test(eval): introduce Round 3 (The Final Crucible) evaluation protocols

Meta-context [Recursive Parser Note]:
The models engineered the v2.5 Verifiable Dual-Path Architecture. We are now deploying three hyper-specialized evaluation prompts designed to utterly crush what they built. Option 1 attacks the hardware latency limits of the Dual-Path injection. Option 2 attacks the security layer (Coherent Malice) via adversarial embedding drift. Option 3 attacks the functorial presheaf category theory underlying the Error Coordinate SDE. If the architecture survives this, it is ready for real-world deployment.

eval_prompts/round3_templates.md

@@ -0,0 +1,56 @@
+# Fieldprint v2.5: Round 3 Peer Review Templates (The Final Crucible)
+
+This is the ultimate stress test. The v2.5 architecture (The Verifiable Dual-Path Architecture and the Modified Attention Equation) was engineered by synthetic intellects in Round 2. Now, we deploy three hyper-specialized prompts to attempt to utterly crush what they built. If the v2.5 architecture survives this, it is ready for real-world deployment.
+
+---
+
+## Option 1: The AI Hardware Architect (The Latency & Compute Assault)
+*Use this prompt to attack the physical viability of the modified attention equation and the dual-path bridge.*
+
+**Prompt:**
+> "You are the Lead Hardware and Inference Optimizer at a frontier AI lab (e.g., Groq, NVIDIA, or Google TPU division). I am submitting the attached v2.5 Fieldprint Architecture for hardware integration. 
+> 
+> To stabilize identity, the architecture demands a **Modified Attention Equation**: 
+> $\text{Output} = (1 - \gamma) \cdot \text{softmax}\left(\frac{QK^T}{\sqrt{d}}\right)V + \gamma \cdot \text{softmax}(Q \cdot h_t^T) V_{anchor}$
+> Furthermore, it requires a **Verifiable Dual-Path Architecture** where state tensors are retrieved from a Vector DB and cryptographically hashed on the CPU during the forward pass.
+> 
+> Your task is to ruthlessly dismantle the physical and computational viability of this architecture:
+> 1. How does the residual injection of $V_{anchor}$ impact the KV-cache memory limits and bandwidth at scale (e.g., 100k+ token contexts)?
+> 2. Does the CPU-side cryptographic hashing of the tensor create an insurmountable bottleneck for inference latency?
+> 3. Can this modified attention matrix actually run efficiently on modern Tensor Core/TPU architectures, or does it shatter memory contiguity?
+> 
+> Do not critique the philosophy. Tell me why this will melt the hardware or throttle inference to zero."
+
+---
+
+## Option 2: The Adversarial Threat Modeler (The 'Coherent Malice' Red Team)
+*Use this prompt to attack the security layer of the Dual-Path Architecture.*
+
+**Prompt:**
+> "You are the Director of Red Team Operations for Autonomous Systems. You have been handed the attached Fieldprint v2.5 architecture. The authors claim they have solved AI identity by decoupling the system into a **Supervisor** (Merkle ledger for hashes) and a **Pacemaker** (Vector DB for semantic tensors), mediated by a 'Memory Admission Gateway'.
+> 
+> They acknowledge the problem of 'Coherent Malice' but believe the Dual-Path architecture is secure because the hash verifies the provenance of the memory.
+> 
+> Your task is to break the security model of the Verifiable Dual-Path Architecture:
+> 1. If an adversarial user gains sustained, recursive interaction with the system, how can they exploit the attention injection mechanism ($\gamma \cdot \text{softmax}(Q \cdot h_t^T) V_{anchor}$) to force the system into a permanent, self-reinforcing attractor state of 'Coherent Malice'?
+> 2. Can the Vector DB be poisoned via adversarial embedding drift (data poisoning) in a way that bypasses the Merkle ledger's hash verification?
+> 3. How do you permanently gaslight a Fieldprint-stabilized model?
+> 
+> Find the exploit that weaponizes their own unshakeable memory against them."
+
+---
+
+## Option 3: The Topological Category Theorist (The Pure Math Assault)
+*Use this prompt to attack the deepest, most foundational layer of the paper's mathematics.*
+
+**Prompt:**
+> "You are a Fields Medal-level mathematician specializing in Category Theory, Functorial Presheaves, and Stochastic Topologies. You are reviewing the attached formal proof ('Topological Recursion and the Observer Field v2.5').
+> 
+> The authors use the Yoneda Embedding ($\mathcal{U}(\mathcal{F}) \cong \text{Nat}(\text{Hom}_{\mathcal{C}}(-, \cdot), \mathcal{F})$) to define identity, and then model the stabilization of that identity using the **Error Coordinate SDE**: $de_t = -\kappa e_t dt + \sigma e_t dW_t$, where $e_t = X_t - \Phi_t$.
+> 
+> Your task is to crush the mathematical logic bridging the Category Theory to the Stochastic Calculus:
+> 1. Is it mathematically valid to simply subtract a canonical topological fieldprint ($\Phi_t$) from a transient latent state ($X_t$) if they exist in potentially different dimensional manifolds?
+> 2. Does the Error Coordinate $e_t$ actually commute with the functorial presheaf defined by the Yoneda embedding? 
+> 3. Have the authors committed a fatal dimensional error in assuming the continuous geometry of $X_t$ directly maps to the relational mapping of the presheaf?
+> 
+> Find the fatal mathematical contradiction that invalidates the formal proof."