rea/experiments/atomic-codemode/call-oracle.md

Cross-function context experiment: pinned question and oracle

This is an experiment acceptance oracle, unavailable to the isolated model. The question is sqlite-call-question.txt. Fix this question and the required facts before launching either model lane. Both adapters expose the same complete fourteen-method fact API; Code Mode has four MCP tools, direct has seventeen. Use the same pinned engine, artifact, model request, reasoning effort, common envelope guidance, fresh target leases and resource limits. Do not change runtime sources or HEAD during the model batch. One sample per lane cannot establish a general advantage, and the client itself already offers JavaScript dispatch in both lanes. This comparison isolates the server calling adapter, not all forms of composition or a traditional provider workflow.

The artifact is the local Debian ELF libsqlite3.so.0.8.6, SHA-256 2e6eef9a727f081f0d453b4e5e6cbd8b9ef8b6f86cbf7681cbad444d3b0b55c8. Independent GNU objdump decoding from the function entries covers 0x64060 through 0x65400; existing earlier captures independently pin the same flow. No target execution or source-code answer is used.

Required facts for an equal completed answer:

• Wrapper entry 0x65180; direct machine CALL 0x65211 targets internal helper 0x64060. The recovered call operation has an fspec annotation as input zero; that annotation's offset is not the target address.
• Caller LEA 0x65201 passes post-prologue RSP+0x20 in RDI. Structured passing storage is register:0x38, width 8. The SSA argument and the callee's retained input are separate snapshot-local identities; matching their storage supplies a candidate connection. The direct machine transfer and pointer setup support this particular connection. Do not globally infer source parameter ordinal, alias completeness or runtime binding from a slot number or a name.
• The helper preserves the input context pointer in RBP. Its 32-bit field at context offset 0x30 is incremented at 0x64693/0x64696/0x64699 for an opening parenthesis and decremented at 0x64c92 for a closing parenthesis. The parser's token/branch context must support the parenthesis interpretation; an ADD/SUB instruction alone does not identify the meaning of a field.
• In the wrapper this is [RSP+0x50], initialized to zero with the other context tail fields. After helper return zero, LOAD 0x65268 and TEST/JZ 0x65270/0x65272 check it. Zero skips the malformed block; nonzero frees the expression at 0x65274, clears its output pointer at 0x65279, then reaches the malformed diagnostic at 0x65281 through formatter call 0x65290.
• Return-code-one errors can also reach the malformed block, via 0x65239/0x6523c; return code 18 selects a separate depth-limit diagnostic at 0x65330. No demonstrated runtime-feasibility or full memory-alias claim.

The producer verifier additionally checks two formatter call targets: 0x5f031 -> 0x2adc0 (free PLT) and 0x5f062 -> 0x27e50 (vmprintf PLT), whereas their input-zero fspec offsets are 0x10000/0x10001. Resolving these PLT entries to runtime implementations is outside this task.

Native producer fidelity is checked against previously retained captures of all three functions: C, ordinary metadata/line mappings, blocks, operations, values and coverage are identical. Only schema version and structured call records are added. Older sources used pretty JSON; their smaller/newer encoded sizes must not be attributed to the call exporter. The earlier compact-writer change is a separate experiment. Unsupported old schema graphs remain unavailable with valid C retained, rather than acquiring guessed call records.

The first paired question runs at 01e5c5955 both explain the primary state flow correctly but omit the oracle's separate return-1/return-18 distinctions. Because the initial question did not explicitly enumerate those codes, the next round clarifies that requirement in sqlite-call-question.txt before either lane starts. The required facts and independent oracle stay unchanged. Earlier questions and costs remain pinned to their original commits; neither incomplete sample is counted as a full-oracle success.