Backend: unembedding cosine emission + Devstral tokenizer fallback

Emit per-token embedding L2 norms for RQ1 Step 3

Emit per-layer top-K SwiGLU intermediate activations for RQ1 Step 5

Emit effective_probability and sampling eligibility per alternative

Expose query_position and query_token_id in /attention/row response

Derive aggregate-row num_heads from request metadata, not live model

Fix empty text and incorrect is_special for Mistral control tokens

Fix score-all classification in vectorised path

Refactor head classification from cascade to score-all-then-rank

Add FFN contribution metrics and gate activation stats to research endpoint

Add tuned lens as supplementary projection mode for logit lens

Rename 'counterfactual' terminology to 'input ablation' in docstrings

Fix margin winner_token in margin_data and extract _classify_stability helper

Add margin-based decision analysis, interventional counterfactuals, and run comparison (v3.0)

Add deep inspection: data-driven pattern classification, attention/MLP tracking, logit lens

Change default model to Devstral and optimise attention extraction

Fix MistralTokenizer loading and token section boundaries

Fix MistralTokenizer not loaded during model switch

feat: add top-k/top-p sampling and detailed logits/probability tracking

fix: Improve token alternatives numerical stability and temperature control

Store matrices as numpy arrays instead of Python lists

Add per-step memory cleanup for large model support

Fix RAM exhaustion for large token generation

feat: add auto_complete parameter for token generation

fix: add QKV extraction support for Mistral/Devstral architecture

feat: implement lazy-loading for attention matrices

Add avg_entropy calculation for attention heads

Revert QKV visualization fixes - need better approach for data streaming

Add safety checks for missing QKV keys

Limit QKV matrices to top 5 heads per layer to reduce response size

Fix QKV matrix extraction for Mistral/Devstral architecture

Fix QKV visualization for Mistral/Devstral architecture

Fix: Import time module at top level for SSE events

Add SSE streaming endpoint for real-time analysis progress

feat: Include token metadata in analysis response

feat: Implement tier-based model filtering by device type

Fix: Add attn_implementation="eager" to model switch function

Add tokenSections boundaries and update system prompt

Fix: Handle MistralCommonTokenizer pad_token setter

Integrate mistral-common for correct Devstral tokenization

Remove mistral_common to fix dependency conflict

Use mistral_common for proper Devstral prompt formatting

Add system prompt support for instruction-tuned models

fix: Simpler prompt format and temperature=0 for Devstral

fix: Sanitize JSON response for NaN/Inf float values

fix: Check chat_template is set before using apply_chat_template

fix: Add chat template support for Devstral instruct model

fix: Convert bfloat16 to float32 for numpy compatibility

fix: Use eager attention for output_attentions support

Add vocabSize to modelInfo response