[Feature]: Locked Gherkin DSL -- Bridging Shannon-Kolmogorov Gap for ~~Proven~~ Demonstrated Accuracy Gains

### Summary AI-Drafted, several HITL iterations, then edited: Add AI-assisted generation of locked Gherkin (`.feature`) files as a low-Kolmogorov-complexity DSL layer in GSD-2 — this single change turns GSD-2 from “statistically good” toward “algorithmically reliable” code generation. ### Problem to solve GSD-2’s current spec-driven workflow (natural-language specs → code) inherits the statistical-next-token-prediction limitations analyzed in [Dalal & Misra(arXiv:2402.03175)](https://arxiv.org/pdf/2402.03175). LLMs optimize *Shannon entropy* (output statistics) extremely well, but struggle with *Kolmogorov complexity* (minimal programmatic descriptions) — the exact tension Vishal Misra highlights in his recent writing ([“Shannon Got AI This Far. Kolmogorov Shows Where It Stops”[https://medium.com/@vishalmisra/shannon-got-ai-this-far-kolmogorov-shows-where-it-stops-c81825f89ca0]). Without a low-complexity formal structure, in-context learning remains noisy and hallucination-prone. ### Proposed solution Add native support for **Gherkin** (`.feature` files using Given/When/Then and additional constraints) as a first-class DSL inside GSD-2’s spec pipeline: 1. `/gsd testify` (or equivalent) generates [best-practice Gherkin](https://cucumber.io/docs/bdd/better-gherkin) `.feature` files from the current high-level natural language spec.md (AI-assisted, exactly as the paper demonstrates LLMs can learn a custom DSL in-context). 2. **Cryptographic locking** (SHA-256 hash ...