# Latency distribution notebook declared inputs

The latency distribution generator now requires `--inputs` JSON for its
retained-format source roots, HDR dependency and source collections. Optional
engine/scenario catalogues and presentation context are explicit inputs.
Relative paths resolve from the declaration, not the invoking checkout.
Outputs use a separate session build root; input/output overlap and output
child symlinks are refused. `--check-inputs` validates without writing reports.
The CrossFIX generator also quotes declared paths safely in notebook cells.

Of the original generator's **60 functions, 56 remain AST-identical**. The raw
CSV discovery loop and HDR converter retain identical AST after replacing only
their discovery-list and jar-path inputs. Notebook construction and the CLI
supply the declared sources and prose. Original percentile, source-quality,
canonicalization, grouping and sample filters remain unchanged.

The exact original generator and campaign configuration are preserved under
`internal/bench/java/fixbench/scripts/latency-notebook-originals-20261003/`,
pinned to `fba97c79a7cacb162d30a5ac949403d06627658e`. Private regression checks
reproduce all **three derived CSVs byte-for-byte** and every original Markdown
cell, normalizing only the generation timestamp and output directory.

## Verification and retained evidence

Source **e0bf1ba9d098e5f0702a3b24901136c07022e0eb** is retained on
`s/codex-fullbench-latency-notebook-source-20261003`. The local combined suite
passes **26 tests and 126 subchecks, zero skips**. A fresh complete **748-file**
stripped archive passes **18 public tests and eight subchecks, zero skips**,
including both real notebook kernels: **12 CrossFIX and 58 latency code cells**.
The enclosing unrelated Git repository and engine marker do not supply inputs.
Every materialized source byte and executable mode is unchanged after the run;
the disposable source materialization is removed.

Receipts, source maps, controller, output and dependency installation log are
sealed in `~/archive/libhft/latency-notebook-20261003-e0bf1ba9d098` and
[repository evidence](evidence/latency-notebook-20261003/SHA256SUMS).
Earlier context/CrossFIX fixture failures remain retained as earlier failures,
not failures of this notebook generation.

Readiness selects **745 files**, with **zero scrub findings, zero private or
unresolved references and zero dead documentation links**. The **84 textual
root escapes** still require dependency classification and actual standalone
SDK/runtime qualification. No engine build, HDR conversion, benchmark, timing,
physical Stage 5, report or public release admission is established by these
rendering checks. Frozen campaigns and Stage 2 adapter contracts are unchanged.

Integration preserves a concurrent recipe auditor change accepting and
validating `rig_config_sha256`, together with its regression. **746 of 748**
qualified archive inputs remain byte-identical; both changed files are outside
the notebook/rendering/kernel inputs. The initial exact-equality assertion
correctly detected these changes before review. Focused pinning checks pass
**39 tests and 135 subchecks**, with one existing unavailable-locale skip.
The input map and log are retained beside the original qualification.