# FIX Antenna .NET: measured-window CPU investigation

A separate SCAN shared-CPU investigation completes the original order-entry and
mixed flows with 20,000 measured cycles each. Independent cold replay checks all
40,000 primary and scheduler-delay values, all original preflight/administration,
typed owners and measured phase partitions. It resolves eight original-role
self/inclusive CPU reports, with zero lost samples and zero raw profile downloads.
All SDK, producer, helper and harness inputs remain unchanged. These instrumented
local loopback pairs grant no timing or physical comparison and export zero cells.

The first worker completes the order-entry pair, both role exits zero, its
original functional/raw replay and a 10,000-bracket calibration of the original
shared Clock against CLOCK_MONOTONIC. It then refuses unresolved managed symbols
in the CPU reader. The original inputs, failed unit and all output remain intact.
Saved .NET JIT dump injection resolves the existing perf recording. The sealed
continuation independently replays that completed pair and runs only the
previously unstarted mixed pair. It preserves the hashes of every original
completed-prefix file before/after; it reruns zero completed pairs and rebuilds
no SDK, producer, helper, harness or calibration product.

The profiles select each owner's exact `measured` to `run-complete` timestamps
from the original phase file, excluding initial preflight, full warmup/prewarm,
setup and teardown. Both roles share the calibrated host clock. Each role has
five shared CPUs (server 0–4, client 5–9), with measurement on the client's first
CPU. .NET perf-map/JIT instrumentation and 199-Hz user-cycle sampling are explicit;
these are diagnostic profiles, not an unchanged physical-matrix rerun.

The resolved reports expose original SDK `PreparedMessageUtil.PrepareMessage`,
field-index removal/storage and buffer-pool paths. In the mixed server profile,
preparation and field-index removal each account for 3.15% self cycles, with
1.32% in `ByteArrayPool.GetByteArrayFromPool`. The mixed client also shows
`NewMessageChopper.ReadMessage` (1.02%) and the complete application acknowledgement
read (0.97%). These percentages describe sampled CPU work, not elapsed latency
or an admitted cross-engine ratio.

Two stripped CoreCLR offsets dominate the flat reports. Mapping their file
offsets through the executable ELF LOAD segment (+0x1000 virtual-address offset)
locates the actual `pause; dec; jne` loop at virtual addresses 0x46bc00/0x46bc02,
rather than misidentifying unrelated code at file-offset addresses. The recorded perf Build ID, matching runtime ELF SHA-256, segment mapping and exact instruction bytes are bound in [the runtime instruction receipt](../fixantenna-cpu-20261026-v2/runtime-instructions.json). Their summed
self-cycle shares are 52.18%/37.63% for server/client order-entry and 59.14%/57.66%
for server/client mixed flow. This confirms substantial runtime spinning; CPU
profiles alone do not prove the source of physical scheduler-delay tails.

The next configuration investigation is explicit original receiver placement.
The SDK exposes public `Config.RecvCpuAffinity`, but its implementation validates
physical CPU IDs against `Environment.ProcessorCount` (the restricted role's CPU
count). On these isolated masks that can silently reject a valid physical CPU.
No SDK patch or processor-count override is made here. Any adapter-level public
listener configuration must pin only its original receiver thread, prove the
chosen CPU belongs to the inherited mask, retain the measurement CPU and all
existing clocks/semantics, and qualify separately before a fresh physical run.
This is a follow-up hypothesis, not an improvement claim.

| Binding | SHA-256 / source |
|---|---|
| Continuation source | `4bf6ff60904aaaaaf5df833998720d0f7b018557` |
| Independent native receipt | `cbec63d843e6656a7b9977a090bd2a9b5b2b294bf79089e7c6018fdd2b1fda86` |
| Source/input manifest | `85f8402bd5f7d29f5cfd67f3adeb86193e48f008a03338f37a960d45dd152e03` |
| Independent collector | `316e6fd8e5f993d649388e4d1dc50d63142e4e19cb5ccd13427735544120d561` |
| Qualified producer DLL | `c0456bc27ba48c8ca5a20feece424c36206082314bbc22c8bddea468ffca511d` |
| Preserved first receipt | `d642380fcd76be277cf4a93a0693272704fe8fed623d44e0bb86ba99ab50a66c` |

The [bounded proof and eight derived reports](../fixantenna-cpu-20261026-v2/review.json)
are committed together; their individual hashes bind their original full reports.
Full original perf/JIT/sample/clock data remain on SCAN under
`~/build/codex-fullbench/libhft/fixantenna-measured-cpu-20261026-v1` and `-v2`.
The second unit `codex-fullbench-fixantenna-cpu-20261026-v2.service` exits
successfully under the outer shared CPU lease. Both completed physical DAC
campaigns and their diagnostic decisions remain unchanged; see the
[physical record](RECORD_2026-10-26_fixantenna_physical.md).