[EagleClaw89]

Why

Operators can see how hard a Spark is working (aggregate tok/s), but not how fast a conversation feels. Aggregate is throughput; per-stream is what a human experiences. This adds the missing half for TensorFold, measured from the engine's own counters.

What

• perStreamTps — 1 / mean seconds per output token over the last window, from tensorfold:request_time_per_output_token_seconds. The histogram moves only on request completion, so the value is held between completions (never flashes 0), falls back to aggregate ÷ active streams, and stays null when nothing has been measured. Same design vocabulary as the existing Strata "active-slot mean", for TensorFold.
• perStreamPrefillTps — effective prefill rate per request. Verified live: during a 35.4 s / 98,854-token prefill nothing in /metrics moves until the completion poll (prompt_tokens_total, the prefill histograms — all frozen; only requests_running changes). The rate is therefore derived as Δprompt ÷ Δprefill-seconds at completion and held, with a live "prefilling" indicator from requests_running > 0 && generation_tokens_running == 0 (verified: generation_tokens_running stays 0 across the whole prefill, > 0 during decode). prompt_tokens_total includes prefix-cache hits and /metrics publishes no cached counter, so the UI labels it "effective (incl. cached tokens)" — a cache-served prompt correctly reads ~140k tok/s while cold prefill on the same box measured 2,803 tok/s (ground truth 98,854 ÷ 35.14 s).
• Running-count alias rule — TensorFold exports both requests_running and num_requests_running with identical values (verified across whole requests). Read the canonical name; use the alias only when the canonical series is absent; never sum. A test pins a single stream reporting 1, not 2.

Sanity gate + hold expiry on the prefill rate. Small prompts (the 1–200-token showcase/bench kind) and prefix-cache servings complete with sub-millisecond prefill times, so their Δprompt/Δprefill-seconds ratio is timer noise: observed live at 109k–313k tok/s against a verified cold-prefill ground truth of 2,803 tok/s (98,854 tokens ÷ 35.14 s, reproduced twice). Windows are therefore accepted only when Δprefill-seconds ≥ 50 ms, ≥ 50 prompt tokens, and an implied rate ≤ 15,000 tok/s (measured cold prefill on a DGX Spark is ~2.8–3.0k tok/s, so 15k is a 5x margin that keeps cache-credited agent windows — observed at 44k–313k — off the tile); rejected windows hold the last valid rate, and the hold expires after 20 s — the same staleness window the existing /health path already uses for its prefill estimate.

Exposed on the Prometheus endpoint as sparkdash_llm_per_stream_tokens_per_second and sparkdash_llm_per_stream_prefill_tokens_per_second. Both UI lines omit themselves when the engine reports nothing, so non-TensorFold cards are unchanged.

Tests

• 9 new probe cases (TPOT window, hold-when-frozen, null when absent, queueOnly isolation, counter-reset re-seed, alias never-summed, prefill completion-window, prefill hold + prefillActive heuristic) + 2 UI cases.
• Green: node --test server/collectors/__tests__/LlmProbe.*.test.js (197 pass), vitest for touched pages, npm run typecheck, npm run build.
• Live on DGX Spark GB10 / TensorFold / Qwen3.8-Flash-Next: single 98,904-token request → requestsRunning: 1, prefillActive: true for the full 35 s (tile says "prefilling"), then prefillTps: 2803.6 + perStreamPrefillTps: 2803.6 held after completion; 2-stream bench → requestsRunning: 2, generationTps: 132.5, perStreamTps: 76.48.

[MiaAI-Lab]

now on main as fa16c54, with the UI in d37dcc4. TensorFold completion-window counters feed per-stream decode and effective prefill fields, retaining noise guards and prefill expiry (server/collectors/LlmProbe.js:1856, :1917). The GB10-specific upper ceiling was removed because effective throughput includes cache credit; a 30,000 tok/s regression failed with the ceiling and now passes. Both commits passed typecheck, build and the full suite, with authorship and changelog credit preserved. (Review notes prepared with Claude Code on behalf of the maintainer.)