Add machine-verifiable assertions to SentienceAgent examples

[rcholic]

Summary

This PR demonstrates the Sentience SDK's AgentRuntime verification system - a declarative approach to agent observability that makes AI agent behavior deterministic, debuggable, and machine-verifiable.

Why This Matters

Traditional AI agents are black boxes: you prompt them, they act, and you hope for the best. When things go wrong, debugging is painful - you're left parsing logs and screenshots trying to understand what the agent saw vs what it did.

Sentience SDK flips this model:

Traditional Agent	Sentience-Verified Agent
"Did it work?" → Check screenshots	"Did it work?" → assert_done() returns true/false
Debug via logs	Debug via structured trace with per-step assertions
Non-deterministic validation	Machine-verifiable predicates
Manual QA	Automated regression testing

Key Features Showcased

1. Per-Step Assertions - Guardrails During Execution

step_assertions = [
    {"predicate": url_contains("news.ycombinator.com"), "label": "on_hackernews", "required": True},
    {"predicate": exists("role=link text~'Show HN'"), "label": "show_hn_posts_visible"},
    {"predicate": not_exists("text~'Error'"), "label": "no_error_message"},
]

• required=True: Agent fails fast if this doesn't pass (prevents drift)
• Semantic selectors: Query by role + text, not brittle CSS selectors
• Per-step validation: Catch issues immediately, not at the end

2. Declarative Task Completion - Know When You're Done

done_assertion = all_of(
    url_contains("news.ycombinator.com/show"),
    exists("role=link text~'Show HN'"),
)

• Combinators: all_of(), any_of() for complex conditions
• Machine-verifiable: No LLM judgment needed - pure predicate evaluation
• Deterministic: Same state → same result, every time

3. Structured Trace Output - Studio Timeline Integration

trace_dir="traces" # Outputs JSON traces for Sentience Studio

• Every assertion result is logged with timestamp, label, and outcome
• Visualize agent execution in Studio timeline
• Debug failures by inspecting exactly which predicate failed and why

Verification Output

`🔍 Verification Summary:
All assertions passed: True
Required assertions passed: True
Task verified complete: True

Assertion Details (4 total):
✅ on_hackernews (required)
✅ show_hn_posts_visible
✅ no_error_message
✅ task_complete (required)
`