fix(report): live-summary Yhteenveto excludes demoted content (ADR-0032/0033)

[MikkoNumminen]

Problem

The whole-window Yhteenveto (the live-summary Overall narrative, served to the desk via /live/report) was synthesized over every structured item, including demoted rasismi/asiaton ones. SynthesizeThemeAsync hands the model a severity distribution and per-item excerpts with severity, so the model:

• named moderated content in the lead summary with a rating — e.g. '<slur>' (korkea);
• recited a severity distribution that counted demoted items (the inflated window total);
• computed the Overall count and trend over demoted items too — and the desk "suunta" KPI tile reads report.overall.directionLabel, so it inherited the leak.

This violated ADR-0032 (demoted = unrated, excluded from severity aggregates) and ADR-0033 §3 (top of the page reflects real feedback only).

Fix

Synthesize the Overall over rated items only (structured.Where(i => !DemotedCategories.Contains(i.Structure.Category))): narrative, severity digest, excerpts, window total, trend, and needs-review count. reportSentiment already excluded demoted content. A window containing only demoted content now yields no Yhteenveto (the collapsed moderation count still shows) instead of a summary written about hostile content.

Server-side in ReportService, so every consumer of Overall inherits it.

Not a separate bug (verified)

The korkea+kriittinen and entries KPI tiles already compute over rated sections in every render path (desk phase A/B, management, snapshot); for genuinely-demoted items they count 0. The reported "5 kriittinen in the moderated category" was the leaked trend tile + the Yhteenveto's own polluted numbers, both fixed here.

Tests

• New regression test SummaryMode_Overall_ExcludesDemotedContent_FromDigestCountAndTrend: asserts the demoted item's id/excerpt/critical severity never reach the synthesis digest and Overall.Count/FeedbackIds are rated-only.
• Full FeedbackIntelligence.Api.Tests: 148 passed, 0 failed.

AI-first docs

Yes — ADR-0033 §3 updated to state the Yhteenveto is part of the rated-only separation.

Known follow-ups (out of scope for this PR)

• Standard /report path (management view + offline snapshot) still LLM-synthesizes demoted-category narratives with severity fed in, so a moderation card there could editorialize a slur's severity.
• Moderation cards render a trend label including the severity-derived paheneva (worsening).
• Desk phase-A transient: a demoted critical can flash as rated for the moment before /schema resolves, then self-heals.