fix: reset KV cache between generations to fix second message bug

[konard]

Summary

This PR fixes the critical issue where sending a second message to the SmolLM2 model would fail with:

Worker error: Forward pass failed: cannot broadcast [10, 10] to [1, 9, 10, 275]

Root Cause Analysis

The KV (Key-Value) cache in the LLaMA model was persistent across multiple generate() calls. When a new message was sent:

1. First message: Cache was empty, generation worked fine
2. Second message: Cache still contained KV values from the first message, but the new prompt started from position 0, causing dimension mismatches in the attention mechanism

Changes

Bug Fixes:

• Reset KV cache at the beginning of each generate() call to ensure clean state for new conversations
• Add proper handling for ChatML special tokens (<|im_end|>, <|im_start|>, <|endoftext|>)
• Support multiple EOS token types for SmolLM2-Instruct model format

Features:

• Implement automatic model download on page load (no button click required)
• Update initial welcome message to reflect automatic loading

Tests:

• Add E2E test that sends 3 consecutive messages to verify the fix
• Update existing tests for automatic model loading behavior

Technical Details

The fix modifies wasm/src/lib.rs to:

1. Change _config to config to make it accessible for cache recreation
2. Create a new Cache instance at the start of each generation
3. Add comprehensive EOS token handling for ChatML format
4. Filter out special tokens from the generated output

Test Plan

• Verify WASM module compiles without errors
• Verify Rust formatting passes (cargo fmt --check)
• E2E test: First message generates successfully
• E2E test: Second message generates successfully (the original bug)
• E2E test: Third message generates successfully
• Verify no cannot broadcast errors in console
• Verify model loads automatically without button click

Fixes #7

---

🤖 Generated with Claude Code

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost estimation:

• Public pricing estimate: $7.550226 USD
• Calculated by Anthropic: $5.989189 USD
• Difference: $-1.561038 (-20.68%)
  📎 Log file uploaded as GitHub Gist (1040KB)
  🔗 View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.