Same move as Timers 1.7.0. The capture tool existed only where the macOS
app build had compiled it, with the module trying swiftc on the booth as a
fallback — NA has been logging "spawn spl-capture ENOENT" since its
catalog install. Under the module guide's rule it now ships built:
bin/spl-capture-macos-arm64 and bin/spl-capture-macos-x86_64, signed with
the Developer ID (hardened runtime, audio-input entitlement), from
native/spl-capture.swift via native/build.sh. captureTool() picks the
binary for this platform; where there is none — Windows, until a Windows
implementation exists — measurement reports itself off in the health line
and the device picker while the module stays mounted. The runtime build
and its "building…" state are gone, and the old top-level binary is no
longer gitignored.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Same deadlock the Timers module hit live on AB (fixed there in 1.5.1):
start() skipped scheduling its retry while a build was in flight, and
ensureCaptureBinary's onReady only fires on build SUCCESS. A failed
build therefore left the listener waiting forever on a callback that
never comes — "Building the capture tool…" indefinitely — and a binary
scp'd in from outside (the required path on AB, whose CLT cannot
compile) was never picked up without a module remount.
onReady stays as the fast path; the 30-second retry is now armed in
every not-ready state, so a failed build keeps re-checking and an
externally supplied binary is found within half a minute.
Reproduced pre-fix (build fails at 0s, binary copied in at 3s, still
untouched at 38s) and verified post-fix (retry at 30s spawns it).
spl-capture --selftest unchanged against the IEC 61672 table.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
NA could not list a single audio interface. The module was installed from
the catalog, which ships source only because the capture binary is
gitignored, and the picker shells out to that binary with --list. So it
returned nothing, forever:
[spl-meter] device list failed: spawn .../spl-capture ENOENT
The missing-binary build existed, but only inside the listener, and the
listener only starts once a device has been picked. No tool meant no device
list, and no device meant no listener to build the tool. Nothing was going
to break that on its own.
The build now lives at module scope and both callers use it: init() asks for
it on mount whether or not a device is configured, and /devices asks again
each time it is called, so opening the admin page is enough to start it. It
is still a plain spawn that never touches the event loop, and health() now
names the situation instead of leaving an empty dropdown unexplained.
Also removed two per-listener build flags left shadowing the module-scope
ones, which would have made the "is a build already running" test always
false and let a second build start under the first.
Tested the way it actually failed: a data directory holding only the six
source files, no binary. The build starts on mount and /devices answers with
the real device list moments later.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The spectrogram scrolled by copying the visible canvas onto itself one step
along. That canvas already held the frequency labels and gridlines, because
they were drawn straight onto it, so every frame smeared them sideways and
left a trail. Arrived with the labels in 0.3.0; the gridlines did it too.
Only measured pixels belong on a surface that scrolls. The history now lives
on its own offscreen canvas that holds nothing but spectrum, and the visible
canvas is repainted from scratch each frame — history first, annotations
over the top. Nothing drawn as an annotation can reach the scroll buffer by
construction, rather than by remembering to clear the right strip.
The layout was muddled in the same place and is fixed with it: the gutter
used to be reserved along the bottom while the labels were drawn down the
left, overlapping the plot. Labels now sit beside the axis they describe —
down the left when frequency runs vertically, with room for "12.5k", and
along the bottom when it runs horizontally.
Verified both orientations by fingerprinting the gutter and the plot two and
a half seconds apart: gutter identical, plot changed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
AB's meter appeared to halt for a day and a half. Nothing had halted: the
capture child had run continuously since the machine was restarted, never
stalled, never exited, and was delivering twenty frames a second. Every one
of them was digital silence, because Dante Via stopped passing audio after
the reboot — the Timers module's LTC listener, on channel 16 of the same
device, had been reporting no signal for exactly as long.
The module's part in that was reporting "ok" through all of it and drawing
a steady -83, which is the -200 floor plus the +117 calibration offset. A
plausible number on a wall is believed, and a green dot beside it says the
number is worth believing. That is worse than showing nothing.
So silence is now a state the module names. Below -150 dBFS unweighted the
input is not quiet, it is absent: a microphone in a silent room still reads
about -83, and even a muted converter dithers. Five seconds of that on every
channel and health() goes to error naming the device and how long it has
been dead, the tiles stop drawing a level and say "No audio" with how long,
and the status dot goes red.
Verified both directions, which is the half that matters: BlackHole with
nothing routed into it is caught in five seconds, and a real microphone in a
quiet room at -49 dBFS is never flagged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deploying 0.3.0 to AB crashed the ProdDash server on restart:
TypeError: Cannot read properties of null (reading 'streams')
at server.js:273 req.on('close', () => state.streams.delete(res))
teardown() nulls `state` and then ends every open stream, which fires each
request's close handler — and the handler reached back through the `state`
it had just discarded. Any re-init with a browser connected would do it, so
saving admin config while a dashboard was open was enough to stop the whole
server, every module with it.
The stream handler now closes over this mount's Set rather than the module's
state: after teardown it deletes from a set nobody reads, which is harmless.
