…al-data (#169)

The previous TestHRV.test_gamma_in_range coupled three orthogonal claims
into one mandatory mid-CI network call:

  1. compute_gamma maths recovers γ ≈ 1 from a 1/f PSD,
  2. _ensure_loaded enforces the 3-subject floor,
  3. real PhysioNet/NSR2DB data lies in the physiological band.

The single point of failure (PhysioNet/wfdb network) made (1) and (2)
non-testable when the network flaked, and forced #168 to bump n_subjects
3→5 — a margin bandaid, not a fix.

This change reduces the test to first principles by splitting along the
three axes:

  • test_compute_gamma_recovers_unity_from_synthetic_1f
      Always-on. Synthesises a 1/f signal through the same VLF-band
      Welch+mask the adapter uses; no wfdb, no network. Catches the
      maths regressing.
  • test_floor_minimum_enforced
      Always-on. Monkeypatches HRVPhysioNetAdapter._load to return one
      subject below the 3-subject floor and asserts the documented
      RuntimeError. Catches floor regressions without network.
  • test_real_data_gamma_in_range
      Opt-in via NEOSYNAPTEX_RUN_NETWORK_TESTS=1. Exercises the full
      wfdb→PSD→γ pipeline on real NSR2DB data; no longer on the PR
      critical path, so PhysioNet flakes cannot block merges. Intended
      to run in a dedicated nightly/scheduled CI lane.

The floor constant (_HRV_FLOOR_MIN_SUBJECTS = 3) is named, not magic.
The CI-run-id comment from #168 is removed — the test no longer
depends on transient network state, so the diagnostic context is
documented in this commit message instead of polluting the source.

Supersedes #168.

Co-authored-by: Yaroslav Vasylenko <neuron7x@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>