Retire obsolete lifecycle recovery and E2E machinery

[ericksoa]

Part of #11255. Focused delivery slice of #11765. Begin after #11695 merges.

Problem

Native-loopback qualification exposed live E2E coverage that still exercises NemoClaw-owned lifecycle and recovery behavior we intend to remove. In the exact-head #11695 run, the supported OpenShell stop/start path passed, while gateway-guard-recovery failed only in its later legacy sleep infinity container-replacement transaction. sandbox-survival failed before gateway restart because retained NemoClaw recovery records conflicted with current Docker identities. Those paths are validating obsolete machinery rather than the basic-onboarder architecture.

The authorization target also coupled its scope contract to a hosted agent turn: the required operator.write state settled successfully, but the job failed when the unrelated OpenClaw/inference turn timed out.

Evidence: https://github.com/NVIDIA/NemoClaw/actions/runs/34902351251 on #11695 head 7993b26a5dffb7980000b27af839532a5a5c0733.

Scope

• Delete the legacy keepalive/container-replacement branch from gateway-guard-recovery and remove production recovery helpers whose only remaining consumer is that branch.
• Refocus or retire gateway-guard-recovery so it covers only behavior NemoClaw still owns after onboarding.
• Replace NemoClaw retained-registry/recovery orchestration in sandbox-survival with OpenShell stop/start plus native agent readiness and state-survival assertions.
• Delete stale recovery records, immutable replacement-transaction state, rollback/handoff code, and fixtures that have no supported runtime consumer.
• Keep the scope-upgrade test focused on authorization. Do not make successful hosted inference the verdict for an already-proven scope transition unless that agent turn is itself the contract under test.
• Remove target-catalogue and workflow wiring for any live target that becomes empty.
• Do not build replacement lifecycle machinery.

Acceptance criteria

• The legacy sleep infinity replacement transaction and its assertions are gone.
• OpenShell is the lifecycle actor for supported sandbox stop/start; native agents own their gateway and background processes.
• sandbox-survival begins from clean state without NemoClaw retained-recovery identity conflicts and verifies only supported ownership boundaries.
• Authorization evidence distinguishes scope settlement from workload/inference availability.
• Obsolete production helpers and fixtures are deleted, with the removed inventory and net line reduction recorded in the PR.
• Existing credential-placement, exact-sandbox targeting, containment, and negative-egress evidence remains strict.
• Deterministic tests cover remaining parsing, migration, and security boundaries; live E2E is reserved for actual OpenShell/native-runtime behavior.
• The PR records the E2E target/job reduction.