feat(integrations/verl): swap adapter to arctic_platform.client (unified client)

[sfc-gh-kganesan]

Depends on / stacked on top of #50 (base branch is sfc-gh-kganesan/skyrl-verl-cortex-compat, not main) so the diff is only the adapter delta. Do not merge before #47 and #50.

Tracked by #49.

Why

arctic_platform/integrations/verl/adapter.py still imports from the pre-unification arctic_platform.rl package. Every call site through that module is async def even though the underlying HTTP/Ray calls are already synchronous, backend selection is hardcoded to backend="local", and there's no way to point the adapter at Cortex without forking it.

This PR does the minimal swap onto the unified arctic_platform.client shipped by #45–#47 + #50. Combined they give verl a serverless-capable backend switchable by a single yaml knob.

What

arctic_platform/integrations/verl/adapter.py

Import swap — arctic_platform.rl → arctic_platform.client:

• ArcticRLClientConfig, create_arctic_rl_client from the unified package.
• ArcticRLRayServerState moved to a TYPE_CHECKING import from arctic_platform.client.transports.onprem_ray, so a Cortex-only install (which never pulls ray in) still imports the adapter cleanly. Reconnect-path parameter typing preserved.

One backend knob, no adapter-side branching:

• _initialize_client reads backend from yaml (remote_backend.backend, default "onprem") and passes it straight through. Dispatch to RayTransport / HttpTransport / CortexTransport happens inside ArcticRLClient.__init__ via make_transport(config) — the adapter has zero knowledge of Cortex-specifics.
• Cortex sub-block (remote_backend.cortex.*) is whitelisted through into ArcticRLClientConfig: cortex_host, cortex_database, cortex_schema, cortex_endpoint, cortex_pat_env_var, cortex_base_url, max_seq_len. Any missing key falls back to CortexTransport defaults or CORTEX_* env vars.
• Dropped the hardcoded backend=\"local\" and log_prob_engine=\"deepspeed\". If a stale yaml still sets backend: local, the client's legacy alias (in feat(rl): Cortex-serverless dispatch via arctic_platform.rl (zero integration change) #50) warns and coerces to onprem.
• CUDA_VISIBLE_DEVICES widening (needed because Arctic's driver actor is scheduled with num_gpus=0) is now gated on backend == \"onprem\" — driver-side GPU placement is meaningless on Cortex.

Sync-under-async:

• All the underlying calls in arctic_platform.client are synchronous. Every await self._client.* becomes a plain call; every await self._send_* becomes a plain call.
• RemoteBackend's interface methods (compute_log_prob, update_actor, generate, save_checkpoint, update_weights, wake_up_inference, sleep_inference, reset_prefix_cache, destroy) stay async def — verl's RemoteBackendTrainer awaits them and swapping the interface would be a breaking change I don't want in this PR.
• Private _send_compute_ref_log_prob / _send_compute_log_prob / _send_update_actor become plain def; their bodies were already effectively synchronous once you strip the awaits.

What this PR does NOT touch

• Payload construction (_prepare_padded_arctic_batch_dict, _no_padding_2_padding_prompt_response, _create_ds_config, _create_ds_worker_config, _send_update_actor's left-pad handling) — byte-identical to main, so on-prem behavior is unchanged.
• Response shape handling (response[\"batch\"][\"log_probs\"], .pop(\"logprobs\"), response[\"metrics\"] aggregation) — same reads; the compat layer in feat(rl): Cortex-serverless dispatch via arctic_platform.rl (zero integration change) #50 makes the Cortex path emit the same shape.
• The RemoteBackend interface. No signature changes visible to verl.

Tests

Existing tests/client/* pass (69 pass; the compat test suite from #50 pins every call site the adapter makes against the client). No new adapter-level tests in this PR — the compat suite already covers the client-facing contract; an integration test lives on the E2E smoke below.

Test plan

• python -m pytest tests/client/ — 69 pass on this branch.
• Adapter syntax + interface preservation (AST check: 9 public methods stay async, 3 private helpers become sync).
• On-prem regression smoke — one Qwen3-0.6B GSM8K run against recipes/rl/skyrl/simple_gsm8k/run_qwen3_0.6b_gsm8k_grpo_arl.sh using remote_backend.backend: onprem. Loss curves vs main must match to within numerical noise.
• Cortex first-run — same recipe with remote_backend.backend: cortex + a populated remote_backend.cortex block, once the Cortex-side blockers in Cortex-serverless enablement for verl + SkyRL via the unified ArcticRLClient #49 §2 land (fwd-no-grad endpoint, log-probs endpoint, metrics block on fwd_bwd/step responses).

Follow-ups (see #49)

• Neutrino GS: POST /{job_id}/forward-no-grad, POST /{job_id}/log-probs.
• Cortex fwd_bwd / step responses grow a metrics: {grad_norm, ppo_kl, pg_clipfrac, pg_loss, kl_loss, kl_coef} block.
• Symmetric SkyRL adapter swap in Snowflake-AI-Research/arctic-skyrl (same import swap, same backend knob).

Made with Cursor

[sfc-gh-kganesan]

Closing this PR — after re-reading the merged upstream integrations (NovaSky-AI/SkyRL#1837, verl-project/verl#6422) I confirmed neither integration should have to change to gain Cortex support.

Both the merged SkyRL integrations/arctic_rl/ and the verl RemoteBackend implementation route through arctic_platform.rl.create_arctic_rl_client(config, server_state) — a single entry point on the Arctic-Platform side. The right move is to make that entry point dispatch to the unified client's Cortex transport when config.backend == "cortex", not to swap either adapter's imports.

Rewiring will land as new commits on #50; this branch will be deleted.