feat: spec_type launch-arg plumbing + ngram-simple benchmark for specialist (#13)

[aismai]

Summary

Resolves #13 — measured, and the honest answer is: ngram-simple is a net slowdown on the specialist workload, so this PR ships the plumbing and the benchmark with the flag off by default, rather than enabling it.

• src/lore/models.py: start_model() now forwards an optional per-model spec_type config key as --spec-type <value> (skipped for the embeddings server). No key set → launch args unchanged.
• configs/models.yaml: spec_type: ngram-simple present on the specialist entry but commented out, with the measured result — enabling it later is a one-line uncomment.
• scripts/benchmark_ngram_spec.py: A/B benchmark modeled on benchmark_spec_decode.py. Runs the real TaskClassifier workload (20 classification queries with the production system prompt, JSON output) + 5 extraction/summarization prompts, temperature 0, and checks output equality between arms as the quality gate.
• tests/test_models.py: 3 new tests (flag present when configured, absent by default, never passed to embeddings). 27/27 pass.
• docs/optimization-log.md: full writeup.

Measurements

Falcon-H1-1.5B Q4_K_M, arm64 Metal, 16 GB, 8K ctx, temperature 0, 25 prompts/arm.

Metric	Baseline	ngram-simple
Avg tokens/sec	13.52	12.17 (-10.0%)
Avg latency/request	8.46 s	10.25 s
Exact-match outputs vs baseline	—	23/25

Why it loses: server logs show draft acceptance mostly 8–23%. The classifier's JSON keys repeat from the prompt, but the values and connective tokens don't, so most drafted n-grams are rejected and verification overhead outweighs accepted tokens. The 2/25 output diffs are batch-size numerics wobble under speculative verification (both arms ran temperature 0).

Setup caveat: measured with mainline llama.cpp (Homebrew b9910) and f16 KV — I couldn't build the turboquant fork in my environment. Both arms share the same binary/KV so the relative delta should carry over, but it's worth a rerun on your M4 with the bundled fork; the script honors env overrides and defaults to the fork binary + turbo4:

.venv/bin/python scripts/benchmark_ngram_spec.py

If your numbers come out positive, enabling is just uncommenting the spec_type line in configs/models.yaml — happy to update the default in this PR if so.