chore: remove concluded experiments and dead docker/sglang-glm5

[PierreLeGuen]

Removes 17 files (-13.5K lines) that no longer add value beyond git history. Every deletion was verified against the live fleet (/machines), prod configs, sibling repos (infra-docs / ansible / compose-manager), recent Slack, and open PRs before removal.

Deleted

June quantization sweep (6 files) — deepseek-v4-flash-int4-autoround-test, deepseek-v4-gemma4-int4-autoround-test, qwen36-gemma4-int4-autoround-test, qwen36-gemma4-int4-compressed-tensors-test, qwen36-gemma4-nvfp4-balanced, qwen36-nvfp4-x8-throughput. The sweep concluded: prod graduated Qwen3.6 to FP8+EAGLE, gemma-4 to FP8-block (prod comments cite ~1.9x over W4A16/QAT), DSV4-Flash to FP4-marlin TP2+EAGLE. None of their served-model-names/SNI domains exist in the live fleet; the ongoing DSV4 quantization campaign runs on tdx outside this repo.

Losing pack variants (2) — gemma4x2-qwen36x2, qwen35-qwen36-gemma4. The 06-01 three-way pack exploration was won by prod/dsv4-qwen36-gemma4.yaml (live on gpu07); the qwen35 co-location was resolved by the customer-driven prod/qwen35-dsv4-flash.yaml (live on gpu30).

OTel harnesses (2) — otel_test, glm51_otel_test. The monitoring label contract they were built to develop landed fleet-wide in v0.0.252 and is enforced by scripts/validate_otel_labels.rb.

Superseded GLM-5.1 configs (2) — GLM-5.1-AWQ-4bit (graduated: prod/GLM-5.1-SGL-AWQ-TP4.yaml runs the same checkpoint/revision/topology plus EAGLE and the loader patch, live on 2 hosts) and GLM-5.1-AWQ-vllm (lost to SGLang; can't run EAGLE). Note: the 4bit file at HEAD registers to the prod glm-5-1 domain despite its header claiming an isolated one — deploying it by accident would join the prod rotation without EAGLE. Deleting it removes a foot-gun.

gemma4-31b-qat-vllm — QAT path superseded by FP8-block in prod; the 06-22 validator commit naming it was mechanical CI coverage, not active work.

docker/sglang-glm5/ (4 files) — bake-weights-into-image experiment. build.sh has referenced a nonexistent Dockerfile.sglang-glm5 since its first commit (never runnable from a clean checkout); the nearaidev/sglang:glm5* images were rolled back from prod in April and nothing references them since.

Deliberately kept

• experiments/GLM-5.1-FP8-TP8-archived.yaml — README documents it as kept for reference, and it is the exact switch-back config for the open customer complaint (2026-07-15) about AWQ quality; infra-docs (inference.md, benchmarks/glm-5.1.md) reference it under its pre-rename name. It now carries a # STATUS: ARCHIVED header per the repo's own convention.
• cleanup-hf-model.yaml, test/isolation-*.yaml, all of prod/, scripts/, .github/ — documented utility, active fixtures, live configs, CI.
• VLLM_CACHE_ROOT=.../gemma4-31b-qat-vllm in two prod files is a cache-directory name on live hosts, not a file reference — left untouched to avoid invalidating warm caches.

Docs

README/AGENT.md experiments/ descriptions updated; the local validation loops gain an [ -f "$f" ] || continue guard (CI already has one) so a sparse experiments/ can't print false INVALID lines.

Notes for reviewers

• PR Decouple OTel log metadata from Datadog labels #142 (codex/decouple-otel-from-datadog-labels) mechanically relabels every yaml, including 11 deleted here — it will hit trivial delete/modify conflicts on rebase; resolution is to drop its changes to the deleted files.
• Open PRs experiments: DSV4-Flash EAGLE 3/1/4 + symm-mem two-shot patch (inline dockerfile) #133 and experiments: DeepSeek-V4-Flash W4A8 on vLLM (model-only, co-tenant-safe, inline dockerfile) #125 add new experiments/ files, so the directory stays in use.
• All four Ruby validators and the compose-syntax check pass locally on this branch.

[github-actions]

✅ OpenCodeReview: No comments generated. Looks good to me.