The same hazard existed anywhere a callback could outlive the mount — the
keepalive ping, a frame arriving from the listener mid-teardown, and the
/measure interval — and each is closed over or guarded now.
Reproduced first (open an SSE client, save config, watch it fall over) and
confirmed fixed against the same sequence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Rebasing onto main landed rules 5 to 11, which this module predates and
broke in three places.
Rule 5 — no explanatory text in a tile's gear: the Banding and Smoothing
fields each carried a paragraph. The Banding one said something true and
worth knowing, that below roughly 200 Hz the window length limits real
resolution however fine the setting, so it moves to the source where the
band plan is built rather than being lost. Option labels shortened with it.
Rule 11 — one control per action: the spectrogram offered its axis both as
a gear field and as a title-bar button. The button stays, since that is
what was asked for, and the field is gone. The setting still persists with
the tile; saveInstanceSettings keeps keys that no schema declares.
Rule 9 — title-bar controls are icons: the axis button was drawing a "⇆"
as a label. It is an inline SVG now, with the tooltip carrying the words
and naming the current direction rather than the action.
Rule 8 also asks that a tile can switch off what it draws, so the readout
gained a Caption toggle and both graphs gained Labels.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The number was capped at 6rem, so a readout given half a wall stopped
growing at the size it reached in a small tile.
Removing the cap was not enough on its own. Sizing text off container units
means choosing a coefficient for some assumed number of characters, and the
value here is not a fixed width: the coefficient that fills a tile nicely
for "-45.9" pushes "-100.5" nine percent out through the side. Measured it
rather than guessed — the tile now sets a trial size, reads the width the
actual text wants, and scales to fit. That runs only when the tile size or
the length of the reading changes, not on every frame.
It now fills about 98% of the width wherever width is the limit and 92% of
the height wherever height is, at every tile size from 180x70 up to
1400x700, with no overflow on either axis. The CSS keeps a conservative
size for the first paint before the measurement has run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every control asked for here is a view choice, so none of them may reach
the capture child — two dashboards watching the same room must be able to
disagree about how to draw it without disturbing the measurement or each
other. That set the shape of the change: the child now sends everything
anyone might ask for, once, and each tile folds it down.
Averaging (was "Reading") is Fast, Slow, Inf and None, and the child
computes and sends all four every frame. Fast and Slow are the IEC time
constants, Inf is Leq since reset, None is the block with no time constant
at all. The admin-wide time constant is gone — it was never a property of
the room.
Banding is 1/3 through 1/48 octave. The wire always carries 1/48, and
because every banding offered divides 48 a coarser band is an exact sum of
finer ones in the power domain. Nothing is resampled and no tile's choice
is visible to another. The transform grew to 16384 points to have bins fine
enough to mean it; below roughly 200 Hz a 1/48 band is still narrower than
one bin, which the Banding help says out loud rather than pretending.
Each tile kind now carries its own settings, which is the fix for the
complaint that started this: the RTA was offering an alarm threshold and a
metric picker that did nothing to it. An entry's instanceSchema replaces the
module's, so a gear popover only holds controls that change that tile.
Also: frequency labels in three tiers that thin out as a tile narrows
instead of smearing, gridlines off by default on both graphs, RTA smoothing
as a damping factor on the drawn value only, a title-bar toggle to swap the
spectrogram's axes, and the threshold blurb removed.
Settings saved by 0.1.0 are mapped forward — a layout on a wall should not
start reading dashes because a field was renamed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A sound level meter that lives in ProdDash rather than beside it. Takes an
audio input the way the Timers module's LTC reader does, and offers four
tiles: the level as a big number, per-channel bars with peak hold, a
third-octave RTA, and a scrolling spectrogram.
The measurement happens in spl-capture.swift, not in Node and not in the
browser. LTC gets away with piping raw samples upstream because the decode
is cheap and single-channel; sound level is filtering and FFTs over every
channel at the device's rate, so the child does the work and only the
answers cross the pipe. That costs nothing in dependencies — Accelerate
ships with macOS, and the whole thing builds in two seconds with a plain
swiftc, exactly like ltc-capture.
The weighting filters are the part worth checking rather than trusting, so
`spl-capture --selftest` prints them against the IEC 61672 table. A- and
C-weighting land within 0.15 dB from 31.5 Hz to 4 kHz and exactly 0 dB at
1 kHz. The top octave is where the bilinear transform runs out of room: at
48 kHz the 16 kHz point sits about 3 dB low even with the pole pair
prewarped, and at 96 kHz that falls to 0.6 dB. It is a rounding error on a
broadband number and a real limit all the same, which is why the selftest
prints both rates instead of a pass mark.
Cross-checked against REW on the same microphone: after removing the fixed
calibration offset the two agree to a standard deviation of 0.5 dB on live
room noise.
Uncalibrated is a state the module says out loud — every tile reads dBFS
and labels it, so nobody mistakes a relative number for sound pressure. The
/measure route averages the calibrator tone in the power domain and hands
back the offset to paste into settings; a module cannot write its own admin
config, so that last step stays manual on purpose.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
