v1.17.0 — Ollama agentic reliability: 0% → 94% benchmarked task success

Ollama provider: agentic reliability overhaul

Head-to-head benchmarked against OpenCode on identical seeded coding tasks (qwen3-coder:30b, M1 Max, artifact-verified — methodology and raw data in the anymodel-umb increment 0018-anymodel-vs-opencode-benchmark):

stack	task success	median turn-around
anymodel 1.16.2 (ollama)	0/18 (0%)	~4s (failed instantly)
OpenCode 1.1.52	14/18 (78%)	~115s
anymodel 1.17.0 (ollama)	17/18 (94%)	~10s

Fixes

• Text-channel tool-call recovery on the native Ollama wire — Hermes / Qwen-XML / paren forms now convert to real tool_use blocks in both non-streaming and streaming turns (parity with the openai provider; streaming buffers the text channel and decides at end-of-message)
• flush() on the NDJSON stream translator — a stream that dies without done:true now still emits message_stop instead of hanging the agent loop
• Assistant tool_call echo normalization — native /api/chat types arguments as a map; the OpenAI-style string echo caused a 400 ("can't find closing '}'") on every request after the first tool use
• Skill fidelity injection for system-less requests — previously built and silently dropped

532/532 tests green (10 new).