diff --git a/changelog.d/10820-diag-feature-scope-comment.md b/changelog.d/10820-diag-feature-scope-comment.md
new file mode 100644
index 0000000000..7e98417858
--- /dev/null
+++ b/changelog.d/10820-diag-feature-scope-comment.md
@@ -0,0 +1,28 @@
+**docs(runtime): correct the `diagnostics` feature comment's scope.**
+`crates/perry-runtime/Cargo.toml`'s comment above `diagnostics = []` asserted
+"None are on a hot path". That is true of the things it enumerates
+(`PERRY_GC_DIAG`, the typed-feedback dump, the v8 heap-snapshot builder,
+`process.report`) and false as a blanket statement, because the feature does
+**not** gate `hot_diag` at all: `lib.rs` declares `pub mod hot_diag;` with no
+cfg, the file contains zero `cfg(feature` occurrences, and `enum_on()` is
+called per string concat (`string/concat.rs:688`, `:1052`) and per property
+enumeration in **every** build, shipped included.
+
+This matters because the comment is load-bearing in people's reasoning: a peer
+session quoted a neighbouring `hot_diag` comment as evidence about generated
+code and proposed a fix on that basis. Comments here are read as evidence, so
+a blanket claim that is only true of an enumerated subset is worth narrowing.
+
+**No performance claim attached, deliberately.** The obvious follow-up —
+collapsing `enum_on()`'s `OnceLock`-probe-plus-`AtomicBool`-load into a single
+three-state `AtomicU8` with a `#[cold] #[inline(never)]` resolve arm — was
+implemented and measured against its exact parent commit: **399.76 → 401.19
+instructions per short concat**, bare-loop control 3.01 / 2.99 in both arms.
+That is +0.4% on a ~400-instruction operation, indistinguishable from zero, so
+the change was reverted rather than shipped as churn.
+
+The instructive part is why the theory was wrong. A peer measured a real win
+(zod −5.2%, an S40 fixture −12.4%) from *deleting* an early-returning
+diagnostic gate. That win came from removing the call and its inlining
+barrier entirely — not from making the gate's body cheaper. Saving one load
+out of four hundred instructions is below what any probe here resolves.
diff --git a/crates/perry-runtime/Cargo.toml b/crates/perry-runtime/Cargo.toml
index 8b5e6d75ce..ed59532439 100644
--- a/crates/perry-runtime/Cargo.toml
+++ b/crates/perry-runtime/Cargo.toml
@@ -133,7 +133,13 @@ bun-cli-utils = [
 # pulled only by them, which dead-strips when unreferenced): GC cycle telemetry
 # (`PERRY_GC_DIAG`), typed-feedback trace dump (`PERRY_TYPED_FEEDBACK`), the v8
 # heap-snapshot builder (`v8.getHeapSnapshot`/`writeHeapSnapshot`), and
-# `process.report`. None are on a hot path. The env-driven dev diagnostics
+# `process.report`. None of THOSE are on a hot path. Note this feature does
+# NOT gate `hot_diag`, which is declared unconditionally (`pub mod hot_diag;`
+# in lib.rs, no cfg) and whose `enum_on()` runs per string concat
+# (`string/concat.rs`) and per property enumeration in every build, shipped
+# included. Measured at under 0.4% of a short concat, i.e. below probe
+# resolution -- so this is a scope correction to the comment, not a
+# performance claim. The env-driven dev diagnostics
 # degrade gracefully when off (auto-optimize leaves this off unless the program
 # uses a heap-snapshot / `process.report` API, which the compiler detects).
 diagnostics = []