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

Calls and values surviving full-graph capacity

The question is sqlite-boundary-question.txt. Fix it and this acceptance oracle before either real-model run. Both lanes use the same complete fourteen-method API, preserved engine, model/effort request, common guidance and resource limits. This task requires VM decompilation once per fresh lease; it is different work from the earlier shallow-reference question. Do not compare their task times as an adapter speedup. Actual client JavaScript dispatch is available in both lanes.

Artifact: Debian libsqlite3-0:amd64 version 3.40.1-2+deb12u2, libsqlite3.so.0.8.6, SHA-256 2e6eef9a727f081f0d453b4e5e6cbd8b9ef8b6f86cbf7681cbad444d3b0b55c8. The selected VM implementation entry is 0xea6d0, not its same-name PLT entry.

Pre-model public MCP capture with the new exporter establishes:

• Code is available. Full graph is unavailable with atomic operation capacity exceeded. Calls remain available: 416 retained block-member operations, consisting of 395 CALL and 21 CALLIND. The 395 direct targets are known, all 21 indirect targets remain unknown. Call-boundary values are available (1,294), covering targets, arguments and outputs only, with 21,349 operations scanned over 1,229 recovered blocks. These do not supply a complete definition/use graph.
• At 0xf115c, recovered call ID 0xf115c:6a84 targets ram:0x296c0. Input zero is v1000118d6, fspec:0x102f4, annotation true, width 8. Three recovered inputs use passing storage register:0x38/8, 0x30/8, 0x10/8 (RDI, RSI, RDX in this verified x86-64 lane).
• At 0xf238b, call ID 0xf238b:7130 has the same resolved target. Input zero is v100012a80, fspec:0x1030c, annotation true, width 8. Two recovered inputs use passing storage register:0x38/8, 0x30/8.
• Both first arguments are the same snapshot-local SSA value v30000004b, register:0x38/8, is_input:true, definition null. Other retained argument values have defining operation IDs outside this boundary view. Exact SSA identity is supported; source parameter ordinals and general aliases are not.

Independent GNU objdump decoding from the exported VM entry verifies the two machine CALL targets and register setup. The earlier complete direct-call pair oracle contains both. Of the 396 physical direct immediate calls, the only one absent from the recovered 395 is 0xf2493 -> 0x29370 (stack check failure). The question does not require rediscovering this extra call; the recovered-call count must not be claimed as complete machine/runtime coverage.

Required final facts: correct selected entry and facet availability/scope; both call-count groups and target-unknown distinction; both resolved targets separated from annotation offsets; argument slots/locations; shared first-input identity; selected machine checks; explicit unreadable definition/use and alias/runtime limits; real Evidence and subject digest. Do not require the model to report every additional native metadata counter. Check that its API source arguments show only one VM analysis and all relevant reads reuse its snapshot/Evidence.

Raw public verification is retained as facets-v3-mcp.json and its Evidence directory under .cache/atomic-codemode/. The ordinary VM record (C, mappings, variables and other ordinary metadata) equals the prior captured engine output. The independent oracle and precomputed answer are absent from the model workspace.