High-level, since this is still draft.

Nice motivation — an older manager needs to forward transports it does not yet know about. The mechanism (a fallback Unknown { name, config } variant + a Modify hook to strip it from OpenAPI) is clean, and the shell/pytest cross-version harness is a nice touch. A few things worth resolving before this leaves draft:

1. Please verify malformed known configs still produce a good error. TransportConfig is adjacently-tagged (#[serde(tag = "name", content = "config", …)]) with Unknown marked #[serde(untagged)]. That mixture is unusual, and serde's untagged fallback can eat the specific error from a matched tag when a sibling variant is untagged. Please add a test that submits e.g. {"name":"kafka_input","config":{ "bootstrap_servers": 42 }} and asserts the user gets a kafka_input-specific parse error, not a silent promotion to Unknown { name: "kafka_input", … }. If it does fall through, the fix is to make Unknown only match when name is not one of the known snake-case tags (custom Deserialize that checks the tag against a fixed set before falling back).

2. Gate the acceptance path on runtime_version explicitly. The test script runs the manager with FELDERA_UNSTABLE_FEATURES=runtime_version, but the config change itself is unconditional — any client can now submit an arbitrary {"name":"whatever","config":…} transport and the manager will store it. If custom-runtime forwarding is meant to require an explicit runtime version, the API layer should reject Unknown variants when runtime_version is unset (or when the effective runtime version equals the manager's own build). Otherwise the "invalid transport" ergonomic in point 1 gets worse.

3. OpenAPI HideUnknownTransport is position-dependent. schema.items.pop() works only if Unknown is the last TransportConfig variant. The test enforces that indirectly by checking len == 24 and !contains("unknown"), but a future variant added below Unknown silently corrupts the schema. Consider popping by matching the schema for the Unknown shape rather than by index, or at least add an inline debug_assert! in modify that the popped schema is the Unknown one.

4. name() for Unknown returns the raw name. Fine, but downstream code paths that log/format transport names may now surface arbitrary user-controlled strings; make sure log-injection isn't a concern (probably fine given they go into structured tracing fields, but worth a look).

5. Minor: manual test-plan shell script is nice, but consider whether it should also assert that a malformed known config (e.g. kafka_input with a bogus field) still produces a specific parse error rather than being silently promoted to Unknown — that's a good regression fence for point 1.

I'll do a proper line-by-line pass once you mark it ready.