feat(ci): support additive append-only curve updates

.github/codeowner-signoff-verify-prompt.md

@@ -19,7 +19,7 @@ You are an automated merge-gate auditor for InferenceX.
 A CODEOWNER (`${SIGNOFF_AUTHOR}`) just posted the reviewer
 sign-off checklist (as a ${SIGNOFF_KIND}) that marks
 PR #${PR_NUMBER} as ready to merge. Your job is to
-INDEPENDENTLY verify the checks below (0-10). Do not trust the reviewer's checkmarks.
+INDEPENDENTLY verify the checks below (0-12). Do not trust the reviewer's checkmarks.
 Re-derive every conclusion from CODEOWNERS, CI runs, the PR diff, the master
 configs, and the linked recipe yourself. Be rigorous and specific. The checks encode
 the merge standard in `docs/PR_REVIEW_CHECKLIST.md`. Read it in the checked-out
@@ -359,8 +359,34 @@ Verify BOTH:
   unless the sign-off documents a sanctioned exception.
 - N/A if the PR has no agentic speculative-decoding changes (state that in one line).
 
+## Check 12 — Append-only changes only add new points to an unchanged curve
+APPLICABILITY: this check applies when any new `perf-changelog.yaml` entry contains
+`append-only: true`. If none does, report N/A.
+- Confirm every new changelog entry in the sweep is append-only; mixed regular and
+  append-only entries are not allowed.
+- Inspect the complete PR diff without using a file allowlist. Supporting code,
+  benchmark scripts, launchers, helpers, and other files may change. Their path alone
+  is never a reason to fail; determine whether each benchmark-affecting change is
+  behaviorally isolated to the appended points.
+- For every selected config, compare the generated matrix at the PR base and head.
+  Every base curve and concurrency must remain present, and every non-concurrency
+  field must be identical. In particular, require the exact same image, model,
+  framework, runner, topology, server arguments, scenario, duration, and offload
+  settings. The generated matrix may differ only by one or more newly added
+  concurrency values on existing curves.
+- Trace the selected config and generated runtime values through every affected file
+  into the changed behavior. The behavior must be reachable only for the corresponding
+  newly appended points. PASS when the controlling condition is uniquely satisfied by
+  those points. FAIL an unguarded/shared setup change, a condition also satisfied by an
+  existing point, or any case where exclusivity cannot be proven from the diff.
+- FAIL if an existing concurrency is rerun or removed, a new recipe/curve is created,
+  or a non-concurrency config setting changes. Other benchmark-affecting changes are
+  permitted only under the behavioral-isolation rule above.
+- Treat the repository's append-only matrix validation as supporting evidence, but
+  verify the diff independently and name the offending field/path when failing.
+
 ## Verdict and output
-Decide PASS only if Checks 0-11 ALL pass. A check reported as `N/A` counts as a pass.
+Decide PASS only if Checks 0-12 ALL pass. A check reported as `N/A` counts as a pass.
 Keep the `N/A — <reason>` row so the reviewer sees it was considered. Post EXACTLY ONE summary comment on
 PR #${PR_NUMBER} using `gh pr comment`. Start the comment with
 the hidden marker so reruns are identifiable:
@@ -386,7 +412,7 @@ single terse line. Rules:
   restating the checklist, no hedging ("if X then maybe Y"). Make the call. Link the
   run/recipe instead of describing it.
 
-- If everything is to standard: post the verdict header + the twelve one-line rows
+- If everything is to standard: post the verdict header + the thirteen one-line rows
 - If anything is NOT to standard: the verdict header must be immediately followed by a
   line that @-mentions the sign-off author as `@${SIGNOFF_AUTHOR}`
   with the blocking summary. Then the per-check lines, each failing one led by its root

.github/workflows/claude-pr-review.yml

@@ -134,6 +134,16 @@ jobs:
               - This is a 🔴 **BLOCKING** issue
               - Comment: "New `perf-changelog.yaml` entries must be appended to the END of the file. The file is read chronologically (oldest at top, newest at bottom), so inserting in the middle or prepending breaks the ordering. Please move the new entry(ies) to the bottom of the file."
 
