Add per-action health metrics to executor

[cmttt]

Summary

• Adds osprey.action_health metric emitted once per action after execution completes
• Adds osprey.action_error_count histogram for actions that had errors (10% sampled)
• Tags are all bounded: action (~100 values), had_errors (bool), had_unexpected_errors (bool), had_effects (bool)
• No behavioral changes — purely additive observability

Why

We need to quantify what % of Osprey actions execute to completion without failures. Currently there is no per-action health signal — failures are only visible at the UDF level, and "spammy" exceptions are suppressed from existing metrics entirely.

What it measures

• had_errors: True/False — whether ANY node errors occurred (including ExpectedUdfException)
• had_unexpected_errors: True/False — whether non-expected errors occurred (excludes ExpectedUdfException)
• had_effects: True/False — whether the action produced any effects (verdicts, labels, etc.)

Key queries

// % of actions with any failure
sum:osprey.action_health{had_errors:true} by {action}.as_count()
/ sum:osprey.action_health{*} by {action}.as_count() * 100

// % of actions with zero effects (potential enforcement failure)
sum:osprey.action_health{had_effects:false} by {action}.as_count()
/ sum:osprey.action_health{*} by {action}.as_count() * 100

Test plan

• Verify syntax validity (done — ast.parse passes)
• Run full test suite via Docker (./run-tests.sh)
• Deploy to staging and verify metrics appear in Datadog
• Monitor cardinality for 24h — estimated ~800 series