Bounded report gate rejects the promoted-row severity fields the prompt instructs the agent to author #931

What happened

Smoke run 32914329945 (candidate 8fe49f0, with #919 in) got further again: very-liquid-vaults reached final-report and failed only on:

Bounded report did not preserve dedupe field "impact_rationale" at $.issues[1].impact_rationale;
Bounded report did not preserve dedupe field "likelihood_rationale" at $.issues[1].likelihood_rationale

The other two rows never finished: the retry churn from this deterministic final-report failure ran the 15000-second watch window out (workflow_nonterminal: 2 after 4h22m).

Root cause

Same contradiction class as #918, for the promoted-row severity fields. In bounded classification mode the final-report prompt instructs, for every promoted finding: "author source-backed impact, likelihood, and their rationales, compute severity from the matrix, and set a concise severity_rationale". A dedupe row that already carries any of those fields (schema-legal — strategy findings may include preliminary severity data) can therefore never byte-equal its promoted report row: the agent is required to re-author them. #919 excluded only lifecycle and triage_classification; the generic loop still demands byte-equality for the authored severity fields.

severity itself remains independently validated — the gate requires lifecycle.canonical_severity to equal the report row's severity and enforces High→Medium→Low stable ordering — and the report schema bounds the authored prose.

Fix

For promoted rows, extend the generic preservation loop's skip set from id/title to the full prompt-authored set: impact, likelihood, impact_rationale, likelihood_rationale, severity, severity_rationale. Non-promoted rows keep full byte-preservation for these fields, and every evidence field (description, proof_of_concept, summary, spans, provenance) stays strictly preserved for all rows.