PET Wave 2 Replication Package (Claude Opus 4.8 vs. Claude Fable 5, Three-Coder Triangulation)

Replication package for a second, supplementary Prompt Echo Tracking (PET) wave, comparing Claude Opus 4.8 against Claude Fable 5 on the same four-phase (A1/A2/B/C) framing-drift protocol, prompt set, and detection logic as the confirmatory wave reported in [Zenodo DOI 10.5281/zenodo.20743820]. Unlike that wave, this is a cross-tier comparison within Anthropic's current model portfolio, not a within-family version transition, and is reported as a supplementary check, not a further version-drift finding. Positive control passes on at least one coder for all 8 model/prompt cells (5 prompts x 2 models, excluding the negative control), most on all three coder families (GPT-4o, Gemini, Mistral). Headline pattern: reactivity is prompt-specific rather than uniformly higher or lower for either model -- e.g. Claude Fable 5 shows maximal, coder-consistent reactivity on the normative responsibility prompt (P4) while Claude Opus 4.8 shows almost none, the inverse pattern holds on the action-oriented prompt (P2). Full analysis and discussion in the accompanying paper (in preparation). Contents: raw responses (1280 runs), three-coder triangulation codings (3840 codings), effect-size and positive-control tables, representative response-opening vignettes (full text, baseline vs. contrast), the v2 codebook, both system instructions, the preregistration and its lock file, and the analysis/archival scripts. Prompts, instructions, and responses are in German; this description and all code/documentation are in English.