PR notes: 1.9.5

Scope

This release makes representative performance and allocation behaviour inspectable, adds exact regression tripwires, and fixes one profile-confirmed internal DSP bottleneck. It intentionally leaves the frozen public 1.9 API unchanged.

Review boundary

• benchmarks/PerformanceBenchmarks.pas and BenchmarkRunner.lpr emit 14 canonical rows covering every roadmap domain and required scale.
• docs/performance-evidence-1.9.5.json owns claims, correctness budgets, exact ceilings, comparison policies, and host-keyed 1.9.4 observations.
• tools/check_performance_evidence.py validates captured or newly executed output and writes the reviewable result artifact.
• EngineeringLib.DSP.pas reuses the tested native radix-2 kernel for supported sizes and retains the explicit portable recurrence fallback.
• Release qualification and CI run the contract tests and checked benchmark; timing review remains advisory while exact storage/correctness gates fail.

Required checks

python tools/test_performance_evidence.py

python tools/check_performance_evidence.py --compiler fpc

python tools/qualify_release.py --release 1.9.5 --compiler fpc \

  --lazbuild lazbuild

The complete FPCUnit suite, example/output contracts, documentation/API gates, historical 1.9.4 numerical-evidence gates, package build, and clean-archive Linux/Windows workflows remain part of qualification.

Non-goals

This patch does not add SIMD, threads, architecture-specific assembly, external benchmark dependencies, public APIs, or universal throughput claims. The next roadmap milestone owns portability and distribution expansion.