Working on the per-hop telemetry piece of #25 + room test report

[aaryanmanchanda]

Network report from my run (Macbook Air M1 + iPhone 13 on Indian consumer broadband)

• Host: M1 MacBook Air, 8 GB, Chrome — local HTTPS build of current main
• Worker: iPhone, Safari, 0.5 GB pledge
• Same-WiFi, ~58 ms RTT (UI-reported), 30Mbps consumer broadband

Qwen3 0.6B Q8: Mac solo 36.6 → Mac + iPhone 17.5 tok/s (avg of 3)
Qwen3 1.7B Q8: Mac solo 20.6 → Mac + iPhone 11.4 tok/s (avg of 3)

Going to try running a room with 2 M1 MacBooks (Air+Pro) and bigger models.

In the meanwhile, I'd love to take the per-hop telemetry piece of #25

Proposed shape, following network-scheduler.md §2.4:

• append hops: [{peer, computeMs, encodeMs, bytes}] to ai-hiddenret / ai-hiddenret-b and the lap return — durations as performance.now() deltas (peer clocks aren't synced), bytes straight off the frame
• host derives transport = lapMs − Σ computeMs − host pack/compute, cross-checked against conns.get(next).rtt (already pinged every 2.5 s)
• surface perHopMs p50/p90 in the room UI and add a cross-network row to bench-log.md with the breakdown; keep it feeding roadmap 21's report schema
• unit test on the aggregation in tests/unit/

Anything in flight I'd be colliding with / any preference on where the aggregation lives? (room.js or room/pipeline.js)

ps: genuinely impressive that everything works over webRTC this seamlessly, thank u for open sourcing this

[Nehanth]

Sorry, this slipped through, and thanks for the numbers from India, that's really useful. Per-hop telemetry still isn't built, so if you're still up for it, yes please take it. Your shape looks good. The pipeline code is now shared between the browser and node hosts (room/pipeline.js, #293), so that's probably the place for the aggregation. Open a PR or an issue when you start so we don't overlap.

[Nehanth]

Hey @aaryanmanchanda, quick update since we didn't hear back: another contributor offered to take per-hop telemetry as part of the benchmark work in #21, so it's covered now. If you'd still like to help, a room test report from your setup (M1 MacBook Air + iPhone on Indian consumer broadband) would be really valuable for the device matrix there, since real cross-country network numbers are exactly what we're missing. Thanks again for the early numbers!