The FastAPI backend kept one analysis session shared by every visitor and
spooled each upload to the server's /tmp, and the browser copied audio into
persistent OPFS storage. None of that fits a tool for unreleased masters, so
the web UI is now a static site that decodes and analyses in the tab:
- web/decode.js: WAV/AIFF PCM and float read directly with libsndfile's
scaling (bit-identical to librosa.load; Chromium's WAV decoder scales
positive int16 by 1/32767), other formats via decodeAudioData on an
OfflineAudioContext at the header-sniffed native rate. ID3 song names.
- web/dsp.js: ports of every Metric.compute with the same numerics
(pyloudnorm K-weighting via lfilter's DF2T recurrence, BS.1770 gating,
EBU LRA, resample_poly's Kaiser filter for true peak, librosa.stft and
amplitude_to_db), streaming passes with no signal-length intermediates.
- web/metrics.js: build_spec ports producing the existing plot payload.
- web/analysis-worker.js + app.js: per-tab track list and result cache, a
worker job queue where the visible metric jumps background prefetch, and
the renderer, which now also draws the metrics' own reference lines.
Deployment is nginx-unprivileged serving web/ read-only with
Content-Security-Policy connect-src 'none', so the page cannot make network
requests at all, and no access log. No secure context is needed, so plain
http on a LAN works. The Python web backend, its smoke test, the web extra
and the OPFS store are removed; the desktop app is unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015UXA39Sh9coAx4FhRkd8R2
Multi-stage image built from the `web` extra only, so it carries no Qt: the
whole web path (webapp, web_session, metrics_json, and the shared analysis
modules) imports no Qt, pyqtgraph or matplotlib canvas. The build stage has the
compiler scipy/numba need; the runtime layer installs the built wheel, so the
same image works from a checkout or a release tarball.
Compose declares no audio volume on purpose. The app keeps no audio, so there is
nothing to persist: `read_only` plus a /tmp tmpfs means request spools are
memory-backed and never reach a disk. Publishing is loopback-only because the
tool has no authentication, and BLAS thread pools are pinned because analysis is
CPU-bound.
Fix three packaging bugs the build surfaced, each of which broke the image:
- `pip install --extra web` is not valid for a source tree; uv wants `.[web]`
and failed with "Requesting extras requires a pylock.toml".
- The wheel shipped no static assets at all. `package-data` keys match
*packages*, and the app modules are top-level `py-modules`, so the `"*"` key
matched nothing - verified by inspecting the built wheel: 14 .py files, zero
HTML/JS/CSS. An installed `ujm-web` would have 404'd every asset. `static/` is
now declared a package (hence its packaging-shim `__init__.py`) so the assets
install next to the modules, which is where `webapp.py` looks for them.
- The Dockerfile's curated COPY list omitted modules that `py-modules` requires,
failing the build with "file <module>.py not found". It copies the checkout,
filtered by .dockerignore.
Move PyQt5/pyqtgraph into a `desktop` extra so the web install is genuinely
Qt-free. Verified: `find` for Qt in a web-only venv now returns nothing, and the
desktop path still resolves and imports.
Add scripts/make_fixtures.py so the smoke suite has no network or ffmpeg
dependency (the tagged MP3 is optional and its assertions skip when absent), and
make the suite tolerate that. Verified against the *installed package* running
the container's own command line, from a directory with no source tree:
61/61 passing.
Also correct README claims that had gone stale: BPM (removed), the matplotlib
canvas (replaced by pyqtgraph), and font_control_widget.py (deleted).
mikkeli
irene