Clip v_mem to min_v_mem before generating the layer output (fixes #236)

[tritsystem]

Checklist before requesting a review

• Tests for the changes have been added (for bug fixes/features)
• Docs have been added / updated (for bug fixes / features)
• I have performed a self-review of my code
• Will this be part of a product update? If yes, please write one line about this on docs/about/release_notes.md

_{Docs: no user-facing doc change — the min_v_mem docstring ("Lower bound for membrane potential v_mem, clipped at every time step") already describes the intended behaviour; this PR just makes the code match it. Release notes are maintained per the tag-based release process, happy to add a line if you'd like one.}

---

Fixes #236.

• What kind of change does this PR introduce?

Bug fix.

• What is the current behavior?

When a LIF or IAF layer is built without a spike function (spike_fn=None) — which is also exactly what the public ExpLeak / ExpLeakSqueeze layers do internally — the layer output is the membrane potential v_mem. When min_v_mem is set, the stored state is clipped to that floor but the returned output is not, because in functional/lif.py::lif_forward_single the clip runs after v_mem has already been cloned into the output tensor:

if spike_fn:
    ...
else:
    spikes = state["v_mem"].clone()   # output copied here (pre-clip)
if min_v_mem is not None:
    state["v_mem"] = relu(state["v_mem"] - min_v_mem) + min_v_mem   # too late for `spikes`
return spikes, state

Reproduction (develop @ 904eb89):

LIF(tau_mem=30., min_v_mem=-0.5, spike_fn=None) : output min = -2.47   state v_mem min = -0.50   -> 68/72 outputs below the floor
ExpLeak(tau_mem=30., min_v_mem=-0.5)            : output min = -2.47   state v_mem min = -0.50   -> 68/72 below the floor
IAF(min_v_mem=-0.5, spike_fn=None)             : output min = -2.49                            -> 68/72 below the floor

The existing test_min_v_mem does not catch this: it uses a spiking LIF and only asserts on layer.v_mem, never on the returned output.

• What is the new behavior (if this is a feature change)?

v_mem is clipped to min_v_mem before the spike / no-spike branch, so the no-spike-fn output respects the floor. On the spiking branch the clip is also re-applied after reset_fn so the stored state on that path is byte-for-byte identical to before for every documented configuration (min_v_mem is None by default and always <= 0 < spike_threshold when set). min_v_mem is a lower bound and a spike is triggered by v_mem crossing the upper spike_threshold, so clipping earlier cannot change which neurons spike — a new test asserts the spike train is bit-identical with and without the floor on a spiking input.

After the fix, every path above reports output min = -0.5000, 0/72 below the floor.

ALIF is not affected: its forward unconditionally dereferences spike_fn.required_states, so it cannot be run with spike_fn=None, and its output is always spikes, never v_mem.

• Does this PR introduce a breaking change?

No. Default min_v_mem is None (no-op). For spiking layers the stored state and spike output are unchanged. The only behavioural change is that non-spiking layers (ExpLeak, LIF/IAF with spike_fn=None) now return outputs >= min_v_mem, which is what the min_v_mem docstring already promises.

• Other information:

Diff: sinabs/layers/functional/lif.py +19 / −4 (mostly the moved block + comments), tests/test_layers/test_lif.py +41.

New tests in tests/test_layers/test_lif.py:

• test_min_v_mem_clips_the_output_without_spike_fn — ExpLeak / LIF(spike_fn=None) / IAF(spike_fn=None) output >= min_v_mem. Fails on develop (output -1.48 fell below min_v_mem=-0.5), passes here.
• test_min_v_mem_does_not_change_the_spiking_path — spike train bit-identical with and without min_v_mem on a spiking input.

black --check sinabs/ tests/ → 135 files unchanged. pytest tests/ (non-hardware modules) → 197 → 199 passed, no regressions.

Before (develop, #236 present)

After (this branch)