# Load source discovery and private deployment fixtures — 2026-10-04

## Scope and preserved inputs

The source boundary moves nineteen actual launcher/installer/build fixtures
from `bench/load/` to `internal/bench/load/`. The one actual Java launcher-stamp
integration method moves separately; the public v4 schedule fixture and its
remaining twenty-two function bodies stay public. All **111 original moved
function ASTs and 388 assertion calls** remain unchanged. Private fixtures
bind only their owning development root and public audit helpers, with a
runner that selects every locally defined class and inherited test case.
Imported fixture classes do not acquire duplicate suite entries.

The complete public manifest still selects the workload, framework, auditors
and generic publication fixtures. No engine, root launcher, installer,
measurement clock, toolenv API, Stage 2 adapter, canonical wire, frozen payload
or old result changes in this pass. The private runtime inputs are three
actual operations scripts and two actual SDK source files; they are source
assertions, not redistributed SDKs or fresh SDK build qualification.

`benchmark_sources.py` uses the existing owning structural loader. Source
checks now resolve their same C++, Java, .NET and Rust workload files in both
development and flat archives. Runtime-affinity and JVM-policy declarations
use that owning source root. A standalone smoke import requires an explicit,
existing results TSV. Its capability, recipe, positive pinning and temporary
publication rules retain their original implementation.

The first public source-binding pass preserves **148 function ASTs and 229
assertion calls**. The ingestion fixture correction changes only its
construction method, preserving its other **112 function ASTs and 184
assertion calls**. It now copies the owning complete harness and selected
support rather than seven scripts that omitted a newly required policy helper.
A clock fixture now uses the selected owning CLI. The shutdown source check
retains every original assertion and regex, checks actual consecutive `using`
scopes, and additionally rejects duplicate lifetime declarations. Ordinary
cold observation setup between those scopes no longer creates a false failure.
The actual compiler-backed shutdown fixtures still run.

## Original failures retained

| Source pin | Host | Original outcome |
|---|---|---|
| `d090870356e3a6c79801d538e03c5690a4e98c6c` | HP shared/high | 235 public tests; 48 failures and one error from the old development-only result-fixture path. The private suite was not reached. |
| `09364e2ca9c9834083f3b0b0d8d4dbb44653021d` | SCAN shared/low | 422 tests; 93 failures and 70 errors. One generated import was inside a module docstring; the copied ingestion fixture omitted the policy/source helper. |
| `da80f6b5fbf8` | HP shared/high | 422 tests; two failures and 248 errors. The fixture's source import had not been inserted, a clock fixture still used the parent root, and the original shutdown adjacency check rejected valid cold observation setup. |
| `98dcb2f1562b2b5c4dfbd25519a2d8331efb7521` | HP shared/high | All 422 public tests and six auditor CLI operations pass. The private suite passes 97 cases and skips its one optional external-baseline case. The outer zero-skip assertion stops the checker. |

These failed receipts are never overwritten or renamed as passing. They are
functional source/fixture outcomes and supply no engine-performance claim.

## Qualification and read-only completion

The final source pin uses the **whole 746-file flat manifest archive**, with
an unrelated enclosing Git repository, private-launcher sentinel and results
sentinel. The private fixture tree includes the same complete harness plus
**28 explicit private files** at that pin. No minimal archive or foreign
engine directory supplies missing dependencies.

The high-priority HP shared lease passes all **422 public tests with zero
skips**, including the actual .NET shutdown compilation and Linux signal/join
checks. Four measured-sink CLI checks (C++, managed/native .NET and Rust) and
two .NET OPEN owner checks pass. Missing explicit standalone results and
missing owning support both refuse before publication. The actual original
private suite runs its C++11 timing-helper compile/execute and all private
shell wrapper/argv/assertion cases without root or real network traffic.

The separate read-only continuation supplies the launcher at the exact
pre-migration pin `66edbbf796ec413455f1a3f2306f8bba47291c06` to the original
optional baseline method. It reuses the existing complete materialization;
its receipt keeps the earlier outer failure and skip visible. Passing tests
and builds are not repeated. The continuation also runs the moved Java stamp
case directly and checks missing owning private-launcher refusal. Final
counts, source-byte/mode checks, cleanup and CPU affinity are recorded in the
continuation receipt.

The first read-only attempt fails its unchanged optional comparison because
that original fixture normalizes only `/tmp/native-pool-launcher-*`, whereas
the outer checker had directed its inert argv fixtures into the session build
root. Its 22 temporary-path differences remain recorded. The second attempt
uses the original `/tmp` fixture root for that one inert test only, with its
existing automatic cleanup; no build or source product is placed there.

The completed receipt verifies **98 distinct private cases**, with the
original 97 passes plus the separately completed optional case and **no
remaining skipped case**. The direct Java-stamp test also passes. All 746
public files in both materializations and all 28 private inputs retain their
bytes/modes, neither source tree contains generated outputs, both unrelated
parent sentinels remain unchanged, and the owning source/build materialization
is removed. The original skip and every failed outer receipt remain visible.

## Retention and admission

Original sources and all source tarballs remain in
`~/archive/libhft/load-source-discovery-20261004/` and
`~/archive/libhft/private-deployment-fixtures-20261004/`. The record evidence
includes the source maps, preserved assertion/AST proofs, checker scripts,
complete command stdout/stderr, failure and continuation receipts, and host
lease logs. Source tarballs are not website download artifacts.

Current readiness selects **743 sources** (746 archive files after templates),
with zero private references, unresolved references, scrub findings and dead
links. The **68 textual root escapes** remain a dependency worklist; source
inventory cleanliness does not qualify every remaining consumer.

This checkpoint supplies no physical Stage 5 promotion, quiet timing grade,
fresh SDK release attestation or standalone harness release. The root
QuickFIX/J v4 and Artio v9 installs, D/OrdType decision, full consumer/licence
review and coordinated public release remain open. The website archive may
mirror this evidence without changing a benchmark ranking or admission label.

## Integration with concurrent main

The normal rebase preserves all other sessions' commits. At integrated pin
`ea65ebe35ca19043fd10aff7f6d524b8b444d1f8`, **all 746 qualified harness inputs
and all 28 private inputs remain byte- and mode-identical** to qualified
`98dcb2f1562b2b5c4dfbd25519a2d8331efb7521`. The retained map compares actual
Git blobs and modes, including export templates. No earlier runtime is
silently credited to different consumer inputs; no new build/test repetition
is needed for this unchanged input set.