epic: Loops that Matter (LTM) -- tracking known issues and improvements

[bpowers]

Purpose

Single source of truth for open work in the Loops that Matter (LTM) subsystem -- find a bite-sized piece to work on without re-investigating subsystem state. For an LTM overview see src/simlin-engine/CLAUDE.md and docs/architecture.md; reference and design docs live under docs/reference/ and docs/design/.

How to use this tracking issue

• Picking up work? Browse the clusters below; every entry is an open issue with the ltm label that stands alone with file:line references and a suggested fix.
• Filing something new? Add it to the cluster whose root cause it shares, as a one-line entry: - [ ] #N -- short title [severity]. Expository detail belongs in the issue, not here.
• Closing something? Tick the box and move the line to the Done section at the bottom.
• Command line: gh issue list --label ltm --state open returns the full open set.

Strategy

Issue-by-issue fixing produced a roughly flat open count across PRs #740, #761, and #775: the LTM shape space (subscript form x reducer kind x module boundary x scoring surface) is combinatorial, and each per-shape fix carves out newly-visible neighbors. So the plan is root-cause phases that each close an entire cluster, not more single-issue batches:

• Phase 1 -- read-slice / shape-expressiveness generalization -- DONE (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784, 2026-06-11; residuals closed out by PRs engine: close ltm phase 1 residuals #795 and engine: enumerate rank element loops #796): AxisRead/read_slice_rows as the single per-axis access vocabulary retired the 10 per-shape carve-outs (ltm: iterated+literal mixed subscript keeps the cross-product element expansion -- phantom cross-element loops carry silent nonzero scores (rescoped: hard compile failure fixed) #525 ltm-augment: is_other_dep_iterated_dim_subscript collapse is magnitude-imprecise for a transposed non-live array dep #526 ltm: arrayed co-agg frozen bare in agg→target partial when a target hoists two arrayed reducers (fragment fails compile, stubs to 0) #751 ltm: subscripted reverse-declared mapped reference classifies DynamicIndex, keeping the cross-product despite the compiler's diagonal #757 ltm: variable-backed partial reducers with broadcast or permuted axes keep the conservative cross-product (loop scores stub to 0) #764 ltm: cross-dimensional partial-reduce link scores ignore Pinned axes in the co-reduced slice (inflated divisor + garbage unread-row scores) #765 ltm: compute_read_slice maps a StarRange over a proper subdimension to Reduced over the full parent extent #766 ltm: extend multi-source agg hoisting to feeder-sub-slice combinations (co-source-closure loops stay warned-zero) #767 ltm: FixedIndex reference into a disjoint-dim ApplyToAll target is loud-skipped though recoverable via the #510 per-element construction #769 ltm: enumerate_agg_nodes hoists whole-extent array-valued RANK into a scalar agg node that cannot compile (agg-routed loop scores zeroed) #771); design doc docs/design-plans/2026-06-11-ltm-shape-expressiveness.md.
• Phase 2 (next) -- module state & pathway visibility (closes cluster B): module-internal state (stocks AND lags) counted consistently by the gates, pin validation, partitioning, and pathway/composite emission.
• Phase 3 -- FFI / surface parity sweep (closes cluster C): mechanical plumbing of data the engine already computes out through libsimlin, TS, pysimlin, and MCP.
• The engine element-map execution fix (engine: executed A2A lowering ignores explicit element maps and resolves positionally (silent wrong results for permutation maps) #756, with engine: different-cardinality element-map reference fails to compile (failed to compile fragments) #753 -- engine-labeled, tracked in epic: backlog burndown -- tracking all open work #733) unlocks cluster D.

Standing invariants, non-negotiable in every phase:

• No silent wrong numbers. An un-scoreable shape degrades to one loud warning plus dropped loop scores -- never garbage values, never quiet zeros.
• Two-surface decisions share one predicate. Detected/scored/pinned surfaces must agree by construction (one model_is_stateless, one element-correspondence helper), not by parallel reimplementation.
• The scoreboard is the open medium+ count and the silent-wrong-number count (currently ~0) -- not the raw issue count.
• Prevention beats archaeology. Extend the generated-corpus invariants (the element_graph_proptest pattern) so new defect classes fail in CI instead of being filed from review. Tracked in cluster I.

