docs(native): explain located call-input binding unknowns

Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com>
This commit is contained in:
N0zoM1z0andgensokyo-dev authored and gensokyo-dev committed 2026-11-11 18:24:58 +08:00
1 parent 87cd7070b6
commit 073b4dcc41
1 file changed
+18
@@ -531,6 +531,24 @@ nodes and edges. Node payloads are bounded to 8 MiB and serialized results to
32 MiB. Pagination returns edges whose source is on the node page and preserves
stable IDs for endpoints outside that page.
After fully examining and retaining a callee's parameters, an unmatched
recovered CALL operand is reported at the caller's call-site address in
`unknowns[].call_inputs`. One group lists its `input_indices` (positive p-code
input positions; input 0 is the destination), `call_node_id`, exact resolved
`destination` selector and caller `evidence_id`. The operands themselves stay
in the retained call node and source Evidence. For example, a three-argument
call with only parameter ordinal 0 recovered retains the index-1 binding and
reports indices `[2, 3]` together. This does not establish a variadic signature
or invent formal parameters.
A depth, function, node or edge refusal, unavailable value flow, or truncated
callee recovery does not establish such an examined gap. Those conditions
keep their existing unknowns and traversal reasons. Call-input groups remain
available when paging omits their call node; follow `call_node_id` in subsequent
pages or the source Evidence. Older traces may omit `call_inputs`. Parameter
and return bindings use canonical procedure addresses established by analyzed
dossiers while preserving the provider's destination selector.
Procedure, node and edge `evidence_id` values identify the source function
observations used by the trace. In an MCP session, retrieve them from the
Evidence bundle for inspection or export. A replayed source keeps the original