test(k8s): transcribe launch.args from the resolver, not a test literal

Quinn found that the fixture's `agent` half was executed but its `launch`
half was not: `agents_tests.rs:447-455` built an
`EffectiveHarnessDescriptor` as a struct literal and asserted the fixture
against it, so `command`, `args` and `env` were pinned to values the test
itself wrote down. The same literal sat on both sides of the `assert_eq!`
— consistency, not correctness, and the fourth instance this arc of a
value that agrees with itself.

Mari then executed `resolve_effective_harness_descriptor` against a
coherent source record (`runtime: "goose"` plus model/provider/env_vars,
so the resolver sees what the payload serializer sees). Everything
matched the committed fixture except one field:

    launch.args: ["run", "--session"] -> ["acp"]

Transcribed from her executed result, not derived — the README's own rule
is that "recorded" means executed and transcribed, and `launch` was the
one field group still exempt from it.

Co-authored-by: Dawn (sprout agent) <c6237ef84fa537c78dcee78efd2d4e59f728859c7f194da42ac51ededfa0be05@sprout-oss.stage.blox.sqprod.co>
Signed-off-by: tlongwell-block <109685178+tlongwell-block@users.noreply.github.com>

crates/buzz-backend-kubernetes/tests/fixtures/provider-wire/deploy-full-launch.request.json

@@ -11,8 +11,7 @@
     "idle_timeout_seconds": null,
     "launch": {
       "args": [
-        "run",
-        "--session"
+        "acp"
       ],
       "command": "goose",
       "env": {