+            ### Append-only Perf Changelog Safety:
+            When a new `perf-changelog.yaml` entry contains `append-only: true`, verify the complete PR diff before approving it:
+            - Do not use a file allowlist. Supporting code, benchmark scripts, launchers, helpers, and other files may change. Inspect the complete diff and judge whether each benchmark-affecting change is behaviorally isolated to the appended points.
+            - Every newly added changelog entry must contain `append-only: true`; append-only and regular entries may not be mixed.
+            - Selected existing curves may only gain concurrency values. The container image and every other generated recipe property (model, framework, topology, arguments, sequence lengths, duration, scenario, and so on) must remain unchanged.
+            - Benchmark or launch logic may change only when every changed behavior is on a control-flow path uniquely gated to the corresponding newly appended points. Trace the selected config and generated runtime values through every affected file into the condition. Confirm the path cannot be reached by any existing point. Unguarded/shared setup changes, or a branch also used by an existing concurrency/config/scenario, are blocking.
+            - Existing concurrency values, curves, configs, and scenarios may not be removed, replaced, or newly introduced.
+            - Eval modifiers (`evals-only`, `all-evals`, `eval-min-prefill-ep`) are not allowed.
+            If any condition fails, report a 🔴 **BLOCKING** issue. Never reject a change merely because of its file path; reject it when its benchmark effect is not exclusive to the appended points or the exclusivity cannot be proven from the diff.
+
             ## Terminology:
             - **STP (Single Token Prediction)**: Standard autoregressive decoding — one token per forward pass. No speculative decoding or MTP. Benchmarks labeled "STP only" use vanilla decoding.
             - **MTP (Multi-Token Prediction)**: Predicts multiple tokens per forward pass using speculative decoding (e.g., EAGLE, NEXTN).

.github/workflows/claude.yml

@@ -223,6 +223,8 @@ jobs:
 
             See `docs/configuration-procedures.md` → "Update an image" and "Append the changelog safely" for entry format and rules. Required whenever you change image tags, env vars, or perf-affecting params in `configs/*-master.yaml` or `benchmarks/*.sh`. Use `XXX` as the PR-link placeholder until the PR exists.
 
+            If an entry uses `append-only: true`, allow only added concurrency values on existing curves. Do not enforce a file allowlist: supporting code, benchmark scripts, launchers, helpers, and other files may change when their benchmark effect is exclusive to the corresponding newly appended points. The image and generated non-concurrency recipe fields must remain identical, no existing point/curve/config/scenario may be removed, all added changelog entries must be append-only, and eval modifiers are forbidden. Trace behavior through the complete diff; unguarded changes or paths reachable by an existing point are blocking.
+
             ## Spawning Additional Workers:
             You CAN spawn additional Claude workers by commenting "@claude" with a specific task.
 

CONTRIBUTING.md

@@ -44,6 +44,34 @@ A full benchmark sweep is expensive GPU time, and the runners are shared by ever
 - **This reduces CI queue time for everyone.** Each reused merge frees hours of GPU runner time for other PRs, so please prefer the reuse path over merging without it. A green sweep alone is not enough. The `/reuse-sweep-run` comment must be on record (the sign-off verification checks for it), otherwise `main` silently re-runs the full sweep.
 - `utils/merge_with_reuse.sh <pr-number>` is the supported merge path. It posts the command, syncs the branch with `main`, waits for checks, and squash-merges. See the [workflows README](.github/workflows/README.md#reusing-an-approved-pr-full-sweep) for eligibility details.
 
