# Synthetic grading fixtures and archive live-matrix discovery

Native and .NET comparison tests now declare independent synthetic hosts,
namespace pairs, counter types and payload admissions. Public Rust grade tests
use independent fictional run declarations. Changing topology, pins, raw
samples, scheduler/admin observations, rate, stability or source identity
still invalidates the applicable test gate. These synthetic inputs establish
no SDK, installation, physical, source-release or timing qualification.

Actual retained Rust run/raw-file regressions, three checked-in private Java
tuning-profile tests, actual .NET build-provenance re-audit and root-only Rust
launcher-arm checks continue under `internal/bench/`. The three Java profile
methods retain AST-identical assertions. Original fixtures and source hashes
are byte preserved in `internal/bench/fixture-migration-originals-20261003/`
at source `40c186acfd68`; bootstrap originals are preserved separately at
`3ebbbdf17d02`. Existing historical run files and SDK/runtime payloads remain
unchanged. Synthetic IP/host fixtures use documented example values.

Shell process tests explicitly load the example route JSON instead of inheriting
private rig facts. Ingest/parameter-check fixtures read their own source files
in either layout. Synthetic .NET build receipts stay separate from the actual
retained-provenance regression.

## Actual complete archive failures and correction

The first 741-file archive checker expired at its 45-second outer limit;
no completed test result is claimed. Its source/controller/log/map remain
sealed. A fresh checker with direct log capture completed and reported
**126 failed entries, 102 passes and 144 passing subchecks**. It exposed the
Java live-matrix CLI's development-only toolenv bootstrap, fixed-parent fixture
reads and a remaining private .NET build receipt dependency. The materializer
also omitted executable modes. All failed input/source bytes remain retained;
no successful results are substituted into either failed generation.

The live-matrix CLI now resolves support through the unchanged inert
`xmachine_config.source_roots` API and refuses enclosing Git identity. Of its
112 existing functions, only `repo_root_from_script` and `git_text` differ;
planning, run/build/SDK handling, numeric, clock and row-grading functions are
AST identical. Its development SDK build producer still requires separately
supplied engine sources/artifacts; no standalone SDK producer is admitted.

Fresh source **7d008a62bfd3** is preserved on
`s/codex-fullbench-fixture-bootstrap-source-20261003`. Its complete
**741-file** flat archive retains exact Git executable modes and passes
**124 tests plus 249 subchecks, zero skips**. This includes actual fresh-process
Java dry-run CLI entrypoints, generic comparison grading, retained-sample
re-audit and inert actual shell process bodies. Sources and modes remain
identical; disposable materialization is removed after verification.

The preceding local full fixture suite passes **184 tests plus 311 subchecks,
zero skips**, including actual private profiles/evidence/root-launcher cases.
A focused bootstrap check passes **49 tests plus two subchecks**, and the added
real-Git owning/ancestor identity test also passes. Original collection-name
and inherited-route assertion failures are retained separately.

## Retention and open gates

Original failed generations and the qualified fresh generation live beneath
`~/archive/libhft/benchmark-fixtures{,-corrected,-qualified}-20261003-*`, with
[sealed repository evidence](evidence/benchmark-fixtures-20261003/SHA256SUMS).
The source refs preserve both implementation generations. Runtime workloads,
shared clock/build APIs, canonical fixtures, SDKs and frozen physical payload
pins are unchanged. No JVM/native benchmark, network or physical run is claimed.

Readiness selects **738 source files**, with **57 scrub findings in five files**,
84 textual source-root escapes and zero dead documentation links. Remaining
scrub work is in historical notebook/presentation generators and the matrix
renderer/tests. SDK closure, fixture alignment, root installs, physical/report
qualification and standalone release admission remain open.

Safe integration preserves 736/741 qualified files exactly. The five differences
are the source-layout explanation and the microbench owner's independent
Stage 2 test/guide boundary. All qualified runtime, comparison fixture,
configuration, workload and support bytes are identical; 39 focused grade/source
checks pass again without skips. Both original failed generations remain intact.
The microbench owner retains all sixteen moved private test-method assertions
and its own actual 57-check standalone archive qualification.