fix(templates): strict prompt rendering + steps.X.output registry

[addadi]

Fixes #40

resolveNewExpression silently returned an empty string for any unknown expression or missing key, and had no steps. prefix handler — the path used by executeTaskNode prompt rendering. So {{steps.X.output}} rendered as an empty prompt which the worker then answered from nothing (or hallucinated from residual session context). No error, no failed step: the failure is invisible at the orchestration layer.

Changes:

• Add renderPromptTemplateStrict which propagates UnresolvedReference, UnknownExpression, and StepNotFound instead of swallowing them. executeTaskNode now uses the strict variant — a missing variable fails the node visibly.
• Parametrize resolveNewExpression with a strict flag; the lenient path is preserved for {% if %} conditional evaluation where empty-as-falsy is legitimate.
• Add a {{steps.X.Y}} prefix handler resolving against a state.steps.X.Y registry. The registry is populated from each step's raw output (buildStepRegistryUpdate) and deep-merged via state.applyDeepUpdates so multi-step chains coexist without last_value clobbering. {{steps.X.output}} works with no per-step output_key configured.
• workflow_validation.checkStateRefs skips the state-schema check for {{steps.*}} refs since the registry is runtime-populated.

Integration test: 2-step sequential — step a emits SENTINEL_XYZ, step b renders "Echo: {{steps.a.output}}" verbatim; a and b coexist in the registry after deep-merge. Full suite green (zig build test, 366/366).