feat: render with ffmpeg alone, from and to URLs the caller presigns

[rainhead]

The renderer is now one ffmpeg command and the function no longer knows what S3 is. The image drops from 2.27 GB to 773 MB, there are no caches to seed, and a render takes 2 s at 1024 MB with a cold start of about a second. On a real production segment the output has production's width exactly (1877×512) and its luminance distribution to within 0.01.

Before and after

Production's image of rpi_orcasound_lab/hls/1790665213/live5238.ts (left) and this branch's (right), a 420-column crop from the middle of the segment. Production is a touch softer from its Lanczos resampling; the floor colour, feature colour and events are the same.

Luminance 10th/50th/90th percentiles: 0.193/0.199/0.323 here against 0.200/0.200/0.324 in production; block-averaged (32 px) correlation 0.97.

How it renders

app.py: showspectrumpic draws the spectrogram as intensity, then pseudocolor applies matplotlib's viridis (showspectrumpic's own viridis starts at black, so it cannot match the archive). The window size follows from the height (n_fft/2 rows) and the width is one column per hop_length samples, sized from the samples actually decoded rather than the container's duration, which undercounts an AAC stream by a frame or two (that cost 16 columns on the segment above before I caught it). The archive's dB window is relative to an amplitude of 0.01 in librosa's STFT; DB_OFFSET converts it to ffmpeg's full-scale reference and was fitted against the production image.

Contract

Request: audio_url (presigned GET), image_url (presigned PUT, or null to render and discard), and optional parameters with the archive's values as defaults. Response: image_size, sample_rate, width, height, the parameters applied, and the field names orcasite has stored since the first renderer. The previous audio_bucket/audio_key/image_bucket/image_key shape still works through the boto3 the base image ships, so this can deploy before orcasite changes; the S3 policy in template.yaml stays until then and goes with it.

The orcasite side (presign both URLs, pass parameters, find the function by config instead of listing every Lambda, delete server/audio_viz) is a separate PR there.

Verification

tests/smoke.py renders a synthetic AAC/MPEG-TS clip through the handler inside the built image as Lambda runs it (unprivileged, root-owned empty /tmp, 1024 MB tier); CI runs it. The comparison above was made the same way with the real segment through lambda_handler.

🤖 Generated with Claude Code

[rainhead]

@dbainj1 and @scottveirs, I broke this component out of the orcasite codebase. Claude then proposed to removing the Python-based orchestration around rendering, to slim it down to its ffmpeg bones. The result is a little different. Is the difference acceptable?