engine: residual ~14 C-LEARN unit diagnostics need a Vensim-parity strictness decision (IF-branch / log-argument / coefficient cancellation)

[bpowers]

Description

After four root-cause fixes on branch fix-clearn-unit-errors reduced the C-LEARN unit-diagnostic flood from 481 spurious diagnostics to ~14 (fixes: Context self-alias registration, units_infer time-unit alias resolution, units_check tolerating unknown dependency units, and skipping macro-marked models during unit checking), C-LEARN still produces ~14 unit warnings that genuine Vensim does not flag as mismatches. These break down as:

• 6-7 computed units 'gtonsc/degreesc/yr' don't match specified units on permafrost/clathrate methane-flux variables (flux_c_from_permafrost_release, ch4_emissions_from_permafrost_and_clathrate) -- a temperature-sensitivity-coefficient unit that does not fully cancel.
• 6 expected left and right argument units to match, but 'dmnl' and '1/ppm' don't on the ph variable (a log/concentration formula).
• 1 IF branches have different units: then '1/gtonsco2', else '1/yr' on global_annual_rate_of_co2eq_2050_to_2100.
• 1 umbrella main-model unit inference [...]: unit_mismatch warning (which underlying conflict it reports is nondeterministic -- see engine: wrld3_resource_unit_alias_should_not_conflict is order-dependent / flaky #474 for the order-dependence root cause).

The open question

It is not yet established whether these are:

1. Simlin being legitimately stricter than Vensim -- e.g. Vensim may not check IF-branch unit consistency, or may not require the argument of a LOG/concentration formula to be dimensionless. If so, we may want to relax those specific checks for Vensim parity (an IF-branch-units check and a log-argument-dimensionlessness check are the two candidate relaxations).
2. Genuine dimensional inconsistencies in the C-LEARN model that Vensim happens to tolerate silently. If so, the diagnostics are correct and the resolution is to document them as known model-authoring imprecision (not an engine change).

Resolving this requires per-variable investigation of the C-LEARN equations against Vensim's actual unit-checking semantics for each of the three distinct message classes (coefficient non-cancellation, log-argument dimensionlessness, IF-branch consistency).

Why it matters

• Unit checking is meant to be a verification safety net; residual noise on a flagship real-world model (C-LEARN) erodes trust in it and can mask genuine errors.
• If Simlin is over-strict relative to Vensim, every imported Vensim model with these idioms inherits spurious diagnostics.
• If these are genuine model errors, surfacing them is correct behavior and the work is documentation, not code.

Components affected

• src/simlin-engine/src/units_check.rs -- the IF-branch consistency check and the log/binary-operator argument-units checks
• src/simlin-engine/src/units_infer.rs -- coefficient unit cancellation during inference

Reproduce

cargo test -p simlin-engine --test clearn_unit_errors -- --ignored --nocapture dump_clearn_unit_diagnostics

Context

Discovered while fixing the C-LEARN unit-error flood on branch fix-clearn-unit-errors (481 -> 14 diagnostics via four root-cause fixes). These ~14 are explicitly out of scope for that branch -- they require a Vensim-parity semantics decision per message class, not a mechanical fix. Distinct from #454 (resource_unit/resource_units alias), #455 (year-vs-implicit-time stock unification), #456 (duplicate_unit per stdlib instance), #457 (stdlib module output units not propagated), and #474 (the order-dependence that makes the umbrella diagnostic nondeterministic). Related: #35 (general unit-checking tracker).