v0.9.0 — the model catalog + hardware-honest local defaults

forge learns to answer "which model, and will it run here?" — and to run
on more than a Mac.

The model catalog (new forge models command family):
· forge models — curated, hand-verified recipes checked against
THIS machine's RAM: engine, size, a fit estimate, and an honest
verified/candidate status (only models forge has actually run are verified).
· forge models show — the copy-pasteable runbook; for a custom-arch
model like sarvam-30b, the strings|grep arch gate BEFORE the 20GB
download, resumable weights, serve flags, the context-match gotcha.
· forge models use — turnkey provisioning for Ollama-native models
(pull, write config, smoke-test → forge run just works). Non-Ollama
engines get the honest runbook, never a half-done job.

Hardware-honest local defaults:
· CPU-only machines (no Apple Silicon, no dGPU) get a RAM-capped ladder
instead of a 9B escalation rung that would swap-thrash an 8GB laptop.
· Windows RAM is now detected — it was silently 0, dropping every Windows
machine to the minimal tier.

Full-fidelity replay (H11): forge replay --to-fixture now also captures a
content-addressed workspace snapshot. Credentials are never archived,
captures are bounded, and everything excluded is an honest fidelity limitation.

docs/models: an open collection of how models run through forge, first entry
sarvam-30b.

Still stdlib-only, zero runtime deps. 667 tests.