rea/experiments/atomic-codemode/byte-batch-study.md

Shared native byte reads: checkpoint

The user cleared the active goal after implementation and real preflight. Work stopped before new model trials. No PR or merge into main is requested. Main was refreshed to 5112fe07a in 2dc3820b6. LeanToken returned not_indexed for this experimental path; scoped native reads were used instead.

Opt in with REA_ATOMIC_NATIVE_BYTE_BATCH=1 and an exact experimental native engine pin. Unset/0 preserves single-window acquisition. The 19 shared methods, 22 Direct tools and four Code Mode tools are unchanged. The byte batch reads only selected static loaded-image ranges; no target execution or modification.

disassemble/kuna_batch.rs shares one existing ConsoleProgram load, range resolution, mapped-end clipping and data renderer across at most 16 windows / 64 KiB selected bytes. It rejects ignored options and bounds encoded output against the existing 8 MiB capture ceiling. Each native row retains its exact request, ordinal and success or window_unavailable error. Global load or report encoding failure fails explicitly. The old read path can initialize again when a bounded anonymous region lacks a name; this is source-derived reasoning, not a measured count of internal translator initialization.

NativeByteBatchReads caches validated exact address/length facts for this target lease, deduplicates missing ranges, batches only missing work and preserves caller order. Source + source_row identify a window in retained raw Evidence; duplicates or reordered calls can share that row. Partial reads remain partial. Successful siblings survive bad per-row data; global failures preserve cached facts. Transient capture failures are not cached here. Close clears the cache. Its retained entries are bounded by source admission (128 sources times at most 16 windows); no global cache or live-memory freshness claim is added.

Old and new byte reports use one normalization owner. Exact JSON numeric tokens preserve u64 aliases and reject rounded fractional counters, with shared KunaJson support; alias disagreements fail instead of publishing contradictory facts. Original-file mapping and runtime memory remain unknown/not-observed.

Executed evidence

• Native disassemble tests: 15 passed, including the two new protocol/budget tests.
• Runtime build passed; final experimental JS tests: 128 passed.
• Kuna reference files/ranges boundary tests: 27 passed.
• Changed-file lint/format, module boundaries, affected-source typecheck and converted actual MCP schema/catalog checks passed.
• Real SDK, Direct and Code Mode LLVM checks passed all prior five literal-query goldens, paging, projection, exact ten independent byte windows and 13 complete retained source hashes per lane.
• Both actual MCP adapters passed SQLite mixed success/error, short reads, repeated exact windows and the full 64 KiB request.
• Same-engine SDK comparison passed reordered/duplicate reuse, mixed cached/new windows, all raw hashes and stale Evidence rejection after reopen.
• Actual producer cancellation preserved cached facts in partial_observation, allowed a fresh retry and completed owned cleanup. Target bytes were unchanged.

The existing full repository typecheck memory refusal is not a pass. No full check:fast/check:pr, full native workspace gate, new architecture/provider parity, or new real-model result is claimed for this phase. Earlier model outcomes retain their own immutable revisions. The prepared byte-batch oracle is unexecuted by models.

Same-engine acquisition comparison

Both runs use engine bce4dc97fda99a028aca6b2b40a0a593c87b5224ac9e03221a69313faba2d3e8, the same LLVM artifact and independent file oracle. Only acquisition is toggled between 0 and 1. Cold means a new lease, not dropped OS caches. These are single scripted observations with no model tokens.

Phase	Single-window ms / CPU s	Batched ms / CPU s
cold_ten_windows	94269.834 / 94.110456	8505.789 / 8.488729
warm_reordered_duplicates	2.598 / 0.002657	10.494 / 0.010407
mixed_cached_and_new_window	6385.777 / 6.374731	6446.780 / 6.438594

Cold ten-window retrieval drops from about 94.3 to 8.5 seconds. The warm sample is slower in the batch run; it is retained without claiming a warm speedup. New unseen windows still incur program initialization. The old workflow retains 11 native sources for the completed task; batching retains two.

Whole task	Wall s	Local family CPU s	Kernel family peak MiB
before	102.304	101.898599	861.9
after	16.533	16.327818	858.1
cancellation	15.324	15.107833	862.8

Counters are verified at payload exit while the observer lives, not final whole- scope counters. Memory includes charged file cache. Heavy work was serial, one CPU, hard 4 GiB/no swap, with early stopping at 3 GiB (3.25 GiB for builds). The first compiled native version was replaced before real provider verification to distinguish request_address_hex from numeric native address aliases; original builds/logs remain. No failed run was relabeled as passing.

Reproduction anchors

The canonical native patch targets Kuna fe01d6294eb0d1053ce343c8943646df88e143fc. Binaries, captures and explicit provider settings remain ignored. Source patch SHA-256: 56d2cd197bfcab9566c055c32ee0aece4805c7708f0a0117a8fa06a768969055.

• native-byte-batch-before-v1.json: ec7c9ab21ef91ff6f965d8e6efc9da97c4a6593d19fde660ba5cb2b1c4459dcb.
• native-byte-batch-after-v1.json: 9737291a03ad7c12700bf124134c4c06d4ef5beb8c31dda102eaa5ae0752aa47.
• native-byte-batch-cancellation-v1.json: 6dd58baed164df0747573b143c31f01d12d5d697ca7c3ff4813b545c355d0d18.
• native-byte-batch-llvm-sdk-v1.json: 04d2ee586f376c41d86d1e9c4508c6b84fcb04d71b7ce2aab6b2580169367a1e.
• native-byte-batch-llvm-direct-v1.json: 53be8d5e5ce0f70d015be98d7ac0b500702fb414bd7a0b41be023a660717eff0.
• native-byte-batch-llvm-codemode-v1.json: d94e171953fcc86a09b38e730564fae0d42068e1d2ac59b5547bd3c4837d5d42.