Open work, clustered by root cause

A. Read-slice / shape expressiveness (Phase 1 residuals)

Phase 1 (PR #784) replaced the identical-whole-extent-slice assumption with the per-axis AxisRead vocabulary. The residuals filed during that work were closed out by PRs #795 and #796; see the Done section.

• ltm: bare reducer argument in a per-element (Ast::Arrayed) slot is declined loudly -- spell the slot's pinned element so the hoisted node is per slot #1041 -- bare reducer argument in a per-element (Ast::Arrayed) slot is declined loudly; spell the slot's pinned element so the hoisted node is per slot [low-medium]

B. Module state & pathway visibility (Phase 2 target)

Module-internal state and module-boundary edges are invisible to one or more of: the stateless gate, pin validation, partitioning, link-score emission, pathway/composite scoring.

• LTM: module stock enrichment misses stocks reachable only through internal feedback cycles #282 -- module stock enrichment misses stocks reachable only via internal cycles [low]
• ltm: a generated reference to an absent cross-module LTM variable compiles cleanly and reads a silent constant 0 #683 -- absent cross-module LTM variable compiles to a silent constant 0 [medium]
• ltm: scalar module output feeding an arrayed reader emits a single scalar constant-0 link score, dropping all arrayed loops through modules #716 -- scalar module output feeding arrayed reader drops all arrayed loops [medium]
• ltm: loops through SMOOTH/DELAY/TREND builtin calls are unpinnable (synthetic module node unnameable and unbridgeable by order_variable_cycle) #747 -- loops through SMOOTH/DELAY/TREND builtin calls are unpinnable [low-medium]
• ltm: sub-model pathway/composite scores through an unscoreable (loud-skipped) or agg-routed edge silently multiply to 0 #768 -- pathway/composite scores through a loud-skipped or agg-routed edge silently zero [low-medium]
• ltm: model_detected_loops has no stateless gate -- stateless-but-cyclic shapes report unscoreable loops and silently drop pins #772 -- detected surface has no stateless gate; pins silently dropped (absorbs ltm: pin on a genuinely stateless model bypasses validation -- typo'd pin gets no invalid-pin warning #770) [low]
• ltm: module-internal PREVIOUS lag is uncounted state on every surface (lag-only sub-model loops unscored and unpinnable) #773 -- module-internal PREVIOUS lag is uncounted state on every surface [low-medium]
• ltm: resolve cycle partitions for state-carrying module instances (Solo backstop under-merges genuinely-coupled module-internal-stock loops) #774 -- partition state-carrying module instances (Solo backstop under-merges) [low-medium]

C. FFI / TS / pysimlin surface parity (Phase 3 target, mechanical)

The engine already computes the data; it never reaches one or more external surfaces.

• ltm: surface synthetic variable names via FFI helpers for advanced users #507 -- expose synthetic-variable name builders via FFI/pysimlin/TS [low]
• pysimlin: surface the user-supplied loop name on Loop objects (pinned loops show only pin{n} ids) #659 -- user-supplied pinned-loop name never reaches Python [low]
• libsimlin: get_links is variable-level / slot-0 only for arrayed models #665 -- get_links is variable-level / slot-0 only for arrayed models [low-medium]
• ltm: no query-time PATHSCORE helper (multiply link scores along a named path, element-aware) #669 -- no query-time PATHSCORE helper (element-aware link-score product) [low]
• 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 missing from exhaustive/pinned loop surfaces [low]
• ltm: discovery-mode agg_recovery_truncated signal never reaches FFI/pysimlin/TS (asymmetry vs exhaustive) #701 -- discovery agg_recovery_truncated never reaches FFI/pysimlin/TS [low]
• diagram: render Rux/Bux loop polarity with confidence-gray treatment (deferred from #495) #731 -- diagram: render Rux/Bux polarity with confidence-gray treatment [low]
• engine: @simlin/engine never binds simlin_analyze_get_loops_runtime (runtime Rux/Bux polarity and real confidence unreachable from TS) #760 -- @simlin/engine never binds simlin_analyze_get_loops_runtime [low-medium]

D. Engine element-map execution followers

Gated on engine #756 (executed A2A lowering resolves element maps positionally; see also #753) -- both engine-labeled and tracked in #733, listed here for visibility. Once fixed, also re-enable the element-map carve-outs recorded in the mapped_element_correspondence rustdocs.

