Closes the open finding from PR #33's Track 1 verification spike.

Before this PR, analyze_core.analyze_loudness instantiated
es.LoudnessEBUR128() with no sampleRate argument (Essentia default
44100), but the full pipeline at analyze.py:1497 passed in a stereo
array from load_stereo at the source file's native rate (no upstream
resample). Any source above 44.1 kHz got K-weighting tuned for the
wrong rate.

This PR:
- analyze_loudness now takes sample_rate: int = 44_100 and threads it
  to LoudnessEBUR128(sampleRate=sample_rate).
- analyze.py:1503 (full pipeline) passes sr from load_stereo through.
- analyze.py:1242 (stem path) hardcodes 44_100 because Demucs writes
  stems at 44.1 kHz regardless of source rate.
- New regression test asserts -23.0 ±0.1 LUFS through analyze_loudness
  at 48 kHz.
- Three mock-based wiring tests (per review feedback) assert
  sampleRate=<value> reaches Essentia, including the verbatim case
  for 96 kHz.
- Docstring on the spike outcome doc marks the finding as resolved.

https://claude.ai/code/session_01Uh2eNT1oqjixcox8UX3BTg

…nt (#43)

* feat(spectral): librosa.reassigned_spectrogram as an opt-in enhancement

Last item from the merged external-repo review's adopt list. Adds a
new on-demand spectral-enhancement kind, 'reassigned', that produces
a sharper spectrogram via per-bin (time, frequency) reassignment.

Why reassignment: vanilla STFT spectrograms smear transients across
the analyzing window (~46 ms for n_fft=2048 at 44.1 kHz). Reassignment
relocates each STFT bin to the local centroid of energy in the
(time, freq) plane — transients land at their true onset and stable
partials collapse to sharp horizontal lines. Useful when a producer
is trying to read attack timing or harmonic detail from the existing
mel/STFT views and finding them too blurry.

Implementation:

- apps/backend/spectral_viz.py — new generate_reassigned_spectrogram
  function. Calls librosa.reassigned_spectrogram with both frequency
  and time reassignment enabled, fill_nan=True (handles silent-region
  numerical instability without dropping points), and renders via
  matplotlib scatter (not librosa.display.specshow — specshow would
  re-rasterize to a uniform grid and undo the sharpening). Magnitude
  floor of -60 dBFS below peak filters the noise wash without losing
  attack detail.

- apps/backend/server.py — registered as 'reassigned' in the existing
  _ENHANCEMENT_GENERATORS dict. Inherits all existing wiring: auth,
  measurement-completed precondition, idempotency, artifact storage,
  MIME types, JSON response envelope.

- apps/backend/tests/test_spectral_viz.py — 3 new tests:
  - produces_single_reassigned_png: shape + non-trivial size
  - output_is_valid_png: magic bytes
  - handles_silent_input_without_raising: regression gate against
    NaN propagation on near-silent fixtures

No new dependencies — librosa is already in requirements.txt
(librosa==0.11.0 exposes reassigned_spectrogram).

Usage:

  POST /api/analysis-runs/{run_id}/spectral-enhancements/reassigned

Returns the new 'spectrogram_reassigned' artifact ID alongside the
existing PNG-style spectrograms. Idempotent: re-posting returns the
existing artifact without regenerating.

Out of scope (deferred):

- Frontend UI toggle to display the reassigned PNG. The backend now
  produces it on demand; the UI wiring is a separate follow-up.
- Raw (time, freq, mag) JSON artifact for client-side custom
  visualization. Could be added as a sibling 'reassigned_data'
  artifact if a future consumer needs it.

Closes the last review-adopt item. Full session summary in the
merge commits for #33 through #42.

https://claude.ai/code/session_01Uh2eNT1oqjixcox8UX3BTg

* perf(spectral): cap reassigned-scatter at 300K points

PR #43 review caught a real perf issue: the original
generate_reassigned_spectrogram fed every STFT cell into
matplotlib.scatter unchecked. At default params (n_fft=2048,
hop=1024, sr=44100), that's:

  ~8M raw points on a 3-minute track
  ~15M raw points on a 10-minute track

Both well above the floor-mask survival count and well above the
threshold where matplotlib's render step starts taking tens of
seconds. The librosa compute cost isn't the bottleneck — the
scatter render is.

Fix:

- New REASSIGNED_MAX_SCATTER_POINTS = 300_000 constant. At
  FIG_DPI=100 / ~1200x400 px figure, 300K is the visual sweet
  spot — sharper-looking than vanilla STFT, renders in under a
  second.

- After the floor mask, the post-mask arrays (times_visible,
  freqs_visible, mags_visible) get subsampled via a seeded RNG
  (np.random.default_rng(0)). Seeded so re-running on the same
  audio produces a byte-identical PNG — matters for idempotency
  tests and content-hash artifact caching.

- New test test_scatter_point_cap_is_enforced_for_long_inputs:
  generates a 60 s multi-component fixture (~1M+ post-mask points,
  comfortably above the 300K cap), patches matplotlib.scatter to
  capture call args without rendering, and asserts the actual
  point count passed to scatter is <= REASSIGNED_MAX_SCATTER_POINTS.

The 'inferno colormap' worth-noting from the review was checked —
no 'inferno' reference exists in CLAUDE.md or anywhere in the
codebase. The reviewer's worth-noting was itself stale; nothing
to fix there.

https://claude.ai/code/session_01Uh2eNT1oqjixcox8UX3BTg

---------

Co-authored-by: Claude <noreply@anthropic.com>