Problem

During /nf:solve Haiku classification (Phase 1b), 300 of 2334 items (12.8%) failed to classify. The classification agent reported "failed": 300 but no error details — we don't know if these were API timeouts, context length issues, or malformed inputs.

Observed Output

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Haiku Classification Results
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Genuine:  11
 FP:       594
 Review:   19

 From cache: 0 | Newly classified: 2034 | Failed: 300
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Expected Behavior

1. Log failure reasons — at minimum a breakdown by error type (timeout, token limit, parse error)
2. Retry failed items — at least once with backoff
3. Report which items failed — so users can manually classify or investigate
4. Persist partial results — currently the 2034 successes are saved, but the 300 failures are silently dropped (they default to "unclassified" which is treated as FP in some flows)

Impact

A 12.8% failure rate means ~1 in 8 items go unclassified. If any genuine issues are among the 300, they'll be missed entirely.