ltm-finding: truncate-before-filter can drop partition-dominant loops

[bpowers]

Summary

rank_and_filter() truncates discovered loops to MAX_LOOPS (200) before partition-aware contribution filtering.

Why this is a problem

LTM interpretation is partition-scoped: loops should be compared within cycle partitions. Global pre-truncation can drop loops that are dominant in small partitions but globally below rank 200, violating partition-local dominance semantics.

Literature / algorithm alignment

Reference and design docs emphasize partition-scoped comparison for relative contribution. Global truncation before partition-aware retention introduces cross-partition interference.

Current implementation behavior

Order in rank_and_filter():

1. global sort by avg abs score
2. truncate to 200
3. partition-aware per-timestep contribution filter

Location:

• src/simlin-engine/src/ltm_finding.rs (rank_and_filter)

Impact

• False negatives for partition-dominant loops in large models.
• Reported loop set can underrepresent smaller disconnected feedback subsystems.

Success criteria

1. Reorder filtering so partition-aware retention is applied before any global cap, or apply per-partition quotas prior to global cap.
2. Define and document deterministic tie-breaking when the final capped set still exceeds MAX_LOOPS.
3. Add tests with >200 loops spanning multiple partitions where a low-global-score but high-partition-share loop is retained.
4. Keep deterministic IDs and output ordering stable.

Notes

This issue concerns post-discovery ranking/filtering only, not the known flow-to-stock timing choice.