Improve model load error handling with structured diagnostics, retry logic, and self-hosting support

[Sumanth-806307]

Problem

Models fail to load across multiple surfaces (chat.webllm.ai, JSFiddle examples, Chrome extensions) as reported in #85. Current error handling lacks:

• Structured error classification
• Automatic retry mechanisms
• Cache recovery logic
• User-actionable error messages
• Self-hosting capabilities

Root Causes Identified

1. Insufficient Error Diagnostics - Generic error messages without classification codes
2. No Retry Logic - Transient network/CDN failures cause hard stops
3. Cache Corruption - No automatic cache clearing and retry
4. No Self-Hosting Support - Users locked into default CDN with no override option

Proposed Solution

Phase 1: Enhanced Error Diagnostics (High Priority)

• Add ModelLoadErrorCode enum (manifest_fetch_failed, artifact_fetch_failed, worker_init_failed, webgpu_init_failed, cache_invalid)
• Implement error classification in webllm.ts
• Add structured error display with actionable guidance
• Include "Copy Diagnostics" feature for bug reports

Files: app/client/api.ts, app/client/webllm.ts, app/store/chat.ts

Phase 2: Retry Logic & Self-Recovery

• Automatic retry with exponential backoff (max 3 attempts, 1s → 2s → 4s)
• Automatic cache clearing on cache_invalid errors
• Progress indication during retries
• Only retry on retryable error types

Files: app/client/webllm.ts

Phase 3: Custom Artifact Source Support

• Add customModelBaseUrl config option
• UI setting in model configuration panel
• Allow users to self-host model artifacts
• Addresses issue [Bug] chat.weblm.ai not loading LLMs #85 reporter's request

Files: app/store/config.ts, app/components/model-config.tsx, app/client/webllm.ts

Phase 4: Documentation

• Troubleshooting guide with error code explanations
• Self-hosting setup instructions
• Updated issue templates with diagnostic fields

Files: docs/TROUBLESHOOTING.md, docs/SELF_HOSTING.md, .github/ISSUE_TEMPLATE/bug_report.md

Acceptance Criteria

• All model load errors map to defined error codes
• Retryable errors trigger automatic retry (max 3)
• Cache corruption triggers automatic clear + retry
• Custom base URL configurable in Settings
• Error messages include actionable guidance
• "Copy Diagnostics" provides complete debug info
• Documentation covers all error codes and self-hosting

Implementation Details

Full implementation plan available in plan-85.md with:

• Detailed code examples for each phase
• Testing strategy (unit, integration, manual)
• Rollout strategy with risk assessment
• Success metrics and monitoring approach

Related Issues

• Addresses [Bug] chat.weblm.ai not loading LLMs #85 - Original bug report
• Improves overall error handling infrastructure

Estimated Effort

Time: 3-4 weeks (1 developer)
Priority: High (affects user experience across all surfaces)
Risk: Low-Medium (Phase 1-2), Low (Phase 3-4)