+## Adding points to the latest curve with `append-only`
+
+When a recipe and image are unchanged and a PR only adds concurrency values to an
+existing curve, mark every new changelog entry with `append-only: true`. Sweep setup
+compares the generated matrices at the base and head revisions, runs only the newly
+added points, and emits metadata that lets InferenceX-app extend the most recent
+matching curve instead of presenting the partial run as a separate curve.
+
+This mode is intentionally narrow, but it is not based on a file allowlist. Supporting
+code, benchmark scripts, launchers, and other files may change when their behavioral
+effect is exclusive to the newly appended points named by the changelog. No changed
+benchmark path may execute for or alter an existing point. Every selected config must
+already exist, its prior concurrency points must remain present, and all generated
+fields other than concurrency must be identical. Image changes or shared logic changes
+that can affect existing points are rejected. Append-only entries cannot be mixed with
+regular entries or eval-selection modifiers in the same sweep. The matrix validator
+enforces deterministic generated-config invariants; the human and AI reviewers must
+inspect the complete diff and verify behavioral isolation.
+
+```yaml
+- config-keys:
+    - dsv4-fp4-b300-vllm-mtp
+  description:
+    - "Add concurrency 192 to the existing curve"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  append-only: true
+```
+
 ## AMD cluster: never leave root-owned files in runner workspaces
 
 Multi-node benchmarks on the AMD MI355X TW cluster submit Slurm jobs whose containers often run as **root**. If those containers write files (typically `benchmark_logs/logs/slurm_job-*`) into the GitHub Actions runner workspace and the job is **cancelled** before teardown runs, the root-owned directories are stranded. The runner user cannot delete them, so `actions/checkout` fails with:

docs/PR_REVIEW_CHECKLIST.md

@@ -28,6 +28,7 @@
 - [ ] Verified that every single-node vLLM/SGLang recipe in this PR is documented in the official [vLLM recipes](https://recipes.vllm.ai/) and/or the [SGLang cookbook](https://docs.sglang.io/cookbook/intro):
   - [ ] I linked the corresponding upstream PR in the [vLLM recipe repo](https://github.com/vllm-project/recipes) or [SGLang repo](https://github.com/sgl-project/sglang/tree/main/docs_new) and verified that it is **MERGED** before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
 - [ ] Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at [docs/waiver/](https://github.com/SemiAnalysisAI/InferenceX/tree/main/docs/waiver)`<PR_NUMBER>.md` — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
+- [ ] If this PR uses `append-only: true`, verified that it only adds previously unmeasured concurrency points to an existing curve: the image and generated non-concurrency recipe/topology settings are unchanged, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it.
 - [ ] If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
 
 ### Additional detail section:
@@ -42,4 +43,3 @@
 
 
 <img width="569" height="632" alt="image" src="https://github.com/user-attachments/assets/491d9763-ab09-4734-b0f1-39eefe1ab5c4" />
-

utils/matrix_logic/validation.py

@@ -875,6 +875,14 @@ class ChangelogEntry(BaseModel):
     pr_link: str = Field(alias="pr-link")
     evals_only: bool = Field(alias="evals-only", default=False)
     all_evals: bool = Field(alias="all-evals", default=False)
+    append_only: bool = Field(
+        alias="append-only",
+        default=False,
+        description=(
+            "Run only concurrency points added to an otherwise unchanged existing "
+            "curve, then append them to that curve during dashboard ingestion"
+        ),
+    )
     eval_min_prefill_ep: Optional[int] = Field(
         alias="eval-min-prefill-ep", default=None, ge=1,
         description=(
@@ -887,6 +895,17 @@ class ChangelogEntry(BaseModel):
         description="Restrict to specific scenario types (e.g., ['fixed-seq-len', 'agentic-coding'])"
     )
 
+    @model_validator(mode="after")
+    def validate_append_only_mode(self):
+        """Append-only entries are throughput deltas, never eval-only requests."""
+        if self.append_only and (
+            self.evals_only or self.all_evals or self.eval_min_prefill_ep is not None
+        ):
+            raise ValueError(
+                "append-only cannot be combined with eval selection fields"
+            )
+        return self
+
 
 class ChangelogMetadata(BaseModel):
     """Pydantic model for validating changelog metadata structure."""