Unify call edge representation across language backends

[tanbing117]

##Background

FM-Agent currently supports multiple language backends (codegraph, regex-based extractors, and other language handlers). However, the call edge representation returned by these backends is inconsistent.

The current interface mainly uses:

{
caller_fqn: {
callee_fqn,
...
}
}

This representation is sufficient for top-down layer generation, but it loses important edge metadata and makes it difficult for future analyses (e.g., security plugins) to consume call graph information consistently.

##Problem

The current call edge interface has several limitations:

1. Edge metadata is lost

Different backends may distinguish different edge types, such as:

normal function calls
constructor calls

However, the current FQN-set representation merges them into the same relation:

caller -> {callee1, callee2}

The edge type information cannot be preserved.

2. Call-site information is unavailable

The current representation only stores:

caller function -> callee function

It does not preserve where the call happens.

This prevents downstream analyses from reliably associating source-level information with a specific call site.

For example:

check(password)
check(token)

Both become:

foo -> check

and the two call sites cannot be distinguished.

3. Backend formats are inconsistent

Different language handlers currently return different formats:

dictionary-based edges:
{
caller: {callee}
}
list-based edge objects:
[
{
"caller": "...",
"callee": "..."
}
]
custom edge objects

Every consumer must understand each backend's format separately.

This makes extending the call graph interface difficult.

##Proposal

Introduce a normalized call edge representation inside the language registry layer.

All language backends should be converted into:

[
{
"caller": FunctionId,
"callee": FunctionId,
"kind": "call" | "constructor",
"language": str,
"span": {
"file": str,
"start_line": int,
"start_column": int,
}
}
]

The normalization layer should:

accept existing backend formats
validate required fields
normalize optional metadata
deduplicate edges at call-site granularity
preserve backend-specific information when available翻译