[antonsoo]

signal[::2] currently keeps the original sampling period, and negative slice starts are treated as literal time offsets. For example, a ten-sample signal starting at 2 s at 4 Hz reports signal[-3:].t_start as 1.25 s instead of 3.75 s.

Normalize the start against the signal length and adjust the sampling rate by direct division for both bare and tuple slices. Add regressions for negative/clipped starts, forward/reverse strides, nested slices and exact sampling-rate preservation. The separate author-list commit follows the first-contribution instructions.

Validation: 665 core tests and 51 subtests pass; Black passes on the changed Python files. On recorded EEGMMIDB S001/R04 data, all 30 slicing configurations match absolute sample-index timestamps after the fix, correcting 18 configurations and 227 event-window selections. All sample values and event-window selections remain correct. Direct rate division avoids the reciprocal-round-trip drift reported in review. This changes __getitem__; the time_slice change in #1905 remains separate.

[zm711]

I have requested @apdavison to help review this.

[apdavison]

This is great, many thanks for catching and fixing this.

I have one suggested improvement - see the inline comments.

[antonsoo]

Changed both slice paths to set sampling_rate = self.sampling_rate / step. The new exact-rate checks cover Hz/kHz units, forward/reverse strides and all three indexing forms. The core suite passes 634 tests and 51 subtests; all 30 recorded EEG slicing cases still match the independent sample-index timestamps and event windows.

[antonsoo]

Merged current master (d50b9b95) into this branch as 8d773c44, retaining the accepted contributor-name correction. The timestamp implementation and its tests are unchanged. All 665 native core tests and 51 subtests pass, as does Black. All 30 recorded EEG slicing reports remain identical to the reviewed direct-rate revision; current master still has the 18 timestamp and 227 event-window mismatches that this fixes.

[antonsoo]

The failing Windows job is TestSegment.test_times, at min(t_starts) while constructing its expected value. I reproduced it on both untouched current master d50b9b95 and this head: random.seed(42201); np.random.seed(42201) makes the first random_segment() contain five image sequences and none of the five child types the test examines. Its expected-time list is empty. The generator, Segment implementation and test are byte-identical on both branches; the other red matrix jobs were cancelled after this failure.

The slicing tests pass in that job. I can prepare a separate fix for the randomized test if helpful.