable at gateway startup or restarts later, and surface collector-side partial acceptance instead of treating every successful gRPC call as full delivery. This is the fifth draft in the split OTLP relay stack.

Related Issue

Extends NVIDIA/OpenShell#2641 and the work in NVIDIA/OpenShell#3196. This draft does not close either item.

Changes

• Configure the collector channel lazily so collector availability does not gate gateway startup
• Reconnect through the tonic channel when a collector starts or returns after an outage
• Keep bounded in-flight export admission from the preceding stack layer
• Distinguish full success from OTLP partial_success
• Log rejected-span counts and collector diagnostics
• Bound the collector-recovery test probe with per-attempt and total deadlines
• Document startup, reconnection, timeout, and partial-success behavior

Testing

• mise run pre-commit passes
• cargo test -q -p openshell-server otel_relay
• cargo test -q -p openshell-server supervisor_otel_export
• E2E tests added/updated (not applicable to this isolated delivery layer)

Checklist

• Follows Conventional Commits
• Commits are signed off (DCO)
• Documentation updated

Stack

• Foundation: rhuss/OpenShell#21
• Explicit relay opt-in: afourniernv/OpenShell#2
• Trusted metadata: afourniernv/OpenShell#3
• Resource bounds: afourniernv/OpenShell#4
• This draft is based on codex/3196-stack-resource-bounds