fix: count a pct-triplet as one character when truncating a prefix

Addresses Copilot's round-9 review on #2035.

RFC 6570 2.4.1 counts the prefix in *characters* and is explicit that a
pct-encoded triplet counts as a single one. Truncation went by code point,
so it cut a triplet in half -- measured on 6acaff70:

  x{+v:5}  v = "%61%62%63%64%65%66"  =>  "x%61%256"

The orphaned "%6" is not a triplet, so encodeAllowReserved then (correctly)
escaped its `%`, and the value went out carrying a byte the caller never
asked for. Now "x%61%62%63%64%65".

Triplets are units only under `+` and `#`. Under every other operator the
encoder escapes a `%` anyway (`%61` -> `%2561`), so the three characters
really are three and grouping them would mis-count the prefix -- asserted
both ways. Astral code points remain single units under both, which is what
`Array.from` was there for.

Also corrects the comment on `readResourceFromTemplate`, which still said
the shared helper drops blank values. It has not since 6acaff70: `params`
passes through as given, so only an absent key omits its expression and a
direct API caller can request `?q=` deliberately. A caller whose values come
from a form drops its own blanks first.

Signed-off-by: cliffhall <cliff@futurescale.com>

clients/web/src/test/core/mcp/uriTemplate.test.ts

@@ -641,6 +641,44 @@ describe("unmetRequiredGroups", () => {
   });
 });
 
+describe("prefix modifier truncation", () => {
+  // RFC 6570 2.4.1 counts characters, and says a pct-encoded triplet counts as
+  // ONE. Truncating by code point cut the sixth triplet in half -- measured
+  // before the fix, `{+v:5}` over "%61%62%63%64%65%66" produced "x%61%256",
+  // because the orphaned "%6" was no longer a triplet and its `%` was then
+  // (correctly) encoded. A byte the caller never asked for.
+  it.each([
+    ["+", "x{+v:5}", "x%61%62%63%64%65"],
+    ["#", "x{#v:5}", "x#%61%62%63%64%65"],
+  ])("keeps pct-triplets whole under %s", (_op, template, expected) => {
+    expect(expandUriTemplate(template, { v: "%61%62%63%64%65%66" })).toBe(
+      expected,
+    );
+  });
+
+  it("counts a triplet as three characters where the encoder escapes it", () => {
+    // Under a non-reserved operator `%` is escaped anyway (`%61` -> `%2561`),
+    // so the three characters really are three and grouping them would
+    // mis-count the prefix.
+    expect(expandUriTemplate("x/{v:3}", { v: "%61%62" })).toBe("x/%2561");
+  });
+
+  it("truncates by code point, not UTF-16 code unit", () => {
+    // An astral character is one character; `slice` would cut the surrogate
+    // pair in half and yield a lone surrogate.
+    expect(
+      expandUriTemplate("x/{v:2}", { v: "\u{1F600}\u{1F601}\u{1F602}" }),
+    ).toBe("x/%F0%9F%98%80%F0%9F%98%81");
+    expect(expandUriTemplate("x{+v:2}", { v: "\u{1F600}\u{1F601}" })).toBe(
+      "x%F0%9F%98%80%F0%9F%98%81",
+    );
+  });
+
+  it("leaves a value shorter than the prefix untouched", () => {
+    expect(expandUriTemplate("x{+v:9}", { v: "%61%62" })).toBe("x%61%62");
+  });
+});
+
 describe("a value defined as the empty string", () => {
   // RFC 6570 distinguishes an *undefined* variable (the expression is omitted)
   // from one defined as "". The expander used to collapse the two, which made

core/mcp/inspectorClient.ts

@@ -4972,11 +4972,17 @@ export class InspectorClient extends InspectorClientEventTarget {
     const uriTemplateString = uriTemplate;
 
