Triage bot: move to gpt-5-mini + reasoning-aware decode + sharper prompts

Upgrade the Discord triage bot's model + decode + prompting.

Model: default LLM_MODEL → gpt-5-mini (was gpt-4o-mini).

Reasoning-aware decode (OpenAiLlmClient): the adapter now picks decode knobs per model family. Reasoning models (gpt-5*, o-series) set reasoning_effort + verbosity (and must NOT send a custom temperature — gpt-5 fixes it at 1); classic models get a low temperature (default 0.2) so the structured JSON decision is stable run-to-run. New config: tesks.llm.reasoning-effort (default low), tesks.llm.verbosity (default low), optional temperature, optional max-completion-tokens. Helpers unit-tested (OpenAiLlmClientTest).

Prompting:

• "Writing the task well" guidance — specific imperative titles (<80 chars), self-contained markdown descriptions (summary + repro/expected-vs-actual/acceptance), no raw-transcript paste, no invented detail; infer priority/labels only from clear signals.
• Stronger duplicate avoidance in the agentic loop: search_tasks once before a CREATE unless plainly novel, and prefer UPDATE/COMMENT when a result already covers the request; get_task to confirm a named ref before UPDATE/COMMENT.

Full backend suite green incl. the agentic-loop tests.