engine: two offset machineries disagree on submodel size with a table-only variable (latent off-by-N)

[bpowers]

Problem

Two independent offset machineries size a module-typed body variable differently when the submodel contains a table-only (lookup-only) variable:

• calc_flattened_offsets_incremental (src/simlin-engine/src/db/assemble.rs) sizes a module-typed body variable by the sum of its sub_offsets entry sizes, which excludes table-only / lookup-only variables (they are skipped from the offsets map).
• compute_layout (src/simlin-engine/src/db/layout.rs) sizes the same submodule by the full sub_layout.n_slots, which does count table-only slots.

If a submodel contains a table-only variable (the standalone lookup-only kind from #606) and is instantiated as a module, the two machineries can disagree on the submodule's size, corrupting result-map slot offsets (an off-by-N in the result map).

Why it matters

This is a latent correctness bug: result-map slot offsets could silently point at the wrong slots for any model that instantiates such a submodel as a module. It produces wrong numbers, not a crash.

Current status

• Not tripped by any in-corpus model today -- no existing fixture has a submodel containing a lookup-only variable instantiated as a module.
• Not introduced by the is_root refactor -- this is pre-existing; the salsa-pipeline-cleanup refactor surfaced it.
• Related to engine: standalone lookup-only variable is lowered to gf(Time) (unit error; phantom Time-indexed series for a bare graphical function) #606 (standalone lookup-only variable handling) but distinct: engine: standalone lookup-only variable is lowered to gf(Time) (unit error; phantom Time-indexed series for a bare graphical function) #606 is about the gf(Time) lowering being a unit error; this is about the two offset machineries disagreeing on submodule size when such a variable is present.

Components affected

• src/simlin-engine/src/db/assemble.rs (calc_flattened_offsets_incremental)
• src/simlin-engine/src/db/layout.rs (compute_layout)

Possible approach

Reconcile the two size computations so both either include or exclude table-only slots consistently. Add a regression test: a submodel containing a lookup-only variable, instantiated as a module, asserting the results-map offsets match the assembled layout (i.e. the two machineries agree).

Severity

medium / latent correctness bug.

Context

Identified during the salsa-pipeline-cleanup refactor.