No open entries: the one LTM-side follower (#754) closed with PR #795. The cluster stays as the marker for the carve-out re-enablement still owed once #756 lands.

E. Discovery & post-sim quality

• ltm-finding: discovery runs at save_step instead of dt #309 -- discovery runs at save_step instead of dt; sub-step transients missed [medium]
• ltm-finding: remove hardcoded 'main' model assumption #313 -- discovery hardcodes the 'main' model assumption [low]
• ltm-finding: SearchGraph clones Ident<Canonical> in per-timestep DFS hot path #481 -- SearchGraph clones idents in the per-timestep DFS hot path [medium]
• ltm-finding: global MAX_LOOPS truncation is partition-blind past 200 partitions (no per-partition floor) #677 -- global MAX_LOOPS truncation is partition-blind past 200 partitions [low]
• ltm-post: dominant-period flicker on data-driven models (instantaneous shares inherit input-data noise) #670 -- dominant-period flicker on data-driven models (needs windowed view) [low-medium]
• ltm: arrayed pinned loops' no-element default aggregate breaks share-based analysis (Run.loops behavior_time_series) #672 -- arrayed pinned loops' no-element default aggregate breaks share analysis [low-medium]
• ltm: surface diagnostic when all loop scores are zero (equilibrium model) #504 -- no diagnostic when all loop scores are zero (equilibrium model) [low]

F. Performance & scale

• ltm: residual ~30k synthesized PREVIOUS helper-aux vars on C-LEARN threaten the VM 65k-slot ceiling #654 -- ~30k synthesized PREVIOUS helper auxes threaten the 65k VM slot ceiling [medium]
• ltm: C-LEARN LTM compile-time dominated by per-fragment fixed overhead (DimensionsContext clone, O(D^2) intern_name, repeated parse) #655 -- LTM compile time dominated by per-fragment fixed overhead [medium]
• engine: LTM element-level circuit enumeration spends most of its time in string canonicalization/rotation churn (cross-product shapes) #955 -- element-level circuit enumeration dominated by string canonicalization/rotation churn on cross-product shapes [medium]
• ltm: link-score generation materializes one full guard-form arm per element of an arrayed target (quadratic in element count) #977 -- link-score generation is quadratic in an arrayed target's element count ((N+1)^2 equation arms per edge) [medium]

G. Capability features (deliberate backlog, not bugs)

• ltm: implement LOOPSCORE and PATHSCORE builtins for user-pinned loops #484 -- implement LOOPSCORE / PATHSCORE builtins for user-pinned loops [low]
• ltm: collapse mixed-diagonal loop families into one arrayed loop score (enumerator + pins) #658 -- collapse mixed-diagonal loop families into one arrayed loop score [low]
• ltm: collapse mapped-diagonal loop families into one arrayed loop score (cross-element check defeats grouping across dimension mappings) #755 -- collapse mapped-diagonal loop families into one arrayed loop score [low]
• ltm: pinned loops cannot express both directed cycles over a shared variable set #674 -- pins cannot express both directed cycles over a shared variable set [low-medium]

H. Doc / decide

• ltm-augment: TIME-bearing reducer bodies break the PREVIOUS(agg) anchor in body-aware partials (nonlinear arm attributes full time-drift to every row) #763 -- TIME-bearing reducer bodies break the PREVIOUS(agg) anchor; decide delta-ratio-bail vs same-sample anchoring, then document the chosen limitation [low-medium]
• ltm-augment: instantaneous and flow-to-stock link scores use inconsistent time windows within a loop #544 -- document the intra-loop window mismatch: instantaneous links score over [t-dt, t] while flow-to-stock numerators score over [t-2dt, t-dt]; the reference doc's 1-DT-shift note (Section 3.2) explicitly defers this and the design doc's divergence enlarge clouds #6 covers only the cross-tool offset (ltm: document flow-to-stock 1-DT-shift convention in user-facing reference #505), so a one-paragraph addition to both is still owed [low]

I. Prevention -- generated-corpus harnesses

The invariants that keep a defect class from being re-filed from review. See the "prevention beats archaeology" invariant above.

• ltm: every-emitted-fragment-compiles harness over a generated corpus (prevention) #782 -- every-emitted-fragment-compiles harness over a generated corpus [medium]

J. Ceteris-paribus wrap correctness

wrap_non_matching_in_previous decides, per argument, what is held at its previous value. Getting that wrong yields silently wrong partials rather than a loud skip.

