diff --git a/src/diagnostic/registry.rs b/src/diagnostic/registry.rs
index c4ceaf12..782879e0 100644
--- a/src/diagnostic/registry.rs
+++ b/src/diagnostic/registry.rs
@@ -242,35 +242,22 @@ do not need braces:
     },
     ErrorEntry {
         code: "ILO-P017",
-        short: "inline lambda captures outer scope",
-        long: r#"## ILO-P017: inline lambda captures outer scope
+        short: "use-import failed",
+        long: r#"## ILO-P017: use-import failed
 
-Phase 1 inline lambdas — `(p:t>r;body)` passed to a HOF like `srt`, `map`,
-`flt`, `fld`, `grp` — cannot close over variables from the enclosing function.
-Every name referenced in the body must be a parameter of the lambda, a name
-bound locally inside the lambda body, or a known top-level function/builtin.
+A `use "path.ilo"` declaration could not be resolved. Possible causes:
 
-**Wrong:**
-
-    rank xs:L n threshold:n>L n
-      srt (x:n>n;-x threshold) xs
-
-The lambda references `threshold` from the enclosing scope.
-
-**Fix A: use the HOF's ctx-arg form.** Every closure-aware HOF accepts an
-optional context value that is threaded through every call:
-
-    rank xs:L n threshold:n>L n
-      srt (x:n c:n>n;-x c) threshold xs
-
-**Fix B: define a top-level helper** that takes the value as a param and use
-`srt fn ctx xs`:
+- The path is not reachable from a file context (inline code via
+  `ilo '<src>'` has no base directory to resolve against)
+- The file does not exist at the given relative path
+- The file could not be read (permissions, IO error)
 
-    diff x:n c:n>n;-x c
-    rank xs:L n threshold:n>L n;srt diff threshold xs
+The diagnostic message identifies the specific failure mode.
 
-Closure capture is tracked as a Phase 2 follow-up; once it lands, free
-variables will be captured by value automatically.
+**Note:** ILO-P017 used to be raised by inline lambdas with captures from
+the enclosing scope. Closure capture now works on every engine (tree, VM,
+Cranelift JIT/AOT) so that path no longer errors; the code was repurposed
+for `use`-import resolution.
 "#,
     },
     ErrorEntry {
diff --git a/src/parser/mod.rs b/src/parser/mod.rs
index 4926a94c..cd4fbb50 100644
--- a/src/parser/mod.rs
+++ b/src/parser/mod.rs
@@ -3588,9 +3588,10 @@ For variable-position list indexing bind the head first: \
     /// returns `Expr::Ref("__lit_N")` so HOFs see a fn-ref identical to a
     /// named helper.
     ///
-    /// Phase 1: closures are rejected. Any reference to a name that isn't a
-    /// param, isn't a local binding, and isn't a known function/builtin
-    /// raises ILO-P017 pointing at the Phase 2 follow-up.
+    /// Free variables in the body (names that aren't params, locals, or known
+    /// top-level fns/builtins) become Phase 2 captures: appended as extra
+    /// params on the lifted decl and snapshot by value via `Expr::MakeClosure`
+    /// at the call site. Phase 2 closure capture works on every engine.
     fn parse_inline_lambda(&mut self) -> Result<Expr> {
         let start = self.peek_span();
         self.expect(&Token::LParen)?;
diff --git a/src/vm/mod.rs b/src/vm/mod.rs
index 22b801b2..086825c8 100644
--- a/src/vm/mod.rs
+++ b/src/vm/mod.rs
@@ -69,12 +69,13 @@ pub enum CompileError {
     UndefinedVariable { name: String },
     #[error("undefined function: {name}")]
     UndefinedFunction { name: String },
-    /// Inline lambda with capture (`Expr::MakeClosure`) cannot lower to the
-    /// register VM yet — HOF dispatch with N captures depends on the parked
-    /// FnRef NaN-tagging effort. Returning this from `vm::compile` lets the
-    /// default runner fall through to the tree interpreter cleanly, rather
-    /// than panicking partway through codegen.
-    #[error("inline lambda capture for `{fn_name}` is tree-only")]
+    /// Inline lambda with more than 255 captures cannot be encoded by the
+    /// register VM's `OP_MAKE_CLOSURE` instruction, which uses an 8-bit
+    /// capture-count field. Phase 2 closure capture is otherwise fully
+    /// supported across every engine; this variant only fires on the
+    /// pathological wide-capture case. Kept around so the compiler can
+    /// surface a structured error rather than panic.
+    #[error("inline lambda `{fn_name}` exceeds the 255-capture VM cap")]
     UnsupportedClosureCapture { fn_name: String },
     /// The register-VM byte-encoded instruction set is capped at 256 live
     /// registers per function (8-bit register field). When codegen for a
@@ -6558,13 +6559,13 @@ impl NanVal {
                 }
             }
             Value::Closure { fn_name, .. } => {
-                // Closures don't round-trip through NanVal — VM/Cranelift HOF
-                // dispatch with N captures is downstream of this PR. The VM
-                // `compile_expr` branch for `Expr::MakeClosure` records a
-                // `CompileError::UnsupportedClosureCapture` before we get here
-                // in any well-formed program, so this sentinel-string fallback
-                // is belt-and-braces for stray Closures threaded through the
-                // tree-bridge dispatcher.
+                // Closures don't round-trip through NanVal as a bare tagged
+                // value — the VM uses a dedicated `HeapObj::Closure` plus
+                // `OP_MAKE_CLOSURE` for native dispatch. This sentinel-string
+                // fallback is only reached if a tree-level `Value::Closure`
+                // leaks into the NanVal conversion path (e.g. through a stale
+                // tree-bridge call site); native compilation goes through
+                // `Expr::MakeClosure` instead.
                 NanVal::heap_string(format!("<closure:{}>", fn_name))
             }
         }