fix(core,studio): a voice track is never a carve bed, and never carves its own group

A narration clip inside a Voiceover group had a carve pointed at that
group — a member ducking the bus it feeds. Three faults, each sufficient
on its own.

**No bed-eligibility rule.** `couldBeCarveSource` has said since it was
written that music and sfx cannot be sources, and it is called from
nowhere — exported, tested, dead. Nothing ever asked the near-end
question: can this track be the BED. `showCarve` only asked "is anything
already carving against me, and is there anything to listen to", so a
voice track was offered the control like any other. Added
`couldBeCarveBed` beside its sibling and wired it in.

**Offering is not applying.** A bed with exactly one candidate carves
itself unasked, which is right for a track named `music-bed` and wrong
for one named `a1` — a decision taken off a name that said nothing is how
a carve appears that nobody remembers configuring. `isNamedCarveBed`
gates self-application on a name that positively reads as a bed; the
picker stays looser, the same split the source side already makes between
`sourceOptions` and `autoSourceIds`.

**A bed was offered its own group.** The candidate scan excluded exactly
one element, the bed itself. Its siblings survived that filter and rolled
up into the very group the bed belongs to, which came back as a
candidate — and being the only one, was applied. The mirror case too: a
group bed's id matches no <audio> id, so nothing stopped a group carving
against itself. `collectCarveCandidates` now takes the bed's id and drops
both it and its group.

An existing carve still shows its module (`carve !== null`), so nothing
already configured becomes unreachable — only newly offered and
self-applied ones are refused.

The bed/relationship predicates moved to `useFxCarveGrouping.ts`, next to
the source-eligibility rules they belong with. That is also what puts
`useFxCarve.ts` back under the 600-line ceiling it crossed here.

Five tests, each mutation-checked against the pre-fix code. Verified live:
selecting `vo-2` renders no carve module; `music-bed` still gets one,
listening to `Voiceover (4)`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

packages/core/src/audioCarve.test.ts

@@ -10,6 +10,8 @@ import {
   clipsOverlap,
   mixCarveSources,
   couldBeCarveSource,
+  couldBeCarveBed,
+  isNamedCarveBed,
   DEFAULT_CARVE,
   normalizeCarveSettings,
 } from "./audioCarve.js";
@@ -538,6 +540,30 @@ describe("classifyAudioName", () => {
     expect(couldBeCarveSource("sfx-explosion")).toBe(false);
   });
 
+  // The near-end rule, which nothing used to ask. `couldBeCarveSource` shipped
+  // with its own doc comment ("music and sfx are out") and no caller; the bed
+  // side had no predicate at all, so a narration clip was offered the carve and
+  // — finding one candidate — had one applied for it, against the group it was
+  // a member of.
+  it("never offers a voice track as the bed, but keeps an unnamed one eligible", () => {
+    expect(couldBeCarveBed("music-bed")).toBe(true);
+    expect(couldBeCarveBed("sfx-riser")).toBe(true);
+    expect(couldBeCarveBed("a1")).toBe(true);
+    expect(couldBeCarveBed("vo-2")).toBe(false);
+    expect(couldBeCarveBed("voiceover")).toBe(false);
+    expect(couldBeCarveBed("narration-3")).toBe(false);
+  });
+
+  // Showing the control is a suggestion; writing the attribute is a decision.
+  // A decision taken off a name that said nothing is how a carve appears that
+  // nobody remembers configuring — so `a1` may be offered but never chosen.
+  it("only self-applies to a name that positively reads as a bed", () => {
+    expect(isNamedCarveBed("music-bed")).toBe(true);
+    expect(isNamedCarveBed("sfx-riser")).toBe(true);
+    expect(isNamedCarveBed("a1")).toBe(false);
+    expect(isNamedCarveBed("vo-2")).toBe(false);
+  });
+
   it("treats underscores as separators, not word characters, for short hints", () => {
     // `\b` treats `_` as a word character, so `\bbed\b` used to miss `bed_01` —
     // an underscore-separated bed classified as "unknown" and could end up

packages/core/src/audioCarve.ts

@@ -188,6 +188,40 @@ export function couldBeCarveSource(...parts: readonly (string | null | undefined
   return kind === "voice" || kind === "unknown";
 }
 
+/**
+ * Could this track be the BED a carve is written onto?
+ *
+ * The other half of `couldBeCarveSource`, and the half nothing used to ask. A
+ * carve makes room in a bed for a voice; a voice track has no room to make for
+ * itself, and offering it the control is offering a track to duck against its
+ * own kind. Observed: a narration clip in a Voiceover group carved against that
+ * group — a member ducking the bus it feeds.
+ *
+ * Loose in the same direction as its sibling: a name that says nothing stays
+ * eligible, because a name is a hint and an author may know better. Only a name
+ * that positively reads as speech is refused.
+ */
+export function couldBeCarveBed(...parts: readonly (string | null | undefined)[]): boolean {
+  return classifyAudioName(...parts) !== "voice";
+}
+
+/**
+ * Does this track's name positively say "bed"?
+ *
+ * Stricter than `couldBeCarveBed`, for the one act the author did not ask for:
+ * applying a carve on their behalf. Offering the control on a track named `a1`
+ * is a suggestion they can ignore; writing `data-fx-carve` onto it is a decision,
+ * and a decision taken off a name that said nothing is how a carve appears that
+ * nobody remembers configuring.
+ *
+ * The same split the source side already makes between what the picker may show
+ * and what `autoSourceIds` may choose unprompted.
+ */
+export function isNamedCarveBed(...parts: readonly (string | null | undefined)[]): boolean {
+  const kind = classifyAudioName(...parts);
+  return kind === "music" || kind === "sfx";
+}
+
 export const DEFAULT_CARVE: HfCarveSettings = {
   enabled: true,
   sources: [],