feat: add llmfit budget subcommand and web Budget Advisor panel

[Clemente-H]

Summary

Adds a hardware budget advisor feature that lets users discover which hardware configurations fit within a purchase budget and see which LLM models each configuration can run — using the same fit pipeline as llmfit fit.

• llmfit budget <MAX_PRICE> — new CLI subcommand. Filters a built-in hardware catalog by price and shows ranked LLM model lists per configuration, with --limit, --min-fit, --json, and --csv support.
• llmfit-core/src/hardware_catalog.rs — static catalog of ~16 real-world hardware configurations (mini PCs, Apple Silicon, NVIDIA/AMD GPUs, ROCm cards) with prices, VRAM, RAM, and CPU specs. configs_within_budget(max_price) returns matching entries. Designed to be easy to extend.
• SystemSpecs::synthetic() — new constructor in hardware.rs for building simulated specs without touching real hardware detection. Used internally by HardwareConfig::to_specs().
• GET /api/v1/budget?max_price=N&limit=5&min_fit=marginal — new REST endpoint in serve_api.rs.
• Budget Advisor panel in the web dashboard (llmfit-web) — form with budget/limit/fit-level controls, renders a per-configuration table of top models using the existing CSS design system.

Example

$ llmfit budget 800
=== Hardware Budget: up to $800 ===
10 configuration(s) found

┌─ Mac mini M4 Pro (24 GB) — $799 │ 24 GB VRAM │ 24 GB RAM │ 14 cores │ Metal
│  Note: Unified memory; 24 GB pool, excellent perf/watt
...
┌─ RTX 3090 (24 GB) PC — $700 │ 24 GB VRAM │ 32 GB RAM │ 12 cores │ CUDA
...

$ llmfit budget 800 --json
$ llmfit budget 1500 --limit 10 --min-fit good

Closes #541 (hardware-aware model selection)

Test plan

• cargo test -p llmfit-core — hardware_catalog unit tests pass (budget filter, empty budget, to_specs no panic)
• llmfit budget 800 — shows ≥1 config with ranked model table
• llmfit budget 800 --json — valid JSON with budget_usd and configurations array
• llmfit budget 0 — prints "no configurations" message cleanly
• llmfit serve → open browser → click "💰 Budget advisor" → enter budget → results appear
• curl "http://127.0.0.1:8787/api/v1/budget?max_price=800&limit=3" — returns JSON

🤖 Generated with Claude Code

[AlexsJones]

Thanks @Clemente-H — this is a substantial, well-structured contribution: the catalog module has tests, the subcommand reuses the existing fit pipeline instead of reinventing it, and the JSON/CSV output follows the house style.

Before it can move forward there's a scope question that's the maintainer's call, plus a few technical points.

Scope (for @AlexsJones): this bakes a curated hardware price catalog into the compiled binary. Prices go stale fast (the catalog itself says "as of mid-2025" — already a year old, e.g. RTX 3090 at $700), vary by region and currency, and every price refresh becomes a code change and release. If the feature direction is wanted, it may fit better as a data file (bundled or fetched JSON with a "last updated" stamp) so the catalog can be refreshed without a release.

Technical notes in the meantime:

1. The NVIDIA L4 (24 GB) Cloud Instance entry prices a monthly cloud cost on a purchase-price axis (its own note admits this). It will distort budget comparisons — drop it, or make cloud options a separate concept.
2. --min-fit silently maps any invalid value to marginal; rejecting unknown values with an error would match the CLI's behavior elsewhere.
3. The 2× RTX 3090 (48 GB total) entry is modeled in to_specs() as a single GpuInfo with 48 GB (count: 1). If the fit pipeline treats one 48 GB device differently from 2×24 GB tensor-parallel (it does for per-device layer fits), this overstates what that config can run.

Parking this for the maintainer's direction on the catalog approach before further iteration — the implementation itself is in good shape.

🤝 Handled by Alex's Repo Steward — replies reviewed by @AlexsJones.

Learn more or run your own: https://github.com/AlexsJones/repo-steward