Generic error strings hide actionable causes #294

Location: runner.py:201
Severity: LOW

“Audio processing failed. Please try again.” is shown for OOM, bad file, GPU fault, and disk-full alike. The stderr tail is already captured in separate() — it just never reaches the user or the error record.
Store the classified cause in Job.error_detail (shown behind a “details” disclosure — error reporting, not a feature). Classify at least: out-of-memory, unsupported-device, disk-full, bad-input.
Any failure is answerable from disk, after the fact, without a reproduction: rotating stemdeck.log (server) + non-truncating backend.log (desktop) + jobs/failed/<id>/error.txt with device, model, timings, and stderr tail.
StemDeck runs htdemucs_6s, fixed, with default settings (no --shifts, default overlap). Assessment of where that sits in 2026:
Verdict: yes — and your instinct about the architecture is exactly what the field converged on. No single model splits a full mix into lead + backing well; the established approach is a second pass over the vocals stem with a “karaoke” (backing-vocal extraction) model:
Phases 1–3 are pure hardening — no feature surface changes, immediately shippable one PR per finding. Phase 4 is the state-of-the-art jump, and lands on infrastructure phases 1–2 make debuggable.

From the health report. Part of #272 (Phase 1 — Evidence).