fix(pipeline): fail-loudly hygiene + operator override hatches

[GrigoryEvko]

Four small hygiene fixes around pipeline construction. Theme: configuration surfaces that silently accept ambiguous input are harder to debug than ones that fail loudly.

StageRegistry.register silently overwrote cls._stages[class_name] on duplicate registration. Two modules decorating distinct Stage subclasses with the same class name would shadow each other invisibly; downstream get_stage(name).import_path lookups would resolve to whichever decorator imported last (load-order dependent). Now raises ValueError listing both import paths. Tests that intentionally re-register can call StageRegistry.clear() between registrations — the existing escape hatch.

PipelineBuilder.add_stage had the same silent-overwrite pattern: self._nodes[name] = factory keyed by a caller-supplied string with no collision guard. The class already exposes a replace_stage method as the explicit overwrite path, so making add_stage raise on duplicate is consistent with intent. A subclass that copy-pastes a parent's add_stage call now fails at construction instead of producing a pipeline where downstream wiring references the original factory by name while the registry holds the shadow.

DefaultPipelineBuilder._contribute_default_deps reassigned the whole self._deps dict via self._deps = {...} literal. A subclass that overrides _contribute_default_deps, adds its own deps to self._deps, and then calls super()._contribute_default_deps() would silently lose everything it added. Switched to per-key self._deps.setdefault(stage, []).extend(deps) — subclass-added deps for unrelated stages survive, and deps for stages the parent also contributes are concatenated rather than overwritten.

select_pipeline_builder, CMAOptPipelineBuilder.__init__, and OptunaOptPipelineBuilder.__init__ all branched on problem_context.is_contextual to decide whether to add context wiring, with no operator override path. When the filesystem-inferred heuristic guesses wrong for a specific run, the only escape was switching to a different pipeline YAML. Added a force: str | None = None kwarg accepting "context" / "default" to all three call sites — overrideable from Hydra via +pipeline_builder.force=default. Invalid values raise so typos surface immediately rather than falling through to the heuristic. Factored validation into a shared _resolve_context_flag helper so the three call sites stay in sync.