docs: Qwen3.8-27B local-model evaluation (EN/TR/DE)

[pretyflaco]

Refreshes docs/local-model-evaluation.md with the Qwen3.8-27B evaluation, run through millet's real two-pass code path (not a harness).

Headline

With the v0.15.1 Pass-1 truncation fix (#31), qwen3.8:27b matches or exceeds the cloud baseline on topic/action/question coverage across English, Turkish, and German meetings — 5/5 format, localized non-English headers, no hallucinated speakers, ~55-100s/meeting on an RTX 3090. It supersedes the earlier 'gpt-oss:20b is best local' conclusion.

Results (via millet, v0.15.1)

Meeting	Lang	Qwen3.8 (T/A/D/Q)	Baseline (T/A/D)	Format
E	EN	20/8/4/7	11/6/3	5/5
F	EN	10/24/5/10	18/20/4	5/5
G	EN	20/16/6/10	10/12/4	5/5
H	TR	10/1/3/7	6/1/0	5/5
I	DE	15/4/3/8	15/3/0	5/5

What changed in the doc

• New 'Qwen3.8-27B Evaluation (2026-08-15)' section + appendix entry.
• Reconciled stale claims: default model (qwen3.5:9b + MILLET_SUMMARY_MODEL, not 'changed to gpt-oss'); gpt-oss demoted to fallback; Sonnet 'only model' line softened; Phase-1 marked shipped.
• Documented the v0.15.1 truncation fix as a prerequisite, and the known 'decisions over-inclusion' weakness.

Caveats (stated in-doc)

• ~2× Sonnet latency; mild over-count of 'decisions'; non-English sample still small (1 TR + 1 DE).

Docs-only; no release. Metrics only — no transcript/meeting content (meetings anonymized E-I).