fix(virality): score last completed day and floor acceleration at 0

[PabloPerezBenavente]

Problem

Narratives were being badged early_surge while showing essentially zero acceleration. Investigation found two stacked issues:

1. Acceleration was ~0 for almost the entire cohort. In prod, 22,208 of 22,256 narratives had acceleration exactly 0.0. The analysis-indicators pipeline defaulted calc_date to date.today(), but its scheduled cron fires just after midnight UTC (observed 00:04:31Z). Acceleration compares carried-forward video stats "as of calc_date" against "as of calc_date - 1", so the delta is driven entirely by videos scraped on calc_date. At 00:04 the current day has barely been scraped → current == prev → acceleration 0.

2. The percentile classifier promoted that noise into the surge band. With the switch to percentile-based alert levels, EARLY_SURGE requires acceleration percentile ≥ 0.95. When 99.8% of the field is 0.0 and a handful are negative (decliners), any microscopically-positive narrative sorts to the top percentile — so flat narratives with accel ≈ 0.0001 were classified as surging.

Changes

• Default calc_date to yesterday (the last completed scraping day) in run_narrative_analysis_indicators_pipeline, so both comparison windows cover a full day of scraping. On prod data this lifts nonzero accelerations from 48 → ~2,288.
• Add a --calc-date YYYY-MM-DD CLI option for backfills and re-runs.
• Floor acceleration_rate at 0. Acceleration measures growth, so a declining narrative reads as "not surging" (0), not a negative rate. This also collapses all decliners and flats into one 0.0 block at the bottom of the percentile ranking, so only genuinely growing narratives rank high — killing the flat-narrative-as-EARLY_SURGE noise.

Verification

• Read-only replay of the comparison logic for a completed day (2026-07-14 vs 2026-07-13) returned 2,288 narratives with nonzero acceleration vs 48 from the buggy 00:04 run.
• mypy clean on changed files.

Notes / follow-ups

• The cron that schedules this pipeline at 00:04 UTC lives outside this repo (not in deployment/). The default-to-yesterday change fixes it regardless, but if that external job passes an explicit date it would override the default — worth confirming where it's defined.
• No unit test added yet for the calc_date default / --calc-date override.

🤖 Generated with Claude Code