fix: apply pre-flatten passes to imported subgraphs

[noron12234]

Problem

Graph::flatten_graph runs four passes over a graph:

1. expand_names_to_individual_items — names: [...] into individual name items
2. flatten_selectors — selector references into the nodes they match
3. convert_reversed_connections_to_forward_connections — source into dest
4. flatten_subgraphs — inline imported subgraphs, prefixing names with the subgraph name

Passes 1-3 were only ever applied to the top-level graph. process_subgraph_node loads an imported subgraph with load_graph_from_uri, which only deserializes the file, and hands it straight to pass 4. A subgraph that uses names, a selector node, or a reversed source connection therefore reaches the flattening logic unlowered, even though that logic assumes all three passes have already run.

Reproduction

sub.json — inside the subgraph, ext_a sends hello to ext_b, written in reverse form:

{
  "nodes": [
    {"type": "extension", "name": "ext_a", "addon": "addon_a"},
    {"type": "extension", "name": "ext_b", "addon": "addon_b"}
  ],
  "connections": [
    {"extension": "ext_b", "cmd": [{"name": "hello", "source": [{"extension": "ext_a"}]}]}
  ]
}

Main graph:

{"nodes": [{"type": "subgraph", "name": "sg", "graph": {"import_uri": "file://.../sub.json"}}]}

Expected after flattening: sg_ext_a --hello--> sg_ext_b.

Actual on main:

{
  "nodes": [
    {"type": "extension", "name": "sg_ext_a", "addon": "addon_a"},
    {"type": "extension", "name": "sg_ext_b", "addon": "addon_b"}
  ],
  "connections": [
    {"extension": "sg_ext_b", "cmd": [{"name": "hello", "source": [{"extension": "ext_a"}]}]}
  ]
}

The connection is still in reverse form, and its source still names ext_a while the node has been renamed to sg_ext_a. It now points at an extension that does not exist, and nothing reports an error.

Impact per feature

Written inside an imported subgraph	Result on main
source (reversed connection)	Silently produces a connection referencing a non-existent extension. No error, message never delivered.
names: [...]	Reaches flattening unexpanded, where name is expected to be Some and names to be None.
selector node	Panics: Unexpected non-extension node in flattened subgraph (subgraph/flatten.rs:493).

The selector case is the most visible one: this code path is reachable from the tman designer HTTP backend and from the C runtime through ten_rust_graph_validate_complete_flatten, so a malformed-but-plausible subgraph file takes down the process instead of returning an error.

Fix

Extract passes 1-3 into Graph::apply_pre_flatten_passes and run them on a subgraph immediately after it is loaded, so an imported subgraph goes through the same lowering as the top-level graph. flatten_graph now calls the same helper, so there is a single definition of what has to happen before flattening.

The other load_graph_from_uri caller (GraphContent::validate_and_complete_and_flatten) already reaches these passes through flatten_graph and needed no change.

Tests

Three tests added to tests/test_case/graph/subgraph.rs, one per feature:

• test_flatten_subgraph_with_reversed_connection
• test_flatten_subgraph_with_names_array
• test_flatten_subgraph_with_selector

Verified they fail on main and pass with this change:

# without the fix
test test_flatten_subgraph_with_names_array ... FAILED
test test_flatten_subgraph_with_reversed_connection ... FAILED
test test_flatten_subgraph_with_selector ... FAILED     # panics at flatten.rs:493
test result: FAILED. 9 passed; 3 failed

# with the fix
test result: ok. 12 passed; 0 failed

The rest of the graph suite is unaffected — graph::reverse (14), graph::flatten_integration (4), graph::names_expansion (6), graph::selector (2), graph::exposed_message (10), graph::graph_info (5), graph::import_uri (6) all still pass.

Possible follow-up

validate_and_complete is also never run on an imported subgraph, so the app-URI rules are not enforced there. I left that out of this PR because it changes which graphs are accepted rather than fixing a mis-compilation. Happy to send a separate PR if you would like that tightened too.

[noron12234]

Closing for now — I want to re-check this myself before asking for review. Will reopen once I'm confident. Sorry for the noise.

[noron12234]

Reopened and rebased onto current main.

Verified the bug is still present. Applying only the new tests to main, without the fix:

• test_flatten_subgraph_with_names_array — fails, names reaches flattening unexpanded
• test_flatten_subgraph_with_reversed_connection — fails, source still references ext_a after the node was renamed to sg_ext_b
• test_flatten_subgraph_with_selector — panics at subgraph/flatten.rs:493

With the fix, all 87 tests under graph:: pass.

On the earlier CI failure: it failed at Setup tgn, not in the Rust code. The same workflow also failed on the cleanup branch around the same time, and #2269 (rust toolchain replacement cleanup) has since landed. This branch now includes it.