Run remote-bench end-to-end via GitHub Actions and design the InferenceX-app ingest trigger

[aistackdev]

Summary

Follow-up to #26/#27 (remote-bench recipe support). Once #27 merges, the remaining gap is: remote-bench.yml produces a real, ingest-shaped agg_*.json artifact, but nothing currently pushes that artifact into InferenceX-app. Local sweeps get there via run-sweep.yml's trigger-agentic-ingest job (a repository_dispatch call with event_type: ingest-agentic-results) — remote-bench.yml is a standalone workflow_dispatch entrypoint that never runs through run-sweep.yml, so that trigger never fires for a remote-bench run today.

Scope

1. Run the full flow for real — dispatch remote-bench.yml via GitHub Actions (not the manual SSH debug loop used for feat: support benchmarking remote/existing inference endpoints (BYO endpoint) #27's smoke test) against a live endpoint, end to end, once merged to vng-benchmark/main.
2. Design the ingest trigger for remote-bench results. Options to weigh (not deciding yet, that's this issue):
   • Add a trigger-agentic-ingest-equivalent job directly to remote-bench.yml, firing the same repository_dispatch on success.
   • Keep remote-bench results out of the automatic ingest path entirely for now (manual/ad-hoc dispatch implies non-canonical/exploratory runs) and revisit once there's real demand for them to show up in the dashboard alongside official sweep results.
3. Known ingest-side landmines to design around, found during feat: support benchmarking remote/existing inference endpoints (BYO endpoint) #27:
   • REMOTE_RUNNER_TYPE must resolve via InferenceX-app's hwToGpuKey() — fine for real datacenter GPUs (h200-nv etc.), but the dev endpoint we smoke-tested against (RTX 5090) isn't in GPU_KEYS at all, so that specific result wouldn't resolve a GPU key on ingest. Not a blocker for real targets, but means we can't fully validate the ingest path against that same dev box.
   • image must be the real deployed image (fixed in feat: support benchmarking remote/existing inference endpoints (BYO endpoint) #27) — otherwise ingest records a meaningless placeholder.
   • Duration matters for canonical-eligibility: --unsafe-override kicks in under 900s (as our smoke tests all were), which the aggregator likely flags as non-canonical the same way it does for local recipes — confirm whether InferenceX-app's ingest treats that the same way for remote-bench rows.

Not in scope

No InferenceX-app schema/mapper changes assumed yet — this issue is scoped to deciding whether/how the trigger should exist, not implementing it. If a mapper change turns out to be needed, that becomes a follow-up PR against SemiAnalysisAI/InferenceX-app, coordinated separately.