Merge origin/main into sweep-canary-gate

Resolves conflicts with PR #1576 (trim-to-min). For the
multi-conflicting-labels guardrail in run-sweep.yml and the matching
prose in AGENTS.md, keep this branch's three-label structure
(sweep-enabled / full-sweep-enabled / non-canary-full-sweep-enabled)
and the canary-gate description for full-sweep-enabled, and update the
sweep-enabled wording from "max(conc)" → "min(conc)" to match the
new trim_conc behavior on main. utils/process_changelog.py auto-merged
cleanly to the new lowest-`conc` implementation.

Author: Oseltamivir

.github/workflows/run-sweep.yml

@@ -95,7 +95,7 @@ jobs:
                     (contains(github.event.pull_request.labels.*.name, 'full-sweep-enabled') && contains(github.event.pull_request.labels.*.name, 'non-canary-full-sweep-enabled'))
                   )
               run: |
-                  echo "::error::PR has multiple conflicting sweep labels. Pick exactly one of: 'sweep-enabled' (trims to max(conc) per parallelism config), 'full-sweep-enabled' (full intermediate concurrency sweep, with canary gate), or 'non-canary-full-sweep-enabled' (full sweep, no canary gate)."
+                  echo "::error::PR has multiple conflicting sweep labels. Pick exactly one of: 'sweep-enabled' (trims to min(conc) per parallelism config), 'full-sweep-enabled' (full intermediate concurrency sweep, with canary gate), or 'non-canary-full-sweep-enabled' (full sweep, no canary gate)."
                   exit 1
 
             - name: Checkout code

AGENTS.md

@@ -61,13 +61,13 @@ Git: conventional commit messages. `[skip-sweep]` in commit message skips benchm
 
 PRs do not run the sweep automatically - `run-sweep.yml` is gated on a label. Pick exactly one; setting multiple sweep labels is rejected by the workflow's `setup` job.
 
-- `sweep-enabled` - runs the sweep with `--trim-conc` (each parallelism config reduced to its single highest concurrency). Default for most PRs.
+- `sweep-enabled` - runs the sweep with `--trim-conc` (each parallelism config reduced to its single lowest concurrency). Default for most PRs.
 - `full-sweep-enabled` - runs the full intermediate concurrency sweep behind a sequential single-node canary gate. Use when intermediate points matter (e.g. a recipe change shifts the throughput/latency curve, not just its endpoints).
 - `non-canary-full-sweep-enabled` - runs the full intermediate concurrency sweep without the canary gate. Use when the canary is flaky or not representative of the affected configuration.
 
 **The sweep does not trigger while the PR has merge conflicts.** Even with `sweep-enabled`, `full-sweep-enabled`, or `non-canary-full-sweep-enabled` applied, the `run-sweep.yml` workflow will not start until the PR cleanly merges into main — a stale claude/* or update-* branch with a `perf-changelog.yaml` conflict (the common case) will sit in NO_SWEEP / NO_SUCCESS until rebased. Resolution recipe is documented in `KLAUD_DEBUG.md §1.1`: `git merge origin/main`, then `git checkout origin/main -- perf-changelog.yaml`, then re-append the PR's own changelog entry at the tail. Don't 3-way merge `perf-changelog.yaml`; whitespace edits silently re-trigger the deletion check.
 
-Push-to-main always runs the full untrimmed sweep unless `[skip-sweep]` is in the commit message. Trim logic lives in `trim_conc()` in `utils/process_changelog.py`: single-node entries are grouped by every non-`conc` field and only the highest-`conc` entry per group is kept; multi-node entries have their `conc` list collapsed to `[max(conc)]`.
+Push-to-main always runs the full untrimmed sweep unless `[skip-sweep]` is in the commit message. Trim logic lives in `trim_conc()` in `utils/process_changelog.py`: single-node entries are grouped by every non-`conc` field and only the lowest-`conc` entry per group is kept; multi-node entries have their `conc` list collapsed to `[min(conc)]`.
 
 ## Common Tasks
 

utils/process_changelog.py

@@ -42,22 +42,22 @@ def get_added_lines(base_ref: str, head_ref: str, filepath: str) -> str:
 
 
 def trim_conc(entries: list[dict]) -> list[dict]:
-    """Trim each parallelism config's concurrency sweep to its highest point.
+    """Trim each parallelism config's concurrency sweep to its lowest point.
 
     Non-full-sweep PRs only need a single concurrency point per parallelism
     config to validate a change runs end-to-end, so the shared cluster stays
     clear. Push-to-main and ``full-sweep-enabled`` PRs skip this reduction.
 
-    The retained value is the maximum configured concurrency — independent of
+    The retained value is the minimum configured concurrency — independent of
     the source ordering of ``conc-list`` / ``conc-start``.
 
     Input comes from ``json.loads(subprocess.stdout)`` so ``conc`` is always
     ``int`` (single-node) or ``list`` (multi-node); other single-node fields
     are hashable scalars.
 
     - Single-node entries: group by every other field and keep only the entry
-      with the highest ``conc`` per group.
-    - Multi-node entries: trim the ``conc`` list in place to ``[max(conc)]``.
+      with the lowest ``conc`` per group.
+    - Multi-node entries: trim the ``conc`` list in place to ``[min(conc)]``.
     """
     groups: dict[tuple, list[int]] = {}
     out: list[dict] = []
@@ -66,7 +66,7 @@ def trim_conc(entries: list[dict]) -> list[dict]:
         if entry.get("prefill") is not None:
             conc = entry.get("conc")
             if isinstance(conc, list) and len(conc) > 1:
-                entry = {**entry, "conc": [max(conc)]}
+                entry = {**entry, "conc": [min(conc)]}
             out.append(entry)
             continue
 
@@ -77,7 +77,7 @@ def trim_conc(entries: list[dict]) -> list[dict]:
     drop: set[int] = set()
     for idxs in groups.values():
         if len(idxs) > 1:
-            keep = max(idxs, key=lambda i: out[i]["conc"])
+            keep = min(idxs, key=lambda i: out[i]["conc"])
             drop.update(i for i in idxs if i != keep)
     return [e for i, e in enumerate(out) if i not in drop]