ltm: resolve cycle partitions for state-carrying module instances (Solo backstop under-merges genuinely-coupled module-internal-stock loops)

[bpowers]

Problem

The #750 fix (branch ltm-fix-batch-2, commit 03a9ec1f) stops None-partition loops from cross-normalizing by giving each one a singleton normalization group (NormGroup::Solo): sound, never pollutes, but it under-merges genuinely-coupled module-internal-stock loops. Two loops through SMOOTH instance A and SMOOTH instance B that share a parent path are genuinely coupled feedback in the same subsystem, yet each now normalizes alone -- both rel_loop_scores pin to sign-preserving +/-1, so their relative contributions cannot be compared even against each other.

The principled refinement (named and rejected-for-scope in the 03a9ec1f commit body): resolve partitions for state-carrying module instances instead of letting them fall to None:

• treat the module instance as a partition node in the stock-coupling graph, and
• map module-internal stock names (sub·stock) to the instance's partition,

so module-internal-stock loops land in the parent partition their instance participates in (or in a shared partition with their true siblings), and only genuinely-uncoupled loops keep the Solo backstop.

A further step composes with #773 / #749: treat PREVIOUS-holding variables as partition state nodes too, so lagged-state loops also get real partitions instead of Solo.

Why this is phase-sized (the honest assessment from the commit body)

• Spans three partition-construction sites: model_cycle_partitions (src/simlin-engine/src/db/analysis.rs ~2548), model_element_cycle_partitions (same file ~3057), and the discovery-graph build (ltm_finding.rs / CausalGraph::compute_cycle_partitions in src/simlin-engine/src/ltm/graph.rs ~1110).
• Changes FFI-visible partition stock sets -- the partition's stock-name set is the documented durable cross-surface identity key (ltm: expose cycle-partition metadata on exhaustive/pinned loop surfaces (SimlinLoop, Model.loops/Run.loops, MCP LoopDominanceSummary), keyed by stock-set identity #685), so adding module-instance/internal-stock names to it is an observable contract change across libsimlin, pysimlin, and MCP surfaces.
• The Solo backstop must stay regardless: PREVIOUS-lagged stockless loops (ltm: stockless PREVIOUS-lagged cycle is scored by the enumerator but rejected by pin validation (false 'no stock' rationale) #749) also land in None and are not covered by module-instance partition nodes alone (unless the further PREVIOUS-as-state-node step is taken).

Why it matters

Restores comparability of relative loop scores for module-heavy models (the exact value partition normalization exists to provide), where the singleton rule currently degenerates every module-internal loop to +/-1. Severity: low-medium; absolute scores unaffected, the Solo behavior is sound just uninformative.

Components affected

• src/simlin-engine/src/db/analysis.rs (model_cycle_partitions, model_element_cycle_partitions)
• src/simlin-engine/src/ltm/graph.rs (compute_cycle_partitions) / src/simlin-engine/src/ltm/partitions.rs (partition_for_loop's module-internal None resolution)
• src/simlin-engine/src/ltm_finding.rs (discovery-graph partition grouping)
• Downstream partition-metadata consumers: src/libsimlin (partition stock sets), pysimlin, MCP (ltm: expose cycle-partition metadata on exhaustive/pinned loop surfaces (SimlinLoop, Model.loops/Run.loops, MCP LoopDominanceSummary), keyed by stock-set identity #685)

Cross-references

• ltm: None-partition loops share one default rel_loop_score normalization bucket, cross-normalizing unrelated module-internal-stock loops #750 -- the None-bucket cross-normalization bug; the singleton-group fix closing it is the backstop this issue refines (commit 03a9ec1f)
• ltm: rel_loop_score normalisation cross-pollutes independent A2A loops #487 (closed) -- the original A2A-slot cross-pollution class
• ltm: expose cycle-partition id/size on discovered loops so singleton-partition loops can be grouped or filtered #671 (closed) -- the singleton-partition +/-1 legibility cost this refinement reduces
• ltm: expose cycle-partition metadata on exhaustive/pinned loop surfaces (SimlinLoop, Model.loops/Run.loops, MCP LoopDominanceSummary), keyed by stock-set identity #685 -- partition-metadata surfacing parity; partition identity is keyed by stock-name set, which this change touches
• ltm: module-internal PREVIOUS lag is uncounted state on every surface (lag-only sub-model loops unscored and unpinnable) #773 / ltm: stockless PREVIOUS-lagged cycle is scored by the enumerator but rejected by pin validation (false 'no stock' rationale) #749 -- PREVIOUS-lagged state; the optional PREVIOUS-as-partition-node step composes with those
• Part of epic epic: Loops that Matter (LTM) -- tracking known issues and improvements #488

Discovery context

Named as the rejected-for-scope principled alternative in commit 03a9ec1f on ltm-fix-batch-2 (the #750 fix), 2026-06-09; filed so the refinement is independently trackable.