refactor(matrix): select srt-slurm recipes with srt-recipe-dir and srt-recipe

[adibarra]

Summary

Multi-node master rows named their recipe through additional-settings: CONFIG_FILE=recipes/...yaml:override_x on the prefill or worker block. Recipe references are now first-class for every srt-slurm config, single- and multi-node:

dsv4-fp4-mi355x-sglang-disagg-agentic-umbp-dspark:
  srt-recipe-dir: dsv4/sglang/mi355x-fp4/agentx
  search-space:
  - conc-list: [4]
    srt-recipe: disagg-umbp-dspark.yaml:override_c4

• Schema: srt-recipe-dir per config key, relative to the single- or multi-node recipe root. Each row has srt-recipe: <file>[:selector] and optional eval-srt-recipe. Selectors are base, override_<name> or zip_override_<name>[i].
• Load-time checks: the file and selector must exist. They resolve under the loaded config's project root, not the tooling checkout. Klaud validates candidates against recipes fetched from the candidate head.
• Rejected: CONFIG_FILE / EVAL_CONFIG_FILE in additional-settings, except on llmd-vllm configs, whose scripts still read CONFIG_FILE.
• Generator: emits one canonical project-relative path (benchmarks/<single|multi>_node/srt-slurm-recipes/<dir>/<file>[:selector]) and reads node counts from it. The old CONFIG_FILE scan and the silent fallback for missing recipe files are gone.
• Workflow and launcher: benchmark-multinode-tmpl.yml exports SRT_RECIPE / EVAL_SRT_RECIPE. A new MultiNodeRequest reads them, and a single helper maps the canonical path to the recipes/...:selector path srtctl receives. Power-rule globs use the canonical form.
• Master configs: migrated mechanically. 459 CONFIG_FILE and 5 EVAL_CONFIG_FILE items became row fields, every config key got srt-recipe-dir, and single-node paths became relative. Also deleted dead settings: SYNTHETIC_ACCEPTANCE* (6 rows, unread) and SLURM_PARTITION (12 rows, overwritten from the cluster record).
• Recipes: deleted the stale CONFIG_FILE header comments in 30 bundles.
• Docs: updated RECIPES.md, CONFIGS.md, architecture.md, eval-agentx-procedures.md, nuke.md, add-model-hardware.md and the debug-runs skill.

Verification

• Matrix equivalence: full-sweep over both masters is identical to the base once old CONFIG_FILE / EVAL_CONFIG_FILE values are mapped to the new fields and the removed dead settings are dropped. That holds for the default (1767 rows) and --all-evals (1797), including node-count and exp-name.
• srtctl input: every multi-node row passes srtctl the same recipes/...:selector path as before. A new fake-Slurm test covers a bundle variant.
• New test: a master config in a temp project resolves a recipe that exists only there.
• Suite: pytest infx/tests passes all 2065 tests, and ruff is clean.

recipe-fingerprint changes once for every point, because the row shape changed. This is expected.

• dsr1-fp8-gb200-dynamo-sglang low-latency row: main pointed at disagg-tp8-stp-variants.yaml with no selector, so srtctl ran all three zip_override_lowlat variants at conc [4, 8, 16]. The launcher now binds exactly one variant, so the row is split by the variants' concurrency ranges: zip_override_lowlat[0] (c8) at conc [4, 8] and zip_override_lowlat[1] (c16-32) at conc [16]. Variant [2] (tuned for c64) is dropped, since the row runs no conc in its range.

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。