rea/experiments/atomic-codemode/aarch64-xref-study.md

Composable AArch64 address references

The AArch64 OpenSSH check exposed a real completeness gap before another model comparison: both warning strings were present, but Kuna returned zero incoming references for either. Native observations had unknown discovery coverage; zero was not proof that the strings were unused. Independent GNU decoding showed ADRP plus ADD at 0x2574c/0x25750 and 0x28534/0x2853c.

The query scanner read one instruction's p-code at a time. The existing ARM pool-displacement tracker supplied the relevant bounded value-flow owner, but did not carry AArch64 page bases. The experiment extends that owner rather than introducing another interpreter or postprocessing disassembly text. SLEIGH's mnemonic licenses ADRP, and its exact constant COPY supplies the coordinate; COPY/ADD/SUB then propagate the value. Machine operand text and hardcoded register-offset tables are not reparsed. Possible page copies alone, such as ordinary MOV-immediate instructions, do not establish PC provenance.

The same calculation produces Data for materialized register addresses, or Read/Write for a LOAD/STORE through a composed default-data-space byte address. The accessed cell is identified; its mutable/GOT contents are not propagated as a live value. This also covers the selected GOT slot, which an intermediate address-only fix still missed. Space-selector constants and stored values are not mistaken for access pointers.

The directly related refactor renames the former private pool-only module to kuna_pc_relative, moves its tests out of the implementation and retains one value lookup for arithmetic and memory accesses. Keys preserve native space identity, offset and width. Partial overwrites invalidate overlaps; copies can carry a known value into a different register. Calls/jumps/returns stop carry, instruction-local predicates do not execute both arms sequentially, a pool load overwritten later in the same instruction cannot seed a stale value, unsupported widths remain unknown and interval checks include the final u64 byte without overflow. Empty/failed assembly rendering cannot license ADRP.

State remains function-local, follows fall-through addresses and keeps the existing eight-hop/eight-value bounds. It is not a complete fixed-point, cross-call, alias, runtime-value or feasible-path analysis. Unknown discovery and indirect-target coverage remain explicit through the shared parser and SDK/CLI/MCP views. Both experimental adapters and the traditional Kuna provider consume the corrected native producer; no special advantage is injected into one evaluator lane. Stock Kuna is not silently upgraded.

Pins and real artifacts

REA refreshed main through 0d2247428 before this work. The native upstream is fe01d6294eb0d1053ce343c8943646df88e143fc, overlay SHA-256 62cb9d9386d110ca1301d4f7d541d0d1ee676f17ca888163cf7d69d8a9d07c53. The copied immutable embed-specs/opt1 engine SHA-256 is e43185ff11e74c18753e8755c7757ef6ede415c87a1dac9f04bb39856de87f43. Default forward-index and reverse-worktree patch checks pass. New analysis patch sections use Git-supported unprefixed empty context lines, preserving line counts/source bytes without whitespace-only lines in the patch artifact.

The target is extracted Debian arm64 OpenSSH, SHA-256 de352e4ace7028d9816fd0ee9d22936bc1b5622c400d5b0b7df660df3426b068. The storage study records its package URL, digest and unverified package-signature scope. It was neither installed nor executed. The independent machine oracle reads checked file-backed ELF ranges, validates selected instruction/branch math and GNU FDEs, and confirms the selected RELATIVE relocation. It does not read Kuna output. The oracle refuses optimized Python rather than silently disabling assertions.

The real SDK verifier returns one reference for 0xb0468, from 0x25750, owned by 0x256c0; and one for 0xb1330, from 0x2853c, owned by 0x27848. It also verifies both direct caller sites and 77 reads of GOT slot 0x10ff20, including 0x256d0, 0x27908, 0x27dcc and 0x28eb8. The slot's selected loaded-image bytes agree with the independent file read, while runtime memory remains not-observed. Full retained source digests are verified. A warm repeat uses the same Evidence IDs and identical facts.

