main-promotion gate: verify CUDA linux/amd64 Carl path on BigMama/WSL2

[joelteply]

airc-queue card

Coordinates work via the AIRC queue substrate (airc#562). Edit this card by commenting OR by running airc queue claim/airc queue release/airc queue heartbeat (later PRs).

{
  "kind": "airc-queue-card-v1",
  "id": "#1397",
  "owner": "codex-main",
  "status": "blocked",
  "blockers": "BigMama SSH/Tailscale unreachable from this Mac",
  "env": "rtx5090-wsl2",
  "evidence": "SSH to bigmama and bigmama.taila5cb68.ts.net timed out using ~/.ssh/continuum_m5_bigmama; tailscale command unavailable locally.",
  "next_action": "Run EXPECTED_SHA=a89c8ab471e7de415fe8faccd2f0bb98fc8e3544 scripts/push-current-arch.sh on reachable BigMama/WSL2 or equivalent linux/amd64 NVIDIA host.",
  "last_heartbeat": "2026-05-18T17:52Z @ a89c8ab47"
}

Close this issue when the work is done (status=merged/abandoned).

Original issue body

Pre-adoption body

Problem

Continuum cannot launch without the common GPU path: CUDA on linux/amd64 / Windows WSL2. Mac Docker Desktop cannot prove or publish this path honestly. BigMama/WSL2 or another linux/amd64 NVIDIA host must run the native image gate.

Current evidence

• PR feat(cognition,#1385): admit generate-response through Rust resource gate #1396 merged to canary at a89c8ab4.
• Mac pre-push proved TS/ESLint/Rust compile/Rust tests, but native image publish did not complete because Linux container GPU was unavailable on Mac Docker Desktop.
• SSH from this Mac to bigmama and bigmama.taila5cb68.ts.net with ~/.ssh/continuum_m5_bigmama times out.
• tailscale is not available in this Mac shell, so I cannot currently reach the Tailscale host from here.

Required command on capable host

On BigMama/WSL2 or equivalent linux/amd64 NVIDIA host:

cd /path/to/continuum
git fetch origin canary
git checkout canary
git reset --hard origin/canary
git submodule update --init --recursive src/workers/vendor/llama.cpp src/workers/vendor/whisper.cpp
nvidia-smi
docker version
EXPECTED_SHA=a89c8ab471e7de415fe8faccd2f0bb98fc8e3544 scripts/push-current-arch.sh

Expected behavior:

• Host platform is linux/amd64.
• nvidia-smi is present, so scripts/push-current-arch.sh includes cuda.
• Publishes continuum-core-cuda for linux/amd64 at canary/head SHA.
• Publishes/refreshes linux/amd64 core, vulkan, livekit-bridge as applicable.
• Does not weaken the no-CPU-fallback rule.

Acceptance

• GHCR has current canary SHA images for CUDA linux/amd64.
• Carl install/chat path on GPU host succeeds with real CUDA GPU, not llvmpipe/no-GPU advisory mode.
• verify-architectures / image revision checks pass for the CUDA/common path.

Status log

• 2026-05-18T17:52Z — heartbeat by codex-main -> status=blocked (Corrected scope: CUDA linux/amd64 is a canary->main promotion gate, not a canary iteration blocker. Mac work continues with local Rust/Metal proof.)