• ltm: the ceteris-paribus wrap leaves a LOOKUP table argument's runtime index live in every link-score partial #984 -- the LOOKUP arm holds the whole table argument verbatim, leaving a runtime subscript index live in every link-score partial [medium]

Done

Compact record; expository detail lives in issue/PR history. The old GH-issue/tech-debt cross-reference table is removed -- every row referenced closed items (historical detail in issue/PR history).

PR #796 (rank element loops, 2026-06)

• ltm: de-hoisted array-valued RANK gets diagonal-only element edges, so rank-mediated cross-element loops are not enumerated #776 -- de-hoisted array-valued RANK gets diagonal-only element edges; rank-mediated cross-element loops now enumerated (PR engine: enumerate rank element loops #796)

PR #795 (Phase 1 residual closeout, 2026-06)

• ltm-finding: discovery mode mints phantom from-side element keys for Bare A2A edges (mapped-dimension and scalar-source legs; loops undiscoverable) #754 -- discovery minted phantom from-side element keys for mapped and scalar-source A2A edges (PR engine: close ltm phase 1 residuals #795)
• ltm: arrayed-owner Pinned/subset broadcast reduce slice loud-skipped; needs the PerElement broadcast rule on variable-backed owners #777 -- arrayed-owner Pinned/subset broadcast reduce slice loud-skipped (PerElement rule on variable-backed owners) (PR engine: close ltm phase 1 residuals #795)
• ltm: duplicated-result-dim aggregate fans out phantom off-diagonal agg->target element edges (warned 0-stub loops) #778 -- duplicated-result-dim agg fans out phantom off-diagonal agg->target edges (warned 0-stub loops) (PR engine: close ltm phase 1 residuals #795)
• ltm: bare-spelled feeder of un-hoisted multi-source reducer gets silent wrong changed-last loop score (per-element partial vs broadcast execution) #779 -- bare-spelled feeder of un-hoisted multi-source reducer scores silently wrong (~3x) via the changed-last partial (PR engine: close ltm phase 1 residuals #795)
• ltm: PartialEquationError link-score skips don't drop dependent loop scores via unscoreable_edges #780 -- PartialEquationError link-score skips don't drop dependent loop scores via unscoreable_edges (PR engine: close ltm phase 1 residuals #795)
• ltm: emit_agg_routed_edges enumerates rows via its own AxisPlan, deviating from the design's I4 single-derivation invariant #783 -- emit_agg_routed_edges re-derives rows via its own AxisPlan instead of read_slice_rows (I4 single-derivation violation) (PR engine: close ltm phase 1 residuals #795)
• ltm: square-source aggs (repeated dim in result_dims) pin by dim name across co-source and agg-to-target halves; only the feeder half is defended #785 -- square-source agg (repeated dim in result_dims) pins by first-match dim name on co-source and agg-to-target halves (PR engine: close ltm phase 1 residuals #795)
• ltm: bare arrayed reducer arg consumed per-slot by execution is hoisted as a whole-array scalar agg (silent wrong link and loop scores) #788 -- bare arrayed reducer arg consumed per-slot by execution hoists as a whole-array scalar agg (PR engine: close ltm phase 1 residuals #795)
• ltm: scalar feeder of a whole-RHS variable-backed reduce emits non-compiling link scores (warned 0-stub loops instead of the #758 drop) #790 -- scalar feeder of a whole-RHS variable-backed reduce emits non-compiling link scores (PR engine: close ltm phase 1 residuals #795)
• ltm: I1-declined multi-source reducer with subset-dim arrayed owner falls through to cartesian derivation (silent +1.0 link scores on unread rows) #791 -- I1-declined multi-source reducer with subset-dim arrayed owner falls through to the cartesian derivation (PR engine: close ltm phase 1 residuals #795)

PR #784 (Phase 1: shape-expressiveness generalization, 2026-06)

One per-axis access model (AxisRead/read_slice_rows); design doc docs/design-plans/2026-06-11-ltm-shape-expressiveness.md.