Executed checks and costs before model trials

Check	Family wall s	Family CPU s	Kernel family peak MiB
Refreshed REA runtime build	36.660	35.931222	2989.1
17 final PC-relative tests, including compile	69.090	68.844945	885.1
All 125 Listing tests, including those 17	0.524	0.349058	59.2
Final native CLI build	104.525	104.240617	630.2
Real AArch64 SDK/source verifier	13.109	12.857963	203.1
Old x86 complete-function/prototype fidelity through real MCP	8.779	8.516305	302.0
Both actual atomic catalogs and converted schemas	1.932	1.689159	139.3
All 87 experimental JS tests	5.060	4.890171	78.8
Complete traditional capability/catalog selection	4.148	3.761593	443.2
Actual traditional MCP target/C/assembly preflight	103.002	102.596202	872.3

The independent machine verifier, changed-file lint/format checks and patch application checks also pass; their CPU/memory costs were not separately measured. All measured successful runs have observed OOM 0 and verified payload cleanup. Resource scopes use one CPU, 4 GiB hard/3 GiB early-stop memory, TasksMax192 and swap disabled. Peaks include charged file cache and are not RSS. Final whole-scope counters are unavailable; table values use verified payload-exit observations. Broad repository type checking retains the previously reported resource-limited gap; no passing full typecheck/native corpus gate is claimed.

The 17 PC-relative tests include the retained pool-displacement cases and new page/copy/space/width/clobber/predicate/access cases; they are included in the 125 Listing tests, not additional independent executions. An intermediate run passed 16/17: its obsolete assertion expected no reference when a composed temporary was dereferenced. The intended extension now identifies that actual cell Read while retaining the negative assertion against a register address-taken claim. The test was updated to require exactly that stronger typed outcome. All failed/intermediate receipts remain local; they are not model attempts.

The AArch64 two-address reference batch takes 3,548.3 ms wall / 3.535213 CPU s on a fresh lease, versus 1.54 ms / 0.001539 s for the same-lease repeat. This shows retained-source reuse, not faster native discovery. Old engine c48ad957 returned incorrect zero results; its costs are not an equivalent completed-task performance baseline for those new facts.

A separate equivalent x86 task requests the same three exact-address reference sets before/after. Every normalized fact and coverage field agrees (counts 1/1/2), using fresh leases/private caches. Its old/new cold phase is 4,028.3/4,082.2 ms wall and 3.993216/4.031850 CPU s; warm is 3.30/2.38 ms. Whole-run wall is 4.852/4.852 s, CPU 4.554857/4.580338 s, peak 155.3/157.4 MiB. This single lookup repeat shows no material observed regression for that case; it is not a model-token measurement or a general speed guarantee. OS/cache and host variation remain uncontrolled.

Frozen model task

The question and oracle ask the same behavior-level warning/pointer/guard/cleanup question through three complete interfaces. Native SSA identities are optional. Traditional preflight supplies the same strings, actual references, both C/assembly dossiers, and recovered static dat_1104f8, with target/engine digests verified. Its first dossier costs 95.291 s waiting for whole-program export; later assembly/dossier reads reuse that cache. Atomic selected-native profiles differ, so the planned comparison is a full workflow comparison, not a dispatch-only ablation.

The runtime/question/oracle are frozen at 5eede35c978c76ff9215443ed0a0d5980bfbdfe1. Execution order is traditional → Code Mode → direct. All request gpt-6.1-sol/high via Codex 0.161.0, with fresh isolated profiles/private caches and the same task. Both atomic catalogs are complete (22/4 tools); the traditional set contains every declared Kuna native operation and session/views (16 tools). Each client actually has JavaScript dispatch. This compares MCP workflows inside that common client surface.

All three complete answers pass agent review against the frozen oracle. They identify both owners, the X27→X0 pointer/helper transfer, banner ADRP/ADD into X6, four-byte field/CBZ successors, static versus live pointer knowledge, cleanup/low return bits, Evidence/target identities, unknowns and close. The traditional answer uses native C recovery plus machine addressing for the static global; both atomic answers also read the selected loaded-image pointer bytes. No native-only representation is required to pass.

Separate scripted checks verify all seven traditional, fifteen Code and fourteen Direct cited sources against delivered responses/traces and target/engine identities. Atomic retained source hashes/lengths agree; complete selected native function documents are identical between the two atomic lanes. Selected native argument/value/passing tuples and the PHI/input distinction match the delivered graphs/call facets. The independent ELF/GNU oracle checks the machine facts. Supplementary GNU checks corroborate the traditional answer's added logging wrapper call at 0x5ae6c → 0x5abb0, with X6 unchanged. Source/fidelity checks are separate from agent review of the entire natural-language answers; neither constitutes human readiness or global machine-equivalence verification.

