sync_pilot/dashboard/source_registry.py

"""Single source of truth for which inference tag sources are surfaced in the UI.

Both the Track Explorer (display) and the GT Review board (seed values) consult
this registry so the two views stay in sync: phasing out an inference path is a
single edit here, after which it disappears from both views with no tag deletion
and no per-view cleanup.

Keys are TrackRecord tag categories (``genre``, ``instrument``, ``vocal``,
``mood``, ``theme``); values are the active tag sources in seed priority order
(first entry is the primary source for GT Review). A category absent from the
registry is left unfiltered, so unlisted categories preserve prior behavior.

This is UI/display-facing and intentionally separate from
``taxonomy/source_policy.py``, which is the prompt-facing evidence trust policy.
"""

from __future__ import annotations

ACTIVE_INFER_SOURCES: dict[str, list[str]] = {
    "genre": ["genre-fusion-v1", "muq-probe-v1", "taxonomy-adapter-v1", "maest"],
    "instrument": [
        # CLAP-on-stem detections (demucs): rich folk-instrument vocab + real
        # confidence — preferred when present (live demucs path), else the probes.
        "stem-window-v1",
        "muq-probe-v1",
        "taxonomy-adapter-v1",
        # Expanded-GT probes carry instrument coverage on tracks that lack a
        # canonical muq-probe-v1/adapter instrument tag (e.g. 038_dert_olur).
        "muq-probe-expanded-v1",
        "passt-probe-expanded-v1",
    ],
    # autoresearch promotion (2026-06-23): taxonomy-adapter-v1 first — its
    # vocal_configuration heads score f1 0.778 vs muq-probe-v1's ~0.12 on the 16
    # reviewed Median GT tracks (the champion vocal win). muq-probe-v1 stays as a
    # fallback for tracks the adapter leaves untagged.
    "vocal": ["taxonomy-adapter-v1", "muq-probe-v1"],
    # autoresearch promotion (2026-06-22): lyla-text-clf first — it scores mood
    # f1 0.731 vs mood-fusion-v0's 0.629 on the 16 reviewed Median GT tracks. lyla
    # is lyric-only, so when a track has no lyrics _pick_tags falls through to the
    # audio+lyric fusion automatically (sensible fallback).
    "mood": ["lyla-text-clf", "mood-fusion-v0"],
    # autoresearch promotion (2026-06-23): theme-fusion-v0 first — it unions the
    # high-precision lyrics-llm themes with lyla recall recovery (score>=0.45),
    # scoring f1 0.732 vs lyrics-llm's 0.591 on the 16 reviewed Median GT tracks.
    "theme": ["theme-fusion-v0", "lyrics-llm", "lyla-text-clf"],
}


def active_sources(category: str) -> list[str] | None:
    """Active sources for a category in seed-priority order, or ``None`` when
    the category is not registered (callers should then apply no filtering)."""
    return ACTIVE_INFER_SOURCES.get(category)


def is_active(category: str, source: str | None) -> bool:
    """Whether ``source`` is active for ``category``. Unregistered categories
    return ``True`` so unlisted categories are not silently hidden."""
    allowed = ACTIVE_INFER_SOURCES.get(category)
    if allowed is None:
        return True
    return source in allowed