• ltm: iterated+literal mixed subscript keeps the cross-product element expansion -- phantom cross-element loops carry silent nonzero scores (rescoped: hard compile failure fixed) #525 -- iterated+literal mixed subscripts classify PerElement; phantom cross-element loops gone (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm-augment: is_other_dep_iterated_dim_subscript collapse is magnitude-imprecise for a transposed non-live array dep #526 -- transposed non-live array dep routes to changed-last instead of freezing the wrong element (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: arrayed co-agg frozen bare in agg→target partial when a target hoists two arrayed reducers (fragment fails compile, stubs to 0) #751 -- frozen co-aggs compile pinned to their slots via per-ident pin maps (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: subscripted reverse-declared mapped reference classifies DynamicIndex, keeping the cross-product despite the compiler's diagonal #757 -- reverse-declared mapped references keep the compiler's diagonal (mapped classification in both declaration directions) (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: variable-backed partial reducers with broadcast or permuted axes keep the conservative cross-product (loop scores stub to 0) #764 -- broadcast/permuted whole-RHS variable-backed reduces mint synthetic aggs (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: cross-dimensional partial-reduce link scores ignore Pinned axes in the co-reduced slice (inflated divisor + garbage unread-row scores) #765 -- co-reduced slices honor Pinned axes; unread rows get neither edges nor scores (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: compute_read_slice maps a StarRange over a proper subdimension to Reduced over the full parent extent #766 -- StarRange over a proper subdimension reads Reduced{subset} (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: extend multi-source agg hoisting to feeder-sub-slice combinations (co-source-closure loops stay warned-zero) #767 -- iterated-dim projection feeders hoist with exactly-complementary changed-last scores (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: FixedIndex reference into a disjoint-dim ApplyToAll target is loud-skipped though recoverable via the #510 per-element construction #769 -- FixedIndex into disjoint-dim ApplyToAll target recovered per-element (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)
• ltm: enumerate_agg_nodes hoists whole-extent array-valued RANK into a scalar agg node that cannot compile (agg-routed loop scores zeroed) #771 -- whole-extent array-valued RANK de-hoisted via the value-shape precondition (PR LTM shape expressiveness: one per-axis access model (epic #488 phase 1) #784)

PR #775 (scoring-correctness and visibility batch, 2026-06)

• ltm: failed implicit-helper fragment compile emits no diagnostic (model_ltm_fragment_diagnostics skips model_ltm_implicit_var_info) #741 -- implicit-helper fragment-compile failures now diagnosed (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: PREVIOUS-captured array-valued non-reducing builtin (RANK) scalarizes into ill-typed stubbed helpers, corrupting link scores #742 -- PREVIOUS-captured array-valued builtin (RANK) scalarization (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: un-hoisted multi-source iterated-dim-feeder reducer emits silent garbage arrayed loop score (combined_read_slice slice-disagreement carve-out) #743 -- un-hoisted multi-source iterated-dim-feeder garbage loop score (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: source->agg link score wrong-signed for hoisted multi-source reducer with sign-flipping coefficient (SUM linear shortcut ignores the body's coefficient) #744 -- wrong-signed source-to-agg score under sign-flipping coefficient (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: recover_agg_hop_polarities never matches subscripted arrayed-agg endpoints, so arrayed-agg loops always classify Undetermined on the scored/pinned surfaces #745 -- arrayed-agg endpoint polarity recovery (no more forced Undetermined) (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: arrayed-cycle detected/scored loop-id namespace overlap lets the runtime id join read another loop's series #746 -- arrayed-cycle detected/scored loop-id namespace overlap (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: model_ltm_variables' stock-free early return is module-blind -- a root whose only state is module-internal silently emits no LTM variables #748 -- module-blind stock-free early return (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: stockless PREVIOUS-lagged cycle is scored by the enumerator but rejected by pin validation (false 'no stock' rationale) #749 -- stockless PREVIOUS-lagged cycle pin/enumeration agreement (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: None-partition loops share one default rel_loop_score normalization bucket, cross-normalizing unrelated module-internal-stock loops #750 -- None-partition loops no longer cross-normalize (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: loop scores through a variable-backed partial-reduce reducer reference never-emitted link-score names, stubbing to 0 #752 -- variable-backed partial-reduce loop scores reference real names (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm: declined element-mapped sliced reducer emits a non-compiling scalar link score on the conservative path (loop scores stub to 0) #758 -- declined element-mapped reducer's non-compiling conservative score (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm-augment: partial builders PREVIOUS-wrap dimension-name subscript indices (Wildcard and Bare shapes), silently stubbing scores #759 -- dimension-name subscript indices no longer PREVIOUS-wrapped (PR ltm: scoring-correctness and visibility batch from epic #488 #775)
• ltm-augment: generate_nonlinear_partial is body-blind for non-bare MIN/MAX/STDDEV reducer bodies (garbage per-row link scores) #762 -- body-aware nonlinear partials for MIN/MAX/STDDEV (PR ltm: scoring-correctness and visibility batch from epic #488 #775)

