rea/experiments/atomic-codemode/capability-map.md

Capability and comparison scope

The experiment has nineteen fact methods behind a shared SDK/CLI owner, exposed as twenty-two direct MCP tools or four Code Mode tools. The structured IR/call/ projection/switch facts require the explicitly pinned experimental Kuna patch; they are not capabilities added to stock Kuna or integrated REA by registration alone. Other providers are not wired to this experimental service.

A fair completed-task comparison first needs sufficient facts on both paths, with a common correctness oracle. A complete universal RE interface is not established here. The selected-case question requires native row/edge identities, so it compares the two atomic adapters; no traditional-provider result for that question is invented. Earlier traditional comparisons retain their own questions, engine profiles and limitations in the trial record.

Analyst question	Composable facts/relations	Executed evidence and limits
Which implementation owns an observed message?	Callable entries/aliases, encoded strings, exact-address xrefs with occurrence identity and function attribution	Both recent SQLite tasks select the owner before capture. Name equality alone is insufficient; discovery/unattributed coverage remains explicit.
Which local condition controls a selected value/call?	Snapshot-local operations, SSA values, indexed operands, definitions/uses with PHI predecessor edges, recovered CFG edges	Four real control-task runs verify the selected guards and machine instructions. Full graph transport checks establish producer fidelity, not global machine equivalence.
What crosses a call boundary?	Native call specifications, argument values and separate passing storage; selected callee inputs and bounded definition projections	Earlier pinned call/boundary trials exercise these relations. Storage matches are candidates, not proof of runtime binding or complete ABI/memory effects; a new exporter hash alone is not a repeated cross-function model trial.
Which cases map to a recovered edge?	Original case identity/label/address, independent native default, flattened case/out-edge relations	All 182 SQLite original targets match independent ELF table bytes. Four real models answer selected cases and distinguish the native default from the machine range guard. No complete case inventory is needed in model context.
What machine bytes support the interpretation?	Batched loaded-image bytes and selected linear instruction windows, each with source/coverage	The case models independently derive signed relative targets. These windows do not establish runtime execution/alignment, a complete machine CFG, live memory or arbitrary file-offset mappings.
Can the answer be audited without returning the report?	Snapshot/source identity, selected fields, batch reads, paging and retained raw/decoded Evidence	Real CLI/MCP checks verify source access and capacity fallbacks. Scope/partial/unknown survive composition; a guest-derived answer keeps its own authority.

The useful grain so far is an identified object plus a relation with its exact qualifiers: an operand needs its slot and PHI binding; a case/out-edge pair needs both identities; an address needs its artifact/address space; an optimized operation needs its snapshot and representation. One small JSON row is not automatically one independently meaningful fact. Machine instructions and normalized operations also cannot substitute for each other.

Selection, projection, batching, paging and bounded traversal compose those facts. The analysis owner derives/validates relations; Code Mode selects and combines them. A query can be fine-grained even when its fresh native capture still computes a whole function. Same-lease reuse is measured separately from that work and does not establish a cross-session cache or native acceleration.

The control-query study independently verifies all 2,845 control-dependence triples in the real SQLite VM through both MCP adapters. Two real models also complete a two-level controller/SSA/path-witness task and distinguish alternative predicates from conjunction. Unknown CFG seeds are explicit, and the snapshot owner reuses a bounded lazy postdominator index. Code Mode has lower cumulative input but higher uncached input and runtime costs in that pair; precise derived facts do not imply a preferred calling adapter.

The OpenSSH value-query study adds exact location lookup and inverse operand occurrences for value identities, including inputs/constants without definitions. It checks every retained selected version/use against the complete source graph; this is separate from machine-byte evidence and does not turn SSA occurrences into a memory-alias proof.

The stripped OpenSSH model comparison also verifies a common warning/caller/stack/guard question through the two atomic adapters and the traditional Kuna workflow. Exact machine facts suffice for the direct answer; extra SSA capture is not required by every RE question. The corrected traditional baseline has its complete declared native toolset. This expands the completed-task evidence while retaining the local scope and cost confounds.

Memory alias dependencies, full machine-flow recovery, indirect runtime binding, feasible paths and nontermination remain unknown. The current evidence does not establish provider parity, broad complex-software superiority or a stable Code Mode cost advantage. The wait repeat shows why fewer MCP requests or polls cannot replace correct answers and whole-task costs.

The path query composes selected recovered edges into one paged witness under explicit exclusions, with zero-edge/not-found/unknown outcomes. Both real SQLite MCP lanes preserve every original edge field; complete native function fidelity is checked. This removes an observed graph-reading burden without establishing feasible runtime paths or an adapter cost advantage.

The completed path-query pair answers the unchanged two-level task correctly in both actual models; each uses four path reads and three selected block reads, without a complete graph inventory in model context. Direct and Code Mode have similar wall time with mixed token/resource costs. Stock 1.833 explicitly rejects the unsupported path query while preserving readable Evidence/code in a real SDK check. These are scoped observations, not stock-provider parity or complete runtime reachability coverage.

The function-boundary study adds independent native prototype entries/return descriptors. On real stripped SSH, prototype counts 4/10 differ from raw input counts 3/44; both actual models preserve that distinction, all exact matches/missing tuples and the selected caller-to-callee/helper transfer. Call-passing storage is separate from the caller value's own coordinate. Native model labels and recovered entries do not prove original source arity/ABI. Complete old function fidelity is preserved except for the new facet. Actual old-native/stock checks return explicit unavailability with readable Evidence. The two completed model tasks have mixed token/time/resource costs; no stable Code Mode advantage or provider parity follows.

The storage study exposes native address-space descriptors and exact varnode_xref names without function decompilation. Width-qualified coordinates compose with prototype/value/passing-storage facts. Real x86-64 and AArch64 OpenSSH catalog checks do not establish big-endian, alias, runtime or full provider parity coverage.

Both real models complete the extended native storage/prototype/call task. Code Mode reduces cumulative/uncached input and wall time in that pair but uses more local CPU and more inner fact reads. The shared metadata capability is verified separately from this adapter observation; earlier mixed pairs remain.

The AArch64 reference study corrects an observed native ADRP+ADD/LOAD/STORE gap through the shared PC-relative value owner. Real arm64 SSH checks verify both warning owners, callers, selected machine words and 77 GOT-slot reads. Coverage stays bounded and does not propagate live GOT contents. Traditional C/assembly preflight can answer the same new behavior-level question, avoiding a baseline missing necessary facts. All three real models complete it, while token/time/resource tradeoffs remain mixed. All preserve the recovered no-return disagreement and use machine facts for the post-call path; a native BTI/veneer identity split is a concrete next investigation.