Share native membership and operand ownership across full and projected facts. Preserve exact phi bindings and definition frontiers when the full graph is unavailable, keep CFG/use coverage explicit, and expose selected reads through the same CLI/SDK/MCP contracts. Verify real SQLite/zlib fidelity, encoded fallback, bounded metadata, and cache reuse; pin the next real-model question and oracle. Co-authored-by: gensokyo-dev <342580529+gensokyo-dev@users.noreply.github.com> Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
4.1 KiB
Phi alternatives through bounded native facts
Fix sqlite-projection-question.txt and this oracle before either model launch.
Both lanes receive the same fourteen methods, native engine, model/effort request,
common guidance and resource bounds. This is a seeded follow-up investigation,
not discovery from an unknown entry. The task differs from counting boundary
calls, so its costs are not a before/after speedup comparison with that task.
Artifact: Debian SQLite 3.40.1-2+deb12u2, libsqlite3.so.0.8.6, SHA-256
2e6eef9a727f081f0d453b4e5e6cbd8b9ef8b6f86cbf7681cbad444d3b0b55c8.
Implementation entry is 0xea6d0, with a distinct same-name PLT entry.
Real public projection verification and comparison with the earlier full native diagnostic graph establish these retained identities. The complete graph remains unavailable; bounded operations/values are available. It is not necessary for the answer to enumerate every global budget/counter or all unselected frontier IDs.
| Call origin / identity | Recovered RSI value | Immediate PHI definition |
|---|---|---|
0xf115c / 0xf115c:6a84 |
v1000340ac |
0xf1152:cf0f |
0xf238b / 0xf238b:7130 |
v100033f25 |
0xf238b:ceb7 |
Both calls resolve through native call specifications to ram:0x296c0.
RSI is recovered input slot two, with passing storage register:0x30/8;
that slot is not an established source parameter ordinal. The immediate
definitions are CPUI_MULTIEQUAL, not a guessed linear register assignment.
| PHI | Operand index | Value ID | Native predecessor / edge | Defining producer |
|---|---|---|---|---|
0xf1152:cf0f |
0 | v1000118a8 |
1087 / 1087:0 |
0xf2086:6a72 |
0xf1152:cf0f |
1 | v10001358b |
1153 / 1153:0 |
0xf1146:7560 |
0xf238b:ceb7 |
0 | v100012a75 |
1142 / 1142:0 |
0xf2381:712c |
0xf238b:ceb7 |
1 | v10001352e |
1158 / 1158:0 |
0xf2437:753c |
Each producer is native CPUI_PTRSUB with two nonannotation constant-space
operands, zero and the following exact u64 offset, both width eight. The addresses
are independently checked through machine LEA instructions; do not treat PTRSUB
as arbitrary integer subtraction or a native origin as a physical branch target.
| Producer origin | Verified address | NUL-terminated format text |
|---|---|---|
0xf2086 |
0x12c600 |
cannot change %s wal mode from within a transaction |
0xf1146 |
0x12c698 |
no such savepoint: %s |
0xf2381 |
0x12c500 |
cannot open savepoint - SQL statements in progress |
0xf2437 |
0x12c538 |
cannot release savepoint - SQL statements in progress |
GNU objdump decoding from the exported VM entry independently verifies all four
LEA origins and addresses. 0xf208d jumps to 0xf1152, whose local setup reaches
the 0xf115c CALL 0x296c0 (e85f85f3ff). The alternate release setup jumps at
0xf2446 to 0xf238b; that CALL targets 0x296c0 (e83073f3ff). Open setup
locally reaches the latter call. These selected static transfers do not establish
runtime path feasibility, memory stability or full function control dependence.
Independent direct file reads check these string bytes only after readelf -lW
binds addresses to identical file offsets in the read-only LOAD segment
0x11a000..0x154bfc. Native loaded-image byte queries have their own source IDs.
The byte sequences, including terminators, occupy 52, 22, 51 and 54 bytes in the
table's order. Adjacent bytes must not be merged into the reported format.
Required final facts: correct entry/target digest and facet scope; both call/PHI links, all four operand-to-producer alternatives with predecessor edge IDs; the four correct addresses and strings; verified LEA/call/jump facts; real Evidence; one VM analysis with subsequent reads reusing its snapshot; clear limits for unexpanded definitions, complete uses/CFG, aliases and runtime behavior. Judge these facts and source arguments separately from answer verbosity or RPC count. The oracle and repository are absent from the isolated model workspace.