Default to 0.6B across API/CLI/demo + CLI --model picker (ADR 0051)

[marzoukbaig14]

What

Make 0.6B the default everywhere Committed downloads or serves; 1.7B stays fully selectable. Supersedes ADR 0048 (1.7B serving-of-record) → ADR 0051.

• engine.py: DEFAULT_MODEL_REPO/FILE → the 0.6B GGUF, DEFAULT_TOKENIZER → Qwen/Qwen3-0.6B (identical vocab, v2-i1 Gate 0). Single source for the bare-download default.
• serving/api.py: registry DEFAULT_MODEL → "0.6b" (eager at startup); 1.7B now lazy on first model:"1.7b" request. A no-model /generate now returns 0.6B — an intentional default flip (coordinated with the portfolio reframing).
• cli.py: --model {0.6b,1.7b} (default 0.6b) mapping to the two GGUF repos; --model-path + COMMITTED_MODEL_* overrides retained; model-aware download-size notice; --help updated.
• Demo: dropdown default follows DEFAULT_MODEL → 0.6B on next rebuild.
• Thread fix (n_threads=2) retained on every instance.

Facts for the model cards (observed, not guessed)

• 0.6B GGUF marzoukbaig14/committed-gguf-0.6b/committed-0.6b-finetuned-Q4_K_M.gguf: 396,704,512 bytes ≈ 396.7 MB (378.3 MiB), quant Q4_K_M (general.file_type=15), arch qwen3. Hub-reported size matches the local build byte-for-byte.
• Both repos public: committed-gguf-0.6b and committed-gguf → private=False ("no token needed" is true).
• Exact CLI commands (after this change):
  • default (0.6B): git diff | committed
  • explicit 1.7B: git diff | committed --model 1.7b
  • the pip install --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu committed @ git+... install line is unchanged and model-agnostic (the package pulls whichever public GGUF the default/--model selects).
• Real 0.6B generation (divide-by-zero guard diff → git diff | committed): fix: divide by zero — valid Conventional Commit. The default-path download notice fired for committed-gguf-0.6b (~380 MB), confirming the CLI default resolves to 0.6B.

Verification

ruff check clean; committed --help shows --model {0.6b,1.7b} (default 0.6b); api.DEFAULT_MODEL == "0.6b"; engine default → the 0.6B GGUF; live 0.6B generation above.

After merge (Part 4 — I'll do it)

Factory-rebuild committed-api and committed-demo (merge alone doesn't pull @main), then verify live: /health shows 0.6B default/eager, no-model → 0.6B (~1s warm), model:"1.7b" lazy-loads, demo dropdown defaults to 0.6B. Note: once this deploys, the portfolio's no-model requests begin returning 0.6B (intended).

Does not touch the portfolio repo; does not write the Hub cards (human does, from the facts above).

🤖 Generated with Claude Code

---

Generated by Claude Code