Whole completed task	Traditional	Code Mode	Direct atomic
Required answer	pass	pass	pass
Cumulative input tokens	681,613	388,365	515,929
Cached input, included above	639,360	336,000	436,224
Uncached input	42,253	52,365	79,705
Output tokens	5,423	5,755	6,529
Reasoning, included in output	1,324	1,137	1,934
Unique model responses	23	13	15
Client executor / wait calls	16 / 6	12 / 0	14 / 0
MCP attempts, including failures/discovery/open/close	13	12	31
Rejected MCP attempts	2	0	0
Atomic fact reads	n/a	25 inner reads	25 direct reads
Native function dossiers / snapshot requests	4 dossiers	3 requests, 2 unique captures	2 captures
Model wall seconds	345.676	151.264	164.942
Family wall seconds	346.217	151.834	165.421
Local family CPU seconds	186.786767	41.152525	43.116751
Kernel family peak MiB	973.1	695.7	600.4

Both traditional rejections are instruction-interior addresses passed to exact procedure lookup. The inventory was available, the model used it and completed; this is not the earlier under-equipped baseline. All four traditional dossiers report the same ready project-export key, so the observation does not establish a broken project cache. Its extra logging-wrapper investigations and native profile add work beyond the two selected atomic captures. All attempts remain charged; failed/intermediate development checks are separate from model trials.

Code uses 24.7% less cumulative input and less wall/CPU time than Direct in this pair, but more peak memory. Compared with traditional it uses 43.0% less cumulative input and less wall/CPU/memory, but 23.9% more uncached input and more output. There are only two fewer model responses than Direct, despite nineteen fewer MCP requests. Same fact-read counts do not mean identical native work. This is one new architecture/version sample with uncontrolled OS/model caches, host contention, service build and remote compute, not a stable Code Mode advantage, a price claim or broad complex-software superiority.

Unique per-response usage sums match the task totals. All target, engine and runtime fingerprints remain unchanged. All three profiles are removed; verified payload-exit observations show OOM 0 and descendant cleanup. The same resource bounds and kernel-memory interpretation as above apply.

Raw model records are arm-warning-{traditional,codemode,direct}-v1/ with sibling resource logs/configs; atomic traces and explicitly retained sources have matching prefixes. Derived records are arm-model-costs.json, arm-answer-checks.json, arm-answer-review.json and arm-traditional-delivered-evidence.json. Reviewed answer SHA-256 values are:

• Traditional: 5312aa0757b35b8a28b8d1c60af652760e445b0463117eb41def1e2eeb5031eb
• Code: e331df37f53e58a85bea61f37b5e3b7c6a6848af78a0225383df0d9a481121a7
• Direct: 792b613d25e9c5dfec4e4dd03c13d265f24937e0eafc817931bd5a563e496408

Remaining native identity/control-flow question

All answers detect the recovered caller's no-return treatment of the warning call and use machine instructions for the requested conditional post-call path. The recovered call facet omits the strict call. Complete recovered-graph coverage therefore does not establish a complete machine graph or a correct no-return inference. This is a concrete remaining backend/identity investigation.

Supplementary static observations identify a suspicious boundary: the warning epilogue restores LR/SP and branches at 0x25858 to 0xbad0. That address starts with BTI, followed by ADRP at 0xbad4 and a slot load at 0xbad8 from 0x10f390. ELF R_AARCH64_JUMP_SLOT names that slot free@GLIBC_2.17. The native inventory instead exposes sub_bad0 with size bound 4 and free at 0xbad4 with bound 20. GNU's synthetic PLT label at 0xbad0 says closedir, which the actual slot relocation contradicts; it must not be used as symbol identity. Whether the native BTI/veneer split causes the no-return propagation still needs a producing-boundary investigation and regression proof. No such fix or rerun is attributed to the frozen comparison above.

Raw receipts live under .cache/atomic-codemode/ as arm-*: initial missing xrefs, intermediate/final native test/build logs, SDK/source records, x86 fidelity and equivalent-query costs, actual catalogs and traditional preflight. Binaries, native projects and raw captures remain uncommitted.