[ameoli]

Superseded by #107, which contains the same commit and changes under the replacement branch.

The fractional-lag fix was already implemented, but the recovery tests compared the same sweep index from separate campaigns and therefore reused receiver-noise streams. A temperature/strain pair could pass without demonstrating recovery from independently noisy acquisitions.

This change completes the remaining validation for #105:

• Use reference sweep 0 and measurement sweep 1 with a shared fiber seed, including the existing 2 mm/1 cm/5 cm strain tests.
• Add 16 full-pipeline cases covering ±17 microstrain and ±5 K, two fiber seeds, and single-ended/balanced detection. Check the complex fiber fingerprint, receiver-noise residual correlation, ADC clipping, median bias, window RMSE, and the unperturbed region.
• Validate the shipped fixed-8x/continuous estimator against 258 known fractional/periodic-boundary shifts across three Rayleigh seeds and two window sizes. Document why this replaces the original configurable-padding proposal.
• Record metrics in pytest JUnit properties and document the tested conditions and limitations.

Validation on Windows / Python 3.12.10 (NumPy 2.5.4, SciPy 1.18.1): full suite 726 passed, 4 existing time-warp warnings; targeted tests 28 passed. git diff --check passed. Across the 16 new pipeline cases, maximum absolute relative median bias was 0.601%, window RMSE 2.781%, and absolute receiver-noise correlation 0.00386. Maximum noiseless circular lag error was 1.04e-8 bins. Negative controls showed correlation 1.0 when reusing a sweep, and the old parabolic estimator recovered 6.733 K from a +5 K input and failed the new recovery criteria.

Production code and estimator behavior are unchanged. These cases validate the stated simulated receiver-noise regime, not experimental accuracy or robustness to arbitrary noise/decorrelation.

Related: #105.