chore(release): v9.10.0 — client-only tool execution, end-to-end SSE …

[veerareddyvishal144]

…streaming, classifier v2

• Remove server-mode tool execution: Lynkr always forwards tool calls to the client (deletes src/tools registry, src/agents, orphaned backends, Code Mode, TOOL_EXECUTION_MODE and dead env knobs)
• Upstream SSE streaming, default on: native Anthropic passthrough (azure-anthropic, z.ai, ollama) + in-flight OpenAI→Anthropic transform with tool-call fragment reassembly; routing badge streams after message_start; kill switches LYNKR_NATIVE_PASSTHROUGH / LYNKR_STREAM_TRANSFORM
• Ollama stream frame filter: strip unsigned thinking blocks, normalize message_start (LYNKR_OLLAMA_BUFFER_RESPONSES=false to stream)
• Difficulty classifier v2: conversation context for short follow-ups, vocabulary-trap negatives, one-band reconcile cap (follow-up eval 60%→84%, SIMPLE→REASONING criticals 3→0)
• Fixes: z-ai tier provider alias fell through dispatch to the Databricks dummy port; tier fallback forwards client auth headers, skips error-answering candidates, ignores non-Anthropic bearer tokens for azure-anthropic; llama.cpp folds in-messages system roles and escalates context overflow; telemetry request_id NOT NULL row drops; provider failures log err.cause
• Side-channel routing guards: message-count regression, tool-history absence, first-message fingerprint drift
• Docs, README, Dockerfile, docker-compose aligned with all of the above