engine: per-source fan-out of unit inference diagnostics (enables per-field underlines deferred from #836) #837

Problem

Unit inference conflicts are surfaced only as a single model-level umbrella diagnostic with variable: None, so the diagram cannot attribute an inference error to any variable, let alone underline the offending span. While fixing #836 we chose "Option A": inference errors get no per-field underline at all, because with only the umbrella diagnostic there is nothing to attribute. Two engine-side follow-ups were identified and deliberately deferred; they are the same work area, so they are tracked together here.

1. Per-source fan-out of unit inference diagnostics

UnitError::InferenceError.sources (src/simlin-engine/src/common.rs:2449) already carries a Vec<(variable_name, Option<Loc>)> -- one loc per involved variable's equation. But the salsa pipeline (src/simlin-engine/src/db/units.rs, the has_declared_units && !inference.conflicts.is_empty() block around line 279) collapses all conflicts into one model-level diagnostic with variable: None. No per-variable diagnostic is ever emitted, so the diagram has nothing to place a per-field underline on.

Emitting one properly-located per-source diagnostic per involved variable would enable correct per-field underlines (the deferred "Option D" of #836).

Constraint: the umbrella exists deliberately -- macro-heavy models can produce hundreds of internal constraint contradictions, and one-diagnostic-per-conflict would flood the report (see the comment in that block). A per-source fan-out needs dedup/capping thought: e.g. dedupe by variable, cap the count, or emit per-variable diagnostics alongside (not instead of) the umbrella. The full conflict list already remains available on InferenceResult for callers that want it.

2. format_unit_error picks the wrong source's Loc

The InferenceError arm of format_unit_error (src/simlin-engine/src/errors.rs:216) sets start/end offsets from sources.first()'s Loc:

let (start, end) = sources
    .first()
    .and_then(|(_, loc)| *loc)
    .map(|loc| (loc.start, loc.end))
    .unwrap_or((0, 0));

But