v0.4.0

Credit is no longer given to a probe that has stopped working. The mechanism comes
from the sibling project NeverDry, which already judged a probe's silence against
the probe's own habits rather than against a constant, and it arrives here tuned
the other way round and with a second channel NeverDry does not have.

The tuning is reversed because the cost is. There, refusing a live probe hands a
zone to a weather estimate on a different scale, silently, every night, which is
worse than a day of delay. Here, a refusal is named on a diagnostic built to be
read, while believing a dead probe puts its readings into the published line and
leaves them there after the probe is fixed. So this side errs towards refusing:
the ceiling on the freshness bar is six hours where NeverDry's backstop is
twenty-four, and a probe that has not yet demonstrated a cadence is judged by the
configured timeout rather than by the ceiling.

The warning of 0.1.0 and 0.3.0 still stands: the error budget is argued from soil
physics and not yet measured against a probe in the ground, and the placement
thresholds are argued rather than measured. The three thresholds added here are
argued too, and docs/design/probe-liveness.md says from what.

Added

• Liveness is now the device's, not one channel's. The silence that decides
  whether a probe may be believed is measured across every entity of the probe's
  device, and no entity is inspected for what it says: that it spoke at all is the
  evidence. Home Assistant writes a sensor's state only when its value changes, so
  a probe on ground that is not moving publishes nothing on its moisture entity
  while its temperature, battery and link quality carry on. Judged on one channel,
  a live probe is declared dead for reporting the same number twice. A probe whose
  entity belongs to no device keeps the temperature channel it always had.
• The bar a silence is judged against is learned from the probe. The longest
  quiet the device has come back from inside a trailing week, doubled, capped at
  six hours. A maximum and not a quantile: a device reporting on change produces
  many short gaps while its readings move and a few long ones that are the only
  evidence of its heartbeat, and any statistic that lets the first outvote the
  second sets the bar below the heartbeat. The learned bar takes over only after a
  full day of watching and four times the longest silence seen; until then the
  configured Probe considered offline after value applies, exactly as before,
  so nothing changes for a new installation or for one that tuned that number. The
  evidence is persisted, because the other half of the judgement is rebuilt from
  the states for free and restoring half a judgement is worse than restoring none.
• A second channel for the electrode that stops while the device keeps talking.
  The radio is fine, the battery reports eighty percent, the temperature follows
  the day, and the moisture reading is the same whatever the soil does. No silence
  exists, so no measure of silence can find it. What finds it is the reference: half
  the reservoir of drying with the index not moving half a point. Samples are
  refused with their own reason, a repair says the device is talking and the reading
  is not, and the fall of a deficit resets the stretch rather than completing it,
  because water arrives as a discontinuity and a delivery would otherwise condemn
  any probe that had not refreshed inside the poll interval.
• A battery floor, five percent by default and configurable, zero turning it
  off. The earliest indicator of both failures above, and a sagging supply shifts
  the index without the soil shifting. An exact zero is treated as no information
  rather than as an empty battery: a device that is talking cannot truthfully be at
  zero, so zero is a device reporting its battery badly, and refusing every sample
  from one of those would cost the calibration for a channel the calibration does
  not use.
• docs/design/probe-liveness.md: the cost asymmetry that sets every threshold,
  the alternatives rejected and the numbers behind them, and what remains wrong.

Changed

• A probe that has stopped measuring is no longer diagnosed as a probe in the
  wrong place. The two produce the same signature, an index that does not move,
  and they do not have the same repair: every placement message ends in some form
  of "consider moving the probe", and a user who digs one up finds it just as
  motionless in the new hole, which reads as confirmation. The placement module's
  no response suspicion is therefore withheld while the electrode is stalled,
  and the reason for withholding it is recorded in the placement entity's evidence
  so a verdict that raised nothing can be told from one that was silenced. Nothing
  takes its place there: a dead electrode is not a placement fault.
• Probe online publishes the silence and the bar it was judged against
  together, plus the entities enrolled as witnesses. Both numbers or neither: an
  age without its bar invites the reader to compare it with a number they invented.
  It stays on for a stalled electrode, because that probe is online; the stall is
  carried by Calibration problem, with the drying travel the index failed to
  respond to.
• The diagnostics download carries the cadence evidence and the witness state
  alongside the samples. The first question about a suspect calibration is whether
  the probe was alive while it was collected, and that cannot be reconstructed from
  the samples afterwards.