[steveclarke]

Two bugs on the same path. #31 is why #29 went unnoticed, so it lands first and
proves #29. Rebased on master after #36.

#31 — writes reported success on send

set and the channel/mix/send verbs printed ✓ set <path> = <value> as soon as
the frame went out. The board silently clamps or rejects a value it will not
accept, so that line described the request, not the result.

Every write now re-reads its paths on a fresh connection and reports what the
board holds. A write that did not take names both values and exits 1:

✗ set line/ch3/filter/hpf: wrote 90, board holds 1

The read-back needs a fresh ZB, so the writing connection closes first — the
in-session read-back quirk means it cannot confirm its own writes. It costs one
extra connection per command, not per write, and a genuine mismatch retries up to
5 times before it reports. On a board with other UCNET clients on it that is real
contention, which is what --no-verify is for.

The fake board now clamps PV writes to 0..1 for known controls, as a real 32R
does. Without that it stored 90.0 for a 0..1 control and the read-back agreed
with itself — the missing fidelity that let this class of bug reach a live desk.

#29 — HPF sent Hz to a control that wants 0..1

taper.HPF is now the log sweep the board follows:

pos = ln(Hz / 24) / ln(1000 / 24)

Calibrated against a 32R holding a configured show: 0.35438603 is 90 Hz,
0.13696200 is 40 Hz, 0.38263556 is 100 Hz. Three positions landing on
frequencies a human would dial.

The three spellings of one state are reconciled: position 0 is both the bottom of
the sweep and how the board stores a filter that is off, so hpf: off, hpf: 0
and hpf: raw:0.0 all compile to 0 (hpf: 0 used to be rejected as "0 Hz out of
range"). Validation moves from a generic 20–20000 band to the board's real
24–1000 sweep. dump --as-config emits Hz instead of a raw position, so a dumped
config round-trips through apply and verify clean.

Humanized reads round to one decimal

The first CI run was green on darwin/arm64 and red on linux/amd64: a
Hz → position → Hz round trip is a log and an exp, so the last bits are the
platform's libm and an exact float compare could only ever be arch-dependent.

Fixed at the representation rather than in the assertion, the same choice #36
made for color. Inverting a 32-bit wire position leaves noise — get printed
90.00000306137774 for the filter and -6.000000847568462 for a fader, the
latter predating this branch — so humanized reads now round to one decimal on the
way out. --raw still shows the exact position and Client.Get still returns the
exact inversion; only presentation rounds. The rule already existed as
boardconfig's round1; it moves to internal/schema next to Tolerance so the
config writer, get, dump, the drift table, and the read-back cannot disagree.

The settledValue test now compares within the control's tolerance — the band
the read-back comparison already uses — and says why, so it is not re-tightened.

Confirmed on hardware

Run against the 32R at 192.168.77.82, targeted writes to line/ch1/filter/hpf
only, read back and restored each time:

Step	Result
set line/ch1/filter/hpf 90 (before the fix)	board holds 1 — the bug, reproduced
set line/ch1/filter/hpf 0.35438649	board holds 0.35438639 (float32 of the same position)
channel 1 hpf 90Hz (before the fix)	✗ set line/ch1/filter/hpf: wrote 90, board holds 1, exit 1
channel 1 hpf 90Hz (after the fix)	✓ set line/ch1/filter/hpf = 90, wire 0.35438639, exit 0
channel 1 hpf 2000Hz	refused: taper: value above control range: 2000 Hz > 1000 Hz

The HPF positions already on the desk (0.35438603 on ch3/ch4, 0.13696200 on
ch7) invert to exactly 90 Hz and 40 Hz — the strongest confirmation of the curve.

ch1 was the only path written and is back at 0, its original value. Confirmed
independently: the live board still verifies clean against the show config.

Not touched

• apply needs no change: it never goes through this write path and already
  runs verifyAfterApply on a fresh connection. A per-write read-back there would
  duplicate it.
• fade is out of scope — it ramps a value rather than setting one.

Tests

490 unit (was 373) and 86 e2e (was 72), plus lint, vet, race, govulncheck, fmt
and tidy.

Closes #29
Closes #31

https://claude.ai/code/session_01Pgcg59QJrz9BCmeEyb4B5U