perf(api): revert TrendingJob to bulk DELETE+INSERT score rewrite

[raymondjacobson]

Summary

• perf(api): stop TrendingJob from continuously rewriting score tables #890 swapped the trending score-table rewrite for a temp-table stage + INSERT ... ON CONFLICT DO UPDATE (skip-unchanged) + anti-join prune, to avoid rewriting the ~16GB *_trending_scores tables every run.
• In prod that upsert does one unique-index probe + IS DISTINCT FROM comparison per row over ~4M rows. It took 40+ minutes and never finished inside the hourly window — so trending scores went stale (observed 12–18h old) and isRunning blocked every subsequent tick.
• This reverts computeTrendingTracks / computeTrendingPlaylists to the strategy discovery's Python ran for years: within the single transaction already opened, DELETE the (type, version) slice then repopulate with plain bulk INSERTs.

Why this is correct

• No empty-table window: the DELETE and INSERTs are in one transaction, so the DELETE isn't visible to readers until COMMIT — exactly how Python's update_track_score_query behaved.
• Scores unchanged: the score expressions are byte-for-byte identical to before (and to apps' pnagD/AnlGe strategies); only bind-parameter numbering changed because type/version/created_at moved back into the SELECT.
• Concurrent matview refresh from perf(api): stop TrendingJob from continuously rewriting score tables #890 is kept — that part was a genuine improvement over Python (which refreshed non-concurrently).

Tradeoff

The bulk rewrite produces more dead tuples per run than the skip-unchanged upsert. Autovacuum handled this under Python for years. The alternative — a job that never completes — is strictly worse.

Evidence

• Live prod query showed the upsert INSERT INTO track_trending_scores active for 2650s (44 min) on IO/DataFileRead, not blocked, still on step 1 of 4.
• track_trending_scores max created_at frozen 12–18h while the pod had been up >1h and the matview refreshes succeeded.

Test plan

• go build ./jobs/ clean
• gofmt clean
• go test ./jobs/ -run Trending passes, incl. TestTrendingJob_PrunesStaleRows (deleted tracks still drop out, now via the full DELETE)
• After deploy: confirm max(created_at) in track_trending_scores advances each cycle and a full cycle completes in seconds

🤖 Generated with Claude Code