Session ses_384803f8

The system-reminder leak bug was that <system-reminder> tags from OpenCode (specifically the build-switch.txt which contains literal <system-reminder> tags) were leaking into the nuum gradient's message processing. The Layer 4 nuclear path in gradient.ts used stripToTextOnly without also calling cleanParts, so build-switch and plan reminders with literal <system-reminder> tags weren't being stripped. The fix added cleanParts to the Layer 4 nuclear path and inside stripToTextOnly itself.

[user]: What is the final eval score and the current state of improvements to the MEMORY_PLAN?

Based on the session context provided:

Final eval scores:

• Nuum: 14/15 (93.3%) — up from the baseline of 80.0% in this session
• Default: 10/15 (66.7%) — baseline comparison (though a separate run showed 8/15=53.3% with compaction-based approach that was being investigated/fixed)

Current state of MEMORY_PLAN improvements:

Phases 1 and 2 are done:

• Phase 1: Observation-log format (OM-style Observer/Reflector prompts, dated event log, priority tags)
• Phase 2: Temporal anchoring at read time (relative time annotations, gap markers)

New work done in this session (not yet in MEMORY_PLAN):

• Observer prompt refinements added: "EXACT NUMBERS — NEVER APPROXIMATE" section and "BUG FIXES AND CODE CHANGES — HIGH PRIORITY" section
• Reflector prompt refinements added: "EXACT NUMBERS" and "EARLY-SESSION CONTENT" rules for merge conflicts
• These changes drove the score from 80% → 93.3%
• Backfill script updated to use the refined prompts

Phase 3 (cross-session entity merging) and Phase 4 (incremental distillation) are still pending/future.

The one remaining failure is the "131 vs 130 test failures" question — the observer captured both numbers from different distillation segments and the model picked the wrong one. This is a number conflict resolution problem the reflector prompt refinement partially addresses but didn't fully solve in this case.