docs: adopt local 350M compute plan

[MaharshPatelX]

Summary

• update the project tracker after merged PR feat: add canonical data and leakage-safe splits #7: Parts 0-4 are complete, progress is 5/17, and Part 5 is next
• make LFM2.5-ModelRouter-350M the only active model scale; larger variants remain future work
• add a local-first compute plan for the RX 7900 XTX with 24 GB VRAM and the existing 16 GB host RAM
• require measured memory failures before any RAM-upgrade recommendation
• keep Part 5 CPU-first and define a bounded ROCm/PyTorch acceptance benchmark for Part 6
• make rented GPUs a documented fallback rather than the default
• align the README, blueprint, reproducibility guide, and Part 3/4 status pages

Updated tracker

• 5 of 17 core parts complete
• Parts 3 and 4 marked Done with PR feat: add canonical data and leakage-safe splits #7 evidence
• Part 5 Baseline Suite marked Next
• Part 5 implementation checklist added

Verification

• ruff format --check .
• ruff check .
• mypy src tests
• pytest -q -p no:cacheprovider (45 passed)
• git diff --check