Reconstruct the validated v0.5.17 rollout stack

[nanjiangwill]

What changed

Reconstruct the validated parts of PR #131 on top of its complete revert in #135:

• pin the unified SGLang v0.5.17 runtime and K3 configuration
• update the weight-update profilers and verified measurements
• pin the rebased Miles runtime and adapt its cookbook recipes
• spread new sessions across rollout replicas with real configured headroom
• recognize a run-scoped engine's already-loaded base checkpoint as version zero
• document the final runtime contracts and measurements

Corrections from #131

• retain Modal's native Modal-Session-ID affinity header end to end
• do not introduce X-Session-Affinity
• preserve the established GRPO-group/session affinity fallback
• remove incremental pin-only and corrective commit churn from the reconstructed history
• remove a stale comment that described trajectory routing contrary to the actual GRPO-group behavior

The v0 initialization is independent of request affinity: a run-scoped rollout engine boots from that run's immutable base checkpoint before the trainer is required to publish latest. Recording that physical state allows base inference while CPU update-cache initialization proceeds; unscoped reconcilers still reject requests until they establish an applied version.

Validation

• 109 tests passed
• Ruff lint and format checks passed
• final SGLang pin: 0c79627e857eec795298a372adadf649209cdf2f
• final Miles pin: f83a68b296ca6f42d04e77e2b42c070c1ab02c70
• retained the successful GLM-4.7 five-step E2E and exact GLM-5.2/Kimi weight-update validation results documented in the repository