fix: preserve singleton conditioning payloads

docs/LEARNINGS.md

@@ -154,3 +154,4 @@
 - Distributed node-cache restore also benefits from fan-out/fan-in structure. The old per-node `await key -> await Dict get -> deserialize` loop made total restore time roughly the sum of all independent Dict reads, while batching those lookups concurrently preserves prompt-order hydration but trims the network-bound portion substantially.
 - Implicit mapped fan-out cannot treat every Python `list` as a batch axis. ComfyUI semantic payloads like `CONDITIONING` are also list-backed, and splitting them per item corrupts the conditioning structure into fragments like `["cond", {...}] -> "cond"` that later explode in `sampler_helpers.convert_cond`. Only explicitly batchable transport types and list-wrapped primitive sockets should participate in that automatic zipping path.
 - Primitive widget sockets are not always scalar once a graph is running. Custom nodes can legally emit `list(INT)` or similar values into an `INT` input and rely on ComfyUI's normal over-list execution behavior, so remote validation should only reject list values that are still raw widget literals, not boundary-fed or upstream-fed runtime values.
+- Singleton-list cleanup on the remote boundary needs the same nuance. It is correct to unwrap transport-added wrappers like `[4] -> 4` and `[[["node", [0]]]] -> ["node", 0]`, but wrong to flatten single-entry semantic payloads like `[[tensor, meta]]` for `CONDITIONING`; that shape change is enough to turn `convert_cond()` into a tensor indexing crash.

docs/UPNEXT.md

@@ -43,5 +43,6 @@ Partial-completion semantics for one proxy would require ComfyUI scheduler behav
 - Online testing confirmed that `RemoteEngine(session_affinity_key)` kept the split static and mapped proxies on the same deployed Modal worker for the tested workflow.
 - Online testing also exposed one implicit-mapped runtime edge: list-backed `CONDITIONING` inputs must stay broadcast across mapped item runs. That regression is now covered so future batching changes do not reintroduce the `convert_cond` failure shape.
 - Online testing also exposed a primitive-socket validation edge: upstream or boundary-fed `list(INT)` values are legitimate for nodes that rely on ComfyUI's over-list execution semantics, so only raw widget list literals should fail early.
+- Online testing also exposed a boundary-normalization edge: singleton wrapper cleanup must preserve single-entry semantic lists like `CONDITIONING` while still collapsing true scalar wrappers and wrapped prompt links.
 - Run with `COMFY_MODAL_STREAM_REMOTE_CONTAINER_LOGS=true` during broader production workflows and watch the mirrored session lifecycle logs for any real session misses under scale-out.
 - If production logs ever show split-proxy session misses under real scale-out, add a stronger affinity fallback or explicit leased-worker/session routing instead of relying on best-effort class-instance affinity.