Fix the regression introduced by the weighted Laya signal (#16):
rebuild now merges exact fingerprints in every engine mode after
evaluating hard blockers, compares candidates against all variants
of a group, falls back to deterministic scoring when Laya is
unavailable, and reports transaction errors safely. Non-regression
tests cover the Laya-mode rebuild path; verified against a real
catalog copy (39,341 streams rebuilt successfully).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Introduce a liveGroupingScoring engine that combines existing
membership signals (normalized name, brand, number, metadata tokens,
language, category/source when relevant) with a bounded Laya
contribution: Laya raises or lowers confidence instead of acting as
a final judge, with guard-rails against obvious contradictions.
Sanitize Laya batch identity inputs in layaService and wire the
weighted scoring into the live projection service. Weights are not
exposed in the Nodecast UI. Cover signal combination, bounds and
contradiction guard-rails with tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
beIN Sports 1 variants stayed separated after projection because
names with region prefixes (C Car, C Reu, Thai), time-shift suffixes
(2h/4h/8h/12h) and language markers (En, Fr, Thai) produced distinct
fingerprints. Extend normalization in the server projection service
and the frontend grouping fallback to fold these into one logical
channel, with tests on the exact reported names. Clarify the Laya
language detection warning so a language-matching failure is not
mistaken for a grouping failure, and bump script versions for cache
refresh.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a Settings action to manually rerun Live TV stream analysis:
it triggers an admin-guarded projection rebuild endpoint that
groups streams into logical channels (covering channels present
before the projection existed) and reruns language/Laya detection,
with progress presented via liveAnalysisPresentation. Harden the
Laya batch path that failed with 'Local AI language matching
failed' by surfacing the underlying cause and falling back to
standard matching cleanly. Guard admin-only rebuild through a new
authz helper and cover routes, service and presentation with tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Move variant grouping work out of Live TV rendering into the
catalog pipeline: a liveProjectionService builds the logical channel
projection (schema in liveProjectionSchema) whenever a source is
added, synced or modified (name change, new address), and the
projection is served through a /live route. ChannelList now consumes
the precomputed projection via a liveChannelProjection client
service, recalculating only when the catalog changes. Cover
projection building, incremental recompute and client consumption
with tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