PR #761 (core-algorithm batch, 2026-06)

• ltm: loop-score builder's PureScalar fast path never routes a loop through a scalar synthetic agg node (silent zero loop score) #737 -- PureScalar loop fast path routes ThroughAgg (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: synthetic agg node for SUM(arrayed[*] * scalar) with a scalar target fails to compile and is stubbed to constant 0 #738 -- scalar-target agg fragment now compiles (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: element-graph projection proptest never generates inlined reducers, so ThroughAgg routing is verified only vacuously #739 -- element-graph proptest generates inlined reducers (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: cyclic_orderings' (m-1)!/2 mirror-skip rationale is incorrect; cross-agg loop count is arbitrary #676 -- cyclic_orderings mirror-skip rationale corrected (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: pin whose only stock is module-internal is wrongly rejected as 'contains no stock' #673 -- module-internal-stock pins validate correctly (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: agg→target link score over-subscripts an arrayed synthetic agg in the broadcast case (agg[D1] feeding to[D1,D2]) #528 -- agg-to-target broadcast projection in link-score equations (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: expand_same_element is over-conservative for mapped dimensions (full cross-product instead of the positional-mapping diagonal) #527 -- mapping-aware same-element element edges (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: mapped-dimension sliced reducers are not hoisted into aggregate nodes (compute_read_slice declines a remapped iterated axis) #534 -- mapped-dimension sliced reducers hoist (positional mappings) (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)
• ltm: exhaustive-mode loops never get runtime polarity reclassification (Rux/Bux unreachable) #679 -- exhaustive/pinned loops get runtime polarity reclassification (PR ltm: core-algorithm batch of root-caused fixes from epic #488 #761)

PR #740 (reviewer-finding batch, 2026-06)

• ltm: LTM_BENCH_TRACE should be cfg(any(test, feature = "ltm_bench")) instead of runtime env check #464 -- LTM_BENCH_TRACE build hygiene (PR ltm: batch of root-caused fixes from epic #488 #740)
• LTM post-sim: preserve emission order for rel_loop_score denominator summation #468 -- loop-partition ordering preserves emission order (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: document flow-to-stock 1-DT-shift convention in user-facing reference #505 -- flow-to-stock 1-DT-shift convention documented (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: clarify or eliminate the unreachable MostlyBalancing tie-break in LoopPolarity::from_runtime_scores #506 -- unreachable MostlyBalancing tie-break clarified (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: model_element_causal_edges' both-scalar fast path bypasses ThroughAgg routing (scalar feeder of a hoisted reducer with a scalar target loses the agg hop) #533 -- both-scalar fast path consults ThroughAgg routing (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: graphical-function monotonicity compares the y-delta, not the slope dy/dx (non-uniform x-spacing) #536 -- GF polarity epsilon compares slope, not dy (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: analyze_model swallows compile errors and returns an empty ModelAnalysis silently #660 -- analyze_model surfaces compile failures instead of empty result (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm-augment: substitute_reducers_in_equation parse-failure fallback returns input unchanged #661 -- substitute_reducers parse-failure fallback made loud (PR ltm: batch of root-caused fixes from epic #488 #740)
• simlin-mcp-core: LTM diagnostics never surface (collect_all_diagnostics runs with ltm_enabled=false) #662 -- MCP diagnostics honor LTM enablement (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: non-Euler (#486) guard over-rejects models whose stocks are in no feedback loop #663 -- non-Euler guard gates on flow-to-stock emission, not stock presence (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm-finding: multiply_score_series/max_abs_score_series silently truncate to the shorter operand #678 -- score-series helpers no longer truncate to the shorter operand (PR ltm: batch of root-caused fixes from epic #488 #740)
• ltm: multi-output-port module pathway/composite ordering is process-nondeterministic (unordered output-port set) #680 -- multi-output-port module pathway ordering made deterministic (PR ltm: batch of root-caused fixes from epic #488 #740)

