feat(cache): guest-side prune, asserted against an independent oracle

Measured first, as thread entry #12 requires: a cold furyrs pipeline is 134.7s,
target/ is 3.0 GiB raw and 2.5 GiB pruned, and the stored entry is 645 MiB. A
warm job pays 8.1s (upload + extract) to skip ~120s of dependency compilation.
Publishing costs 15.7s but only on a miss. Build it.

The assertion in the plan turned out to be worthless and is replaced. It
re-scanned for */.fingerprint/<member>-* after deleting exactly those paths, so
asked to catch a planted fury-core fingerprint it exited 0 -- the prune had
removed the evidence before the check looked. An assertion sharing its globs
with the deletion can only prove that rm ran; it cannot see the failure that
matters, which is a glob that never matched the artifact in the first place.

The replacement asks a different oracle. Cargo records in each .d depfile the
sources an artifact was built from: dependencies point into the registry,
anything built from workspace source points back into the workspace. After a
correct prune, zero surviving depfiles reference workspace source outside
target/. Verified negatively too -- an artifact named totallyunrelated-abc123.d
that no prune glob matches, declaring a dependency on fury-core/src/lib.rs,
makes the script exit 2 and refuse to publish.

zstd is already present on the template, so no rebuild is needed; the gzip
fallback stays for the plain LinuxBuild2404.

src/ci_orchestrator/assets/cargo-cache-prune.sh

@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+# Strip every workspace-crate artifact out of target/, then prove none survived.
+#
+# The cache is only sound while it contains dependency artifacts alone. A
+# workspace crate's artifact carries the code under test, and reusing one can
+# turn a red build green -- the worst failure a CI system has, because it is
+# silent. So this script ends by asserting its own work and exits non-zero if
+# the assertion fails; the caller must treat that as "do not publish".
+#
+# Exit codes: 0 clean prune, 1 usage/environment error, 2 assertion tripped.
+set -euo pipefail
+
+TARGET_DIR="${1:?usage: cargo-cache-prune.sh <target-dir> <manifest-path>}"
+MANIFEST="${2:?usage: cargo-cache-prune.sh <target-dir> <manifest-path>}"
+
+[ -d "$TARGET_DIR" ] || { echo "prune: no such target dir: $TARGET_DIR" >&2; exit 1; }
+command -v cargo >/dev/null || { echo "prune: cargo not on PATH" >&2; exit 1; }
+
+# --no-deps restricts the listing to workspace members, which is precisely the
+# set whose artifacts must never be cached.
+members="$(cargo metadata --no-deps --format-version 1 --manifest-path "$MANIFEST" \
+    | python3 -c 'import json,sys; print("\n".join(p["name"] for p in json.load(sys.stdin)["packages"]))')"
+
+if [ -z "$members" ]; then
+    echo "prune: cargo metadata listed no workspace members" >&2
+    exit 1
+fi
+
+# Incremental state is large, machine-specific and never reusable across
+# clones; examples and docs are not dependencies of anything we build.
+rm -rf "$TARGET_DIR"/*/incremental "$TARGET_DIR"/*/examples "$TARGET_DIR"/doc 2>/dev/null || true
+
+prune_one() {
+    name="$1"
+    snake="${name//-/_}"
+    # Fingerprints and build-script output live in per-crate directories.
+    find "$TARGET_DIR" -maxdepth 3 \
+        \( -path "*/.fingerprint/$name-*" \
+        -o -path "*/.fingerprint/$snake-*" \
+        -o -path "*/build/$name-*" \
+        -o -path "*/build/$snake-*" \) \
+        -prune -exec rm -rf {} + 2>/dev/null || true
+    # Compiled outputs: rlib/rmeta/d in deps/, plus binaries at the profile root.
+    find "$TARGET_DIR" -maxdepth 3 \
+        \( -name "$snake-*" -o -name "lib$snake-*" -o -name "lib$snake.*" \
+        -o -name "$name" -o -name "$name-*" \) \
+        -exec rm -rf {} + 2>/dev/null || true
+}
+
+while IFS= read -r name; do
+    [ -n "$name" ] || continue
+    prune_one "$name"
+done <<EOF
+$members
+EOF
+
+# ── The assertion ────────────────────────────────────────────────────────────
+#
+# This deliberately does NOT re-use the name patterns above. An assertion built
+# from the same globs as the deletion can only prove that `rm` ran; it is blind
+# to the failure that actually matters, which is a glob that never matched the
+# artifact in the first place. Asked to catch a planted fingerprint, such a
+# check "passes" because the prune deleted the evidence before looking.
+#
+# So ask a different oracle: cargo records in each `.d` depfile the source
+# files an artifact was built FROM. A dependency's depfile points into the
+# registry; anything built from workspace source points back into the
+# workspace. That is the property we actually care about, and it holds however
+# the artifact happens to be named.
+ws_root="$(cd "$(dirname "$MANIFEST")" && pwd)"
+abs_target="$(cd "$TARGET_DIR" && pwd)"
+
+# Every depfile names its own output under target/, so that prefix is excluded;
+# what remains is genuine source input.
+leaked_paths="$(find "$abs_target" -name '*.d' -print0 2>/dev/null \
+    | xargs -0 -r grep -h -oE "${ws_root}/[^ :]*" 2>/dev/null \
+    | grep -v "^${abs_target}/" \
+    | sort -u || true)"
+
+if [ -n "$leaked_paths" ]; then
+    echo "prune: REFUSING to publish -- artifacts still reference workspace source:" >&2
+    printf '%s\n' "$leaked_paths" | sed 's/^/  /' | head -20 >&2
+    exit 2
+fi
+
+echo "prune: ok -- $(printf '%s\n' "$members" | wc -l) workspace member(s) stripped," \
+     "no surviving artifact references workspace source"