Broker startup does not retry Relaycast workspace_busy admission 429

[khaliqgant]

Problem

Relay broker startup/admission treats Relaycast HTTP 429 workspace_busy as terminal after one attempt. That makes otherwise healthy RelayFlow/Fleet qualification runs fail while the workspace write-admission lane is briefly saturated.

Clean evidence

Relay PR #1665 proof run 34409605164 provisioned a fresh Daytona step sandbox, installed the exact Relay 11.10.4 broker, seeded the workspace, and then failed before the proof arm could start:

mcp-args --register failed (exit 1): registration for base-prover was rate-limited;
retry after 60s: Workspace write capacity is busy; retry with backoff
(code: workspace_busy; attempts: 1)

No product assertion ran, so this is orchestration-infrastructure failure rather than a feature failure.

Expected contract

• Retry only HTTP 429 carrying the exact workspace_busy code.
• Keep unrelated 429 responses terminal.
• Bound the retry count and total delay.
• Preserve status, code, message, and attempts on exhaustion.
• Do not replay the complete startup handshake after request-scoped retry exhaustion because it contains unkeyed workspace/agent-registration POSTs.
• Add an exact base/head RelayFlow proof case.

Scope

This is distinct from #1715, which covers worker /api/spawn preregistration and HTTP 503 database_overloaded. It is also complementary to AgentWorkforce/cloud#3507, which handles the Cloud executor direct mcp-args --register invocation.

Related upstream incident: AgentWorkforce/relaycast#389.