docs: rewrite reasoning/termination/events/evaluation/state concept pages

[fede-kamel]

Summary

Second concept-page batch using the same template as PR #29 and the
provider pages: pitch / when to pick / getting started / capabilities
(each runnable) / gotchas / source / see also.

• reasoning.md (59→141) — lead with the wrong-premise problem, three add-ons each with a runnable snippet, fixed combination order, common gotchas.
• termination.md (65→139) — lead with the `|` / `&` algebra; runnable predicates; inspect-what-stopped recipe; built-in conditions table; custom conditions; gotchas (operator precedence, `ConfidenceMet` requires `reflexion=True`, etc).
• events.md (63→191) — reference page covering every `LocusEvent` subtype with the actual fields from `src/locus/core/events.py` (`ThinkEvent.reasoning`, `ToolStartEvent.arguments`, `ReflectEvent.assessment`, `GroundingEvent.score`, …). Multi-agent and causal events documented.
• evaluation.md (79→139) — align doc with real `EvalCase` API. The previous version showed an `expected={...}` predicate dict that doesn't exist in the framework; replaced with the actual `expected_tools`, `expected_output_contains`, `expected_output_not_contains`, `max_iterations`, `max_duration_ms` fields. Added LLM-as-judge pattern.
• state.md (61→123) — explain immutable `AgentState` (`tuple` / `frozenset` everywhere, `with_*` builders), `to_checkpoint`/`from_checkpoint` round-trip, reducers for graph branches.

Test plan

• `hatch -e docs run mkdocs build --strict` — clean (0 warnings, no broken anchors)
• `pre-commit run --files docs/concepts/{reasoning,termination,events,evaluation,state}.md` — codespell and markdownlint both pass
• DCO sign-off on the commit