feat(contracts): implement interval temporal logic and happens-before ordering (#729) - #754

Resolves #729

Summary

Extends agentwatch.contracts with a Lamport happens-before partial order engine and interval temporal logic. Distributed multi-agent workflows can produce false refutations when strict wall-clock ordering ignores clock drift, network jitter, or asynchronous arrival. This change combines declared span links, intra-stream sequence monotonicity, bounded clock uncertainty intervals, and event-time watermark buffering.

Design Details

1. Happens-before partial order engine

   • Constructs a DAG from intra-stream sequence ordering and supported DeclaredLink relationships.
   • Supports source-ID resolution and transitive reachability.
   • Detects cycles and contradictory causal evidence.
   • Classifies relations as BEFORE, AFTER, CONCURRENT, or CONTRADICTORY.

2. Interval timestamp logic

   • Models bounded timestamp uncertainty using TimestampInterval.
   • Evaluates strict interval precedence and overlap.
   • Prevents overlapping, causally independent events from causing false REFUTED verdicts.

3. Out-of-order watermark buffer

   • Tracks an event-time watermark and bounded retention horizon.
   • Classifies arrivals as ACCEPTED, LATE, DROPPED_TOO_OLD, or DUPLICATE.
   • Identifies affected targets for incremental recomputation and evicts expired data.

4. Evaluator integration

   • Detects causal cycles and returns CONFLICTING_EVIDENCE.
   • Integrates causal ordering into PRECEDING_EVENT and EVENTUALLY_WITHIN while preserving time-window constraints.

5. Compatibility

   • Adds public exports in agentwatch/contracts/__init__.py.
   • Includes regression tests for existing sequence-gap and coverage behavior.

Test Evidence

• 19 tests in test_happens_before_order.py.
• 120 deterministic out-of-order simulation streams.
• Reported contract suite: 161 passed.
• Reported full suite: 1,395 passed, 7 skipped, 0 failed.
• Reported coverage: 89% across agentwatch.contracts.
• Reported Ruff checks and formatting passed; Bandit reported zero findings.

Test and quality-gate results above reflect the recorded implementation verification report.

Scope

This PR addresses AgentWatch Issue #729.