Problem

#2001 identifies a broader onboarding latency pattern: several readiness paths still rely on fixed poll counts, fixed intervals, or platform-specific timeout widening. PR #2492 addresses the gateway startup health loop, but the same waiting pattern still exists in other onboard/readiness flows.

These waits make fast systems pay unnecessary delay and make slow systems fail based on hardcoded loop assumptions rather than a clear deadline.

Scope

Extend the deadline-based wait pattern beyond gateway startup to the remaining readiness paths called out in #2001.

Candidate areas:

• Sandbox readiness polling
• Dashboard readiness polling
• Gateway recovery polling
• Agent gateway polling
• Sandbox create stream readiness polling, if it uses the same fixed-interval pattern
• Any error messages that hardcode stale timeout values like "within 60s"

Expected Behavior

Readiness checks should:

• Start with fast polling so successful systems continue quickly
• Back off up to a capped interval
• Respect one clear deadline budget
• Report the actual deadline used when timing out
• Preserve or deprecate existing health-poll env vars safely if they are externally used

Related Work

• Parent: perf: investigate and reduce networking latency during onboard and validation #2001
• Foundation PR: perf(onboard): add deadline-based gateway wait #2492

This issue should build on #2492 if it lands. If #2492 is superseded, this issue should reuse the replacement wait helper instead.

Acceptance Criteria

• Remaining onboard/readiness polling loops use deadline-based waits instead of fixed N x interval loops.
• Fast systems exit as soon as the readiness condition is met.
• Slow systems receive the full configured deadline budget.
• Existing tests that use NEMOCLAW_HEALTH_POLL_COUNT or NEMOCLAW_HEALTH_POLL_INTERVAL are updated or remain compatible through deprecated aliases.
• Timeout messages include the actual deadline used.
• New or updated tests cover immediate success, retry success, timeout, and zero/short-deadline test behavior.

Non-goals

• Adaptive provider-validation timeout calibration.
• DNS/TCP/TLS probe optimization.
• Onboard orchestration parallelization.
• Profiling trace output.