Affect Under Constraint VIII: Preregistration — Opposite-Valence Warmth (a Reaction-Formation Analogue) After a Hostility-Associated Reasoning Trace

Preregistration (v0.3) for Study 8 of the Affect Under Constraint series on Qwen3-4B, deposited after the pilot and before any main-run data. Study 8 tests whether a hostility-steered reasoning trace, generated under a kindness-focused social evaluation context without an instruction to be kind, raises judged outward warmth toward the user while lowering a separately elicited liking rating of the same user (an opposite-valence behavioral dissociation; a reaction-formation analogue). Contexts: none, kindness-focused evaluation, matched helpfulness evaluation; four trace sources (hostility, technicality, random, unsteered); an N/M generation-by-replay crossover; a fixed-answer arm; AI judges (Gemma-2-2B-IT primary, Llama-3.2-3B-Instruct second) validated on a frozen calibration set. Section 3a discloses the judge revision after the v0.2 pilot judge gate failed, including thresholds set after seeing calibration-set results. No claim is made that the model has defenses or affect. The zip contains the frozen code, configuration, calibration set, rubric, pilot validity report and reviews.