docs: research local AI options for M3 Pro 36 GB

[bearmug]

Summary

• correct the target machine from the supplied M3 Max assumption to the locally observed M3 Pro 36 GB
• assess North Mini Code, Qwen3.6 OptiQ/MTP, Gemma 4, and DiffusionGemma against current primary sources and exact pinned artifact sizes
• document the two active llama.cpp services, near-full swap, and the deleted-but-open Qwen GGUF restart risk
• propose an approval-gated, one-model-at-a-time benchmark plan with pinned tasks, equal repetitions, parser/cache checks, and raw JSONL records
• add a structured candidate/service inventory with revisions, licenses, evidence status, runtime targets, source URLs, and byte totals

No models were downloaded, packages installed, services restarted, or intensive inference runs performed.

Validation

• git diff --check
• custom Python validation: 8 CSV rows / 14 required fields, unique candidates, byte-to-GB arithmetic, pinned revisions, and all CSV Hugging Face source URLs returning HTTP 200
• fresh focused reviewers: no blockers or majors after fixes
• codex review --uncommitted: no actionable correctness findings

npm run check and npm test were attempted, but this isolated worktree has no installed dependencies. Reusing another worktree's node_modules produced stale workspace-resolution failures unrelated to this docs-only diff; CI remains authoritative for repository checks.

Notes

• Vendor benchmark/speed claims are labeled as hypotheses, not local results.
• The live port-8004 Qwen process must not be stopped until its pinned replacement and launcher are verified.
• Any operational follow-up requires explicit approval; this PR is research/documentation only.