[gspark-etri]

Background

The four-model serving task needs native RNGD evidence separate from research-engine PD results. On the tested SDK2026.4 registry configurations, exact Qwen3-0.6B, Qwen3-4B and Gemma3-1B are rejected; KONI builds and serves but changes greedy output under concurrency. This PR records those limits, reproduction clients and controls. It does not implement a production engine fix.

Changes

Publish one experiment record, portable HTTP clients and fixture checks, a SHA256 source manifest and small evidence summaries. Retain failed measurements and distinguish execution, exact-token parity, strict answer quality, cache reuse and lifecycle behavior.

Evidence	Observed result
Native KONI build	105/105 kernel entries; FXB hash and frozen model revision recorded
Cache-off HTTP matrix	24 conditions, 344 responses; 63/168 token and 63/168 SSE-text mismatches against single-request references
Default/off/off/default profile-guided batching	344 additional responses; each arm has 29/42 token mismatches; disabling the flag does not fix drift
Log-probability controls	168 responses, 84 instrumented; output multisets unchanged by instrumentation; all 58 divergent instrumented responses expose competing probabilities at first shared-history divergence
Cache-off frozen quality	Sequential 31/36 answers, 34/36 formats; concurrent 32/36 answers, 35/36 formats; two repeats each
Cache-on matrix	336 responses, 24 successful resets; positive reuse hits in 24/24 conditions; cold-start/reuse output multisets equal in only 19/24
Cache-on frozen quality	144 responses; sequential 32/36 answers, 35/36 formats; concurrent 31/36 answers, 34/36 formats. First/reuse tokens equal 36/36 within each mode; cache-off comparison changes an 8K case's JSON compliance
Context boundary and recovery	Cache off/on, 30 requests: 6 valid edge probes accepted, 8 invalid probes HTTP400, 14/14 normal recoveries token-exact
Active stream disconnect	Cache off/on, 12 engine abort counter increments; 72/72 concurrent survivors and 12/12 recoveries token-exact; final running/waiting gauges zero

Verification

Evidence head: 052f84f; base main fd89265. One RNGD, TP8, context9216, furiosa-llm2026.4.0 b1baffa75, compiler2026.4.0 ae27d44cd8. Model/FXB hashes, exact commands, settings, timings, raw artifact hashes and cleanup records are documented in experiments/2026-10-25-native-rngd-coverage.md and experiments/evidence274/.

Final review checks rerun from the published source copies:

python3 experiments/evidence274/selfcheck_native201_http_profile.py
python3 experiments/evidence274/selfcheck_native201_benchmark.py
python3 experiments/evidence274/selfcheck_native201_quality.py
python3 experiments/evidence274/selfcheck_native274_logprobs.py
python3 experiments/evidence274/selfcheck_native274_quality_concurrent.py
python3 experiments/evidence274/selfcheck_native276_cache.py
python3 experiments/evidence274/selfcheck_native277_boundary.py
python3 experiments/evidence274/selfcheck_native278_disconnect.py

All eight checks pass. These are measurement-client fixtures, not substitutes for the hardware outcomes above. All27 source-manifest hashes and git diff --check verified. Changed-file sensitive scan clean; the full-repository scan retains seven findings in six unchanged baseline files and is not clean. Owned hardware experiments are stopped and Pod deletion verified.

Compatibility and limits

Evidence-only change; no engine behavior or default changes. Exact registry rejections are version/configuration scoped. Single timed batches do not establish stable or equivalent-output speedups. Frozen36-case synthetic quality is not a broad quality benchmark. Reported top-k probabilities do not identify a compiler/kernel root cause.

Lifecycle inputs repeat valid token1000. Streaming victims close at first nonempty content, observed at44/80generated token IDs, so this is mid-decode cancellation rather than first-token or prefill cancellation. Abort counters and idle gauges do not prove physical KV allocation reclamation. Earlier native batch/shape drift remains unresolved.

Related Issue

Related to #201, #274, #276, #277, #278, #261 and #263. Does not close the overall four-model task. Production export/cache fixes and their acceptance remain separately tracked in #271 and #273; transport implementation is in PR266.

AI-assisted work: prepared and verified with Codex; maintainer review and merge required.