     // Expand through the shared helper in ./uriTemplate.js so this and the web
-    // Resources form cannot disagree. It drops blank values so an unfilled
-    // optional field reads as *undefined* (the expression disappears) rather
-    // than as the empty string (a valueless `?topic=`), and it covers the
-    // expression shapes the SDK's own expander gets wrong -- `{a,b}`, `{;id}`,
-    // and the `{id:3}` prefix modifier (#1919).
+    // Resources form cannot disagree. It covers the expression shapes the SDK's
+    // own expander gets wrong -- `{a,b}`, `{;id}`, and the `{id:3}` prefix
+    // modifier (#1919).
+    //
+    // `params` is passed through AS GIVEN. Only an *absent* key omits its
+    // expression; a key present with `""` is a defined RFC 6570 value and
+    // expands (`{?q}` -> `?q=`, `{;q}` -> `;q`), which is what lets a caller
+    // request those URIs deliberately. A caller whose values come from a form
+    // -- where an untouched field is indistinguishable from a deliberately
+    // empty one -- drops its blanks with `definedValues` first, as the TUI's
+    // ResourceTestModal does.
     let expandedUri: string;
     try {
       expandedUri = expandUriTemplateStrict(uriTemplateString, params);

core/mcp/uriTemplate.ts

@@ -434,25 +434,56 @@ function encodeUnreserved(value: string): string {
   );
 }
 
+/**
+ * The two operators whose expansion preserves RFC 3986 *reserved* characters
+ * and existing pct-triplets rather than encoding them (RFC 6570 §3.2.3, §3.2.4).
+ */
+function allowsReserved(operator: string): boolean {
+  return operator === "+" || operator === "#";
+}
+
 /** Encodes one value for its operator: reserved characters survive `+` and `#`. */
 function encodeValue(value: string, operator: string): string {
-  return operator === "+" || operator === "#"
+  return allowsReserved(operator)
     ? encodeAllowReserved(value)
     : encodeUnreserved(value);
 }
 
+/**
+ * Splits a value into the units a prefix modifier counts.
+ *
+ * RFC 6570 §2.4.1 counts *characters*, and is explicit that "a pct-encoded
+ * triplet counts as a single character" — so `{+v:5}` over `%61%62%63%64%65%66`
+ * must stop after five whole triplets. Splitting on code points alone cut the
+ * sixth in half, and the trailing `%6` was then no longer a triplet, so
+ * `encodeAllowReserved` correctly encoded its `%`: the value went out as
+ * `%61%256`, a byte the server never asked for.
+ *
+ * Triplets are units only under `+` and `#`. Everywhere else the encoder
+ * escapes a `%` anyway (`%61` becomes `%2561`), so the three characters are
+ * three characters and grouping them would mis-count the prefix.
+ *
+ * The `u` flag makes the alternation match by code point, so an astral
+ * character is one unit rather than a surrogate pair that truncation could
+ * split — the reason `Array.from` was used here in the first place.
+ */
+function prefixUnits(value: string, operator: string): string[] {
+  if (!allowsReserved(operator)) return Array.from(value);
+  return value.match(/%[0-9A-Fa-f]{2}|[\s\S]/gu) ?? [];
+}
+
 /**
  * Applies a prefix modifier, then encodes.
  *
- * Truncation is by *code point* (`Array.from`), not by `slice`: RFC 6570 counts
- * the prefix in characters, and `String.prototype.slice` counts UTF-16 code
- * units, so it can cut an astral character in half and yield a lone surrogate.
+ * Truncation counts the units {@link prefixUnits} defines rather than using
+ * `String.prototype.slice`, which counts UTF-16 code units and so can cut an
+ * astral character in half and yield a lone surrogate.
  */
 function renderValue(value: string, spec: VarSpec, operator: string): string {
   const truncated =
     spec.maxLength === undefined
       ? value
-      : Array.from(value).slice(0, spec.maxLength).join("");
+      : prefixUnits(value, operator).slice(0, spec.maxLength).join("");
   return encodeValue(truncated, operator);
 }