Triage 2026-06-10

• LTM needs to support arrays #273 -- arrays umbrella, superseded by this epic's cluster structure (triage 2026-06-10)
• LTM: improve/generalize module support #274 -- modules umbrella, superseded by this epic's cluster structure (triage 2026-06-10)
• ltm: pin on a genuinely stateless model bypasses validation -- typo'd pin gets no invalid-pin warning #770 -- stateless typo'd-pin silence, consolidated into ltm: model_detected_loops has no stateless gate -- stateless-but-cyclic shapes report unscoreable loops and silently drop pins #772 (triage 2026-06-10)

Earlier work (2026-04/05 triage + LTM arrays hardening + review batches)

• ltm: extend composite module scoring beyond implicit stdlib modules #306 -- composite scoring extended to user-defined dynamic modules
• ltm: black-box module link scoring collapses magnitude to 1 #307 -- black-box module link score uses true delta-ratio
• ltm: dedup by node set merges distinct directed loops #308 -- loop dedup keys on canonical edge-sequence rotation
• ltm-finding: truncate-before-filter can drop partition-dominant loops #310 -- rank_and_filter truncation reordered after partition-aware filtering
• ltm-augment: parse fallback can silently break ceteris-paribus logic #311 -- partial-equation parse-failure fallback made loud
• ltm: fixed max depth 20 truncates module pathway enumeration #312 -- module-pathway hard-coded depth-20 cap
• engine: CrossElement over-approximation for fixed-index subscript refs in LTM causal graph #448 -- FixedIndex NxN over-approximation replaced by per-reference RefShape
• ltm: JohnsonState per-SCC allocation sized to whole-graph node count #460 -- JohnsonState allocation sized to SCC, not whole graph
• engine: gate wrld3_element_level_enumeration_is_uncapped behind cfg(not(debug_assertions)) #465 -- WRLD3 uncapped-enumeration test gated #[ignore]
• LTM auto-flip warning unreachable through simlin_project_get_errors #466 -- LTM auto-flip warning reachable via simlin_project_get_errors
• LTM layout: add integration/unit coverage for detected.truncated fallback branch #467 -- detected.truncated fallback branch removed
• ltm: polarity analysis returns Unknown for all array reducers #480 -- polarity analysis for array reducers
• ltm: enumerate variable-level loops first, expand only cross-element subgraph #482 -- variable-level loop enumeration before element expansion
• ltm-augment: STDDEV/RANK partials silently fall back to delta-ratio #483 -- STDDEV analytic ceteris-paribus partial
• ltm: surface polarity confidence ratio for Rux/Bux classification #485 -- polarity confidence ratio surfaced
• ltm: emit compile-time error when ltm_enabled is true and integration is non-Euler #486 -- compile-time error for LTM + non-Euler integration
• ltm: rel_loop_score normalisation cross-pollutes independent A2A loops #487 -- per-slot A2A loop partitions (no cross-pollution)
• ltm: GF strict-monotonicity check (EPSILON=1e-10) flags numeric-import noise as Unknown #492 -- GF polarity y-range-relative epsilon
• ltm: surface MostlyReinforcing/MostlyBalancing in JSON/TS/WASM layers #495 -- Rux/Bux + polarity confidence through FFI/TS/pysimlin
• ltm: assign_loop_ids relies on HashMap iteration order for multidigraph sibling cycles #497 -- loop-id stability via canonical edge-sequence tiebreaker
• ltm: support per-element graphical functions in static link polarity analysis #502 -- per-element graphical functions get per-element link polarity
• ltm: cross-element loops should normalize by Δ-aggregate, not diagonal A2A link score #503 -- cross-element loops normalize via synthetic aggregate nodes
• ltm: degenerate link score for disjoint-dimension arrayed-to-arrayed edges with per-element target equations #510 -- disjoint-dimension per-element link scores
• ltm-augment: subscripted-A2A-reference link-score partial fails to compile (PREVIOUS arg must be Var) #511 -- iterated-dims explicit subscript classifies Bare
• ltm: sliced reducer sub-expressions are not hoisted into aggregate nodes (stay on the wildcard-link-score path) #514 -- sliced reducer sub-expressions hoist into read-slice agg nodes
• ltm: MAX_AGG_PETALS=8 caps (and drops) cross-element-through-aggregate loop recovery; per-subset (not per-ordering) enumeration #515 -- cross-agg petal budget with deterministic truncation + Warning
• ltm: cross-element loops through aggregate nodes always classify as Undetermined (agg-hop link polarity is always Unknown) #516 -- agg-hop polarity derivable for monotone reducers
• engine: SUM(PREVIOUS(arr[*])) silently evaluates to 0.0 under an active A2A dimension (no LoadPrev-of-array-view path) #517 -- SUM(PREVIOUS(arr[*])) evaluates correctly under A2A dims
• ltm: unify the element-graph and link-score reference-site walkers behind a single classification IR #520 -- unified reference-site walkers behind ClassifiedSite IR
• ltm-finding: discovery mode (discover_loops) is intractable on World3 (390s / 22GB, no result) #540 -- discovery tractability on World3 (strongest-path DFS blowup)
• engine: bare arrayed name inside nested PREVIOUS() in an apply-to-all equation fails to compile #541 -- nested-PREVIOUS bare arrayed name no longer hard-fails compile
• ltm-post: a single NaN loop_score poisons every rel_loop_score in its cycle partition #542 -- NaN loop score no longer poisons its partition
• ltm-finding: rank_and_filter ranks/truncates loops by raw avg_abs_score instead of partition-relative contribution #543 -- ranking by partition-relative contribution
• ltm: synthetic-fragment compilation uses an empty tables map, so link scores into graphical-function targets stub to 0 #546 -- fragment compiles get the model's real tables map
• ltm: Piece 3 fragment-diagnostics positive tests are coupled to the lifetime of real fragment-compile bugs #547 -- test-only fragment-failure injection hook
• ltm: composite-reference link score into a stdlib-macro module fails to compile in exhaustive mode, so loops through SMOOTH/DELAY stub to 0 #548 -- exhaustive-mode composite-reference fragments compile
• ltm: enumerate_module_pathways has no cap on the number of pathways (exponential synthetic-var blowup) #649 -- MAX_MODULE_PATHWAYS budget with truncation Warning
• ltm: expose cycle-partition id/size on discovered loops so singleton-partition loops can be grouped or filtered #671 -- partition id/size exposed on discovered loops

