mathlib-fp 1.9.8 representative workflow qualification

Version 1.9.8 validates that separately mature domains feel like one library in applications not written to mirror the internal architecture. This guide documents the three qualified workflows, the machine-readable contract, and the automated clean-archive journey that reproduces them.

The three workflows

1. Sensor pipeline — examples/24_sensor_pipeline.pas

Domains: MathBase, EngineeringLib, StatsLib, TimeSeriesLib.

The program loads a bundled 48-reading CSV (examples/data/sensor_readings.csv), validates every reading for finite, in-range values, smooths with a windowed-sinc FIR and estimates a Welch spectrum, summarises with streaming and descriptive statistics, flags an injected anomaly and fits a linear trend with time-series analysis, rejects a non-finite reading through its validation path, round-trips the series through versioned binary interchange, and writes workflow-exports/sensor_report.txt.

• Indexing: zero-based dynamic arrays throughout; the CSV has one reading per line and blank lines are ignored.
• Ownership: TDoubleArray results are value-owned copies returned by value.
• Defaults: TOnlineStatistics.Create rejects non-finite input by default.
• Limitations: the FIR filter lengthens the output by its impulse response; statistics over the smoothed series therefore span that extended length.

2. Numerical modelling and optimisation —

examples/25_numerical_modelling_optimisation.pas

Domains: MathBase, NumericsLib, OptimizationLib.

The program validates embedded modelling data, recovers an exact linear fit and a monotone PCHIP interpolant, solves the bracketed root sqrt(2), minimises a smooth objective with unconstrained conjugate gradient and a bounded L-BFGS solve, and then exercises two diagnostics: a root solve starved of iterations reports non-convergence (iteration limit), and a negative-degree fit request raises EModellingError. The fitted parameters are round-tripped through binary interchange and the report is written to workflow-exports/model_report.txt.

• Defaults: TOptimizationOptions.Defaults supplies tolerances; the example overrides MaxIterations and bounds explicitly.
• Cancellation: not exercised; these solvers terminate on tolerances or iteration limits, which the diagnostics make visible.
• Limitations: BisectionResult raises EInvalidArgument when a bracket has no sign change; the example demonstrates the iteration-limit path instead.

3. Reproducible probability/finance analysis —

examples/26_probability_finance.pas

Domains: MathBase, ProbabilityLib, StatsLib, FinanceLib.

The program seeds a local TLocalRandom (xoshiro256**) and simulates 64 market/asset returns, estimates the asset-return distribution, runs a one-sample t-test and a CAPM-style OLS regression, computes NPV/IRR and a project decision, rejects an invalid standard deviation with EProbabilityError, round-trips the cash flows through binary interchange, and writes an interpretation to workflow-exports/finance_report.txt.

• RNG ownership: TLocalRandom is caller-owned and never touches the RTL global generator; the fixed seed makes the run reproducible.
• Limitations: seeded results are reproducible within a single build/platform; cross-platform floating-point rounding is not claimed to be bitwise identical.

Machine-readable contract

workflow-qualification-1.9.8.json is the stable, host-independent manifest. Each workflow records its source, success marker, exercised domains, bundled fixtures (with a size bound), required diagnostic output, named numerical bounds (parsed as line_prefix plus a minimum/maximum range), and exported artifact paths.

tools/workflow_qualification.py validates the manifest and rejects missing, malformed, duplicate, unsafe, or absolute paths; missing fixtures; oversized fixtures; missing diagnostic paths; and invalid numerical expectations. It is covered by tools/test_workflow_qualification.py.

The automated clean-archive journey

tools/check_workflow_qualification.py performs, for each workflow:

1. compile with FPC 3.2.2 using only src/ on the unit path; 2. copy bundled fixtures into an isolated work directory (sources are never mutated); 3. run the workflow twice from that directory; 4. verify the success marker, every diagnostic, and every numerical bound; 5. verify exported artifacts exist, are non-empty, and are byte-identical across the two runs.

The recorded result JSON names the exact platform and compiler_version that ran; it makes no claim about any other platform. The checker uses only Python's standard library and needs no network access.

Evidence

Generated qualification output is written to build-temp/workflow-qualification/results.json (or the path given by --result). Local Windows x86-64 FPC 3.2.2 evidence is summarised in QUALIFICATION_1.9.8.md. Exact Linux and Windows clean-archive candidate artifacts are produced by CI and are not pre-recorded in the repository.