feat: local-model-fit domain — catalog, hardware profiles, scoring engine, and UI

[Copilot]

Self-contained local-model-fit domain that ranks models by how well they fit a user's hardware. Four layers: catalog ingestion → hardware profile → deterministic scoring → UI consumers. One canonical ranked result set powers all views.

Core domain (web/src/local_model_fit/)

• types.ts — HardwareProfile, ModelCatalogEntry, ModelVariant, RankedModelFit (canonical result shape with score 0–100, tier S/A/B/C/D/F, fit label, memory %, reasons[])
• modelCatalog.ts — 30+ normalized models from Ollama/llama.cpp/LM Studio ecosystem (chat, code, reasoning, vision). Static v1 seed with documented ingestion strategy for future automation
• hardwareProfiles.ts — 21 presets (NVIDIA GPUs, Apple Silicon, CPU-only) + buildCustomProfile() for manual VRAM/RAM overrides
• hardwareProfileDetection.ts — WebGPU adapter probe → platform/UA heuristics → navigator.deviceMemory fallback
• scoreModelFit.ts — Deterministic heuristic: effective memory (VRAM for discrete, 75% unified for macOS, 60% RAM for CPU-only) → memory ratio → score with quant bonus. No ML, fully explainable
• rankModelFits.ts / filterModelFits.ts / summaries.ts — Composable ranking, filtering (search/tags/families/tiers/fits), tier counts

Hooks & store

• useHardwareProfile() — detect/set profile
• useRankedModelFits() — two-layer memo: score on profile change, filter on UI interaction
• useModelFitSummary() — tier counts + one-liner
• localModelFitStore — Zustand with persist for hardware profile + view mode

UI (web/src/components/local_model_fit/)

• LocalModelFitPage — hardware selector → tier summary → search/filter bar → card/list toggle
• HardwareProfileSelector — preset dropdown, auto-detect, custom VRAM/RAM inputs
• TierSummary — clickable tier chips that toggle as filters
• LocalModelFitCardGrid / LocalModelFitList — two renderers, same RankedModelFit[]
• FitBadge — reusable tier chip for later integration into model menus

Usage

import { useRankedModelFits, useHardwareProfile } from "../local_model_fit";

const { profile } = useHardwareProfile();
const { results } = useRankedModelFits();
// results: RankedModelFit[] — sorted, filtered, ready to render

Route: /local-model-fit (lazy-loaded, protected).

Tests

43 unit tests across scoring, ranking, filtering, and summaries — all pure functions, no mocking needed.

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[cloudflare-workers-and-pages]

Deploying nodetool-staging with    Cloudflare Pages

Latest commit:	a735fa6
Status:	🚫  Build failed.

View logs

[cloudflare-workers-and-pages]

Deploying nodetool with    Cloudflare Pages

Latest commit:	a735fa6
Status:	🚫  Build failed.

View logs