diff --git a/.github/workflows/codspeed.yml b/.github/workflows/codspeed.yml
index 2222f3a47..7f840264f 100644
--- a/.github/workflows/codspeed.yml
+++ b/.github/workflows/codspeed.yml
@@ -46,10 +46,11 @@ name: CodSpeed
 # is indistinguishable from that one — so check there first, before reading
 # anything here as broken.
 #
-# That machine is metered, which is the whole of why this is fenced: `main`
-# always, a pull request only when somebody applies `trigger:bench` — the label
-# bench.yml is already driven by. Bare metal should not run fork code unasked
-# either.
+# That machine is metered, which is the whole of why this is fenced: every merge
+# to `main` that changes code, a pull request only when somebody applies
+# `trigger:bench` — the label bench.yml is already driven by. Bare metal should
+# not run fork code unasked either. The one merge that does not qualify is the
+# release commit; see `paths-ignore` under `on:`.
 #
 # What it buys is the signal the tree has nowhere else. bench.yml gates on
 # memory, this job's other half measures memory, and the wall-clock ladder is a
@@ -89,6 +90,36 @@ name: CodSpeed
 on:
   push:
     branches: [main]
+    # **Only what cannot change what the benchmark executes.** A merge that
+    # leaves the built model, the workload and the installed versions identical
+    # would spend a metered macro run reproducing the previous commit's number.
+    # The release commit is the standing case — release-please writes CHANGELOG
+    # and manifest and touches no source, the git tag being the version — and at
+    # the alpha cadence those are a large share of merges.
+    #
+    # That principle is the whole of the list, and it excludes some entries that
+    # look obviously safe. `pyproject.toml` and `uv.lock` decide which polars is
+    # installed, and a floor moving under us is precisely what this series is for.
+    # `bench/` is the workload. This file is the instrument. None of the three
+    # may be added here however documentation-shaped a particular diff to them is.
+    #
+    # `paths-ignore` skips only when *every* changed file matches, and merges here
+    # are squashes, so a pull request touching docs and `src/` together is still
+    # measured. The failure mode to watch is the reverse: a path added to this
+    # list that turns out to reach the build hides a regression in silence.
+    #
+    # A skipped commit has no number, which the concurrency note below calls a
+    # hole the next pull request falls into. It is not one for these paths: the
+    # hole sits at a commit whose executable tree is identical to its parent, so
+    # the fallback compares against the same code it would have. The cost is
+    # cosmetic and worth naming because it reads like a fault — a pull request
+    # merged after one of these is footnoted `unexpected-base`.
+    paths-ignore:
+      - '**.md'
+      - docs/**
+      - .release-please-manifest.json
+      - LICENSE
+      - .gitignore
   # required for a report to have somewhere to land. `labeled` is on the list
   # because the default triple stops at `reopened`, and without it applying
   # `trigger:bench` to an open pull request would re-trigger nothing — the