Problem

Journal events capture what happened but lack the structured metadata needed for post-mortem analysis:

• No duration tracking on unit executions — can't identify slow units
• No error classification — have to grep raw messages to understand failure modes
• No resource metadata — can't correlate issues with specific models or GSD versions
• Stuck-detection recovery is invisible in the journal — no events emitted when it fires

Solution

Enrich journal events with OTel-inspired fields:

• iteration-start — resource block (gsdVersion, model, cwd) + causedBy from stuck recovery
• unit-start — sessionId + messageOffset for session correlation
• unit-end — durationMs, error/errorType classification (structured errorContext preferred over regex)
• New stuck-detected event with level and reason

Forensics scanner surfaces slow units (>2min) and error distributions as anomalies.

Implementation

PR: gsd-build#2618

Builds on the structured ErrorContext from gsd-build#2612 — error classification in unit-end prefers errorContext.category over regex heuristics.