feat: impact × uncertainty signal map in the prospective demo

[picaultj]

Summary

Implements "better characterize weak signals (axes impact vs. uncertainty)" — the fourth idea in upstream issue rte-france/BERTrend#59 — as a foresight-style signal map in the prospective demo's signal-analysis view.

Approach (heuristic, data-derived)

The prospective demo loads signals from saved parquet files (scalar metrics only, no popularity time series), so both axes are derived from the metrics already present — no LLM call, no schema change, and it works retroactively on existing models:

• Impact = normalized Latest_Popularity (magnitude of attention)
• Uncertainty = 1 − normalized(evidence), where evidence blends Docs_Count and Source_Diversity — a thinly-supported topic (few documents / few sources) is treated as more uncertain

Both axes are min-max normalized to [0, 1] across the model's combined signal set. (A future upgrade could replace/augment these with LLM-scored impact — see the earlier design discussion; this PR is the low-risk first slice.)

Changes

• signal_characterization.py (new): Streamlit-free helpers
  • compute_signal_characterization(dfs_topics) → per-topic dataframe with impact / uncertainty
  • plot_signal_characterization(char_df, ...) → Plotly quadrant scatter (x=uncertainty, y=impact, color by signal type, size by document count, dashed median lines at 0.5)
• dashboard_signals.py: renders the map in an expander above the signal tables.
• i18n_translations.py: FR/EN strings (signal_map_title, signal_map_help, impact_axis, uncertainty_axis).
• tests/apps/test_signal_characterization.py (new): unit tests.

Notes / limitations

• The impact/uncertainty formulas are heuristics flagged for calibration; the caption explains what the axes mean to users.
• Uncertainty here reflects evidence volume, not popularity volatility, because the prospective demo's stored dataframes don't carry the popularity time series (that would be available in the live weak-signals demo, a possible follow-up).

Tests

tests/apps/test_signal_characterization.py: 9 passed — normalization bounds, higher-popularity→higher-impact, more-evidence→lower-uncertainty, signal-type labelling, constant-metric midpoint, missing-column handling, and figure construction.

---

Generated by Claude Code