[bpowers]

FFI/parity surfacing cluster -- PR #732

Tackled the "engine knows but consumers can't see it" cluster: the engine computed rich loop analysis that died at the Rust boundary. PR #732 surfaces it across C-FFI / WASM / TS / pysimlin / MCP:

• ltm: discovery-mode agg_recovery_truncated signal never reaches FFI/pysimlin/TS (asymmetry vs exhaustive) #701 -- agg_recovery_truncated to SimlinDiscoveryResult + pysimlin + MCP
• ltm: surface MostlyReinforcing/MostlyBalancing in JSON/TS/WASM layers #495 -- Rux/Bux variants + polarity_confidence (FFI/WASM/TS/pysimlin/MCP; diagram rendering deferred to diagram: render Rux/Bux loop polarity with confidence-gray treatment (deferred from #495) #731)
• ltm: exhaustive-mode loops never get runtime polarity reclassification (Rux/Bux unreachable) #679 -- reclassify_loops_from_results gets its first production caller via the new sim-bearing simlin_analyze_get_loops_runtime(sim) FFI
• ltm: expose cycle-partition metadata on exhaustive/pinned loop surfaces (SimlinLoop, Model.loops/Run.loops, MCP LoopDominanceSummary), keyed by stock-set identity #685 -- cycle-partition metadata on the exhaustive/pinned surfaces (SimlinLoop.partition, pysimlin Model.loop_partitions, MCP LoopDominanceSummary.partition)

These four close on merge of #732 (Fixes).

Closed as already-fixed by PR #705 (commit 26a641e)

Verified present on main and closed: #696 (discovery cross-agg recovery), #698 (discovery wrong polarity through multi-output modules), #699 (overstated zero-edge-exclusion claim + stale test comment).

New follow-up filed

#731 -- diagram: render Rux/Bux loop polarity with the confidence-gray treatment (deferred from #495; the diagram has no existing polarity consumer, so it's net-new design work, not parity plumbing).

Note on #685 (documented in the PR): the exhaustive surface partitions at variable granularity (pop), discovery at element granularity (pop[nyc]), so the partition stock-set is a reliable cross-surface key only for scalar models -- pinned by a test.