Use None, not inf, for "no limit" resample settings

[cboulay]

Why

float("inf") is not valid JSON. Any tool that serializes a Settings object trips over it — ezmsg-dashboard#4 hits this on our defaults, so it fails before the user has touched anything.

A sweep of the repo turned up exactly one Settings field defaulting to infinity, declared twice:

Location	Field	Was
resample.py	ResampleSettings.max_chunk_delay	np.inf
resampleconcat.py	ResampleConcatSettings.max_chunk_delay	np.inf (forwarded)

Plus one near-miss: reference_reset_after_chunks defaults to a finite 3, but its docstring told users to pass float("inf") to disable reset recovery — straight into the same bug. Two tests did exactly that.

What

• max_chunk_delay: float = np.inf → float | None = None (no limit — never extrapolate on wall-clock alone).
• reference_reset_after_chunks: float = 3 → int | None = 3. Default behaviour unchanged; only the "disabled" spelling moves from inf to None.
• Both mirrored on ResampleConcatSettings; its now-unused numpy import is dropped.
• New _as_limit() helper folds non-finite values onto None, so pipelines already passing inf keep working rather than breaking on upgrade.
• The publisher's wait loop drops np.isfinite(timeout) for a plain is not None branch.

None also matches the convention already used across this repo for optional numerics (resample_rate, downsample.target_rate, align.gain, kaiser.ripple), and gives a dashboard an unambiguous "unset" to render.

Considered and rejected: a large finite sentinel

float32 max (~3.4e38) was the other candidate. These are float64 scalars in wall-clock-seconds arithmetic, so it's an arbitrary magic number that still renders as garbage in a UI field — and it would silently flip a code path, sending the publisher into asyncio.wait_for with an absurd deadline instead of the plain event wait.

Not changed

ResampleState.last_write_time keeps -np.inf. It's state, never serialized into a settings payload. The remaining inf hits in the repo are a np.nextafter direction argument, a benchmark min-tracker, and two comments — all correct as-is.

Testing

test_resample_reset_disabled_can_stall is parametrized over [None, float("inf")] so the back-compat path stays covered. The integration test and its docstring move to None.

• ruff check clean
• 4050 unit tests pass (6 skipped)
• 7 resample integration tests pass
• Verified both Settings classes now serialize under json.dumps(..., allow_nan=False), which previously raised ValueError: Out of range float values are not JSON compliant

Note for the dashboard side

This fixes the defaults. A user can still type inf into a float field at runtime, and this repo will accept it (normalizing internally) — so if ezmsg-dashboard serializes settings before handing them to the processor, it still needs its own guard.