Qwen3.8 RTX 5090 quant bakeoff friction log

Time	Category	Command or stage	Observation	Workaround	Durable follow-up
2026-09-03T18:37:23Z	repeated-command	control-plane discovery	Repository-specific controller tools were not exposed in this session.	Use the checkout-pinned local CLI for managed recipe lifecycle and retain exact envelopes.	Revisit controller tool availability independently of model qualification.
2026-09-03T18:51:29Z	unsafe-default	Gittensor SGLang no-speculation startup at 320,960	SGLang refused a context above the checkpoint-derived 262,144-token native limit unless SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 was set. The publisher's 320,960 figure describes maximum KV capacity, not a safe native model contract.	Retain the failed startup and set the managed recipe to 262,144; measure spare KV independently.	Keep model context and engine KV-pool capacity as separate fields in comparison docs.
2026-09-03T18:58:09Z	compatibility	Gittensor SGLang DSpark startup at advertised 165,169 tokens	The exact pinned target and drafter loaded and allocated 142,219 KV tokens each, but draft CUDA-graph capture failed with mat1 and mat2 shapes cannot be multiplied (14x5120 and 2560x248320).	Reject this exact target/drafter/SGLang v0.5.18 lane without a lower-memory retry; the failure is a hidden-width incompatibility, not OOM.	Revisit only with a publisher-confirmed mutually compatible target, drafter, and runtime revision.
2026-09-03T19:47:00Z	harness-control	cometkim NInfer no-speculation preflight	The recipe disables thinking globally at launch, while the first harness attempt also sent chat_template_kwargs.enable_thinking=false; NInfer correctly rejected the unsupported per-request option with HTTP 400. The accidentally continued timing calls failed identically and are retained as invalid-control artifacts.	Declare thinking_mode=unsupported to the harness so it sends no per-request template option; retain launch-level --no-thinking as the effective control.	Add endpoint capability discovery so a failed preflight cannot be followed by timing in the same shell sequence.
2026-09-03T20:30:00Z	quality	cometkim NInfer MTP3 bounded tool suite	Coding and triage checks passed 3/3, but every strict tool call omitted the required string zip argument.	Reject this arm as the general-purpose winner despite its 228.0 tok/s warm decode lead.	Revisit only with a parser/template change treated as a new configuration and rerun the complete gate.
2026-09-03T20:45:00Z	quality-risk	Gittensor SGLang FP8 KV startup	The runtime reported no calibrated FP8 KV scales and used 1.0 defaults. Functional and bounded quality checks passed, but broad fidelity equivalence is not established.	Keep the target-only lane as a measured TTFT challenger, not a promotion candidate.	Calibrate or validate FP8 KV scales and repeat broader quality/SWE gates.
2026-09-03T21:30:00Z	restoration	campaign teardown	All challenger recipes were unloaded and the exact starting Unsloth revision and llama.cpp image were restored. Fresh smoke, JSON, and 20/20 shared-prefix tools passed.	None.	Keep promotion and routing changes behind the separate human gate.

The campaign is complete; later follow-ups should append new dated rows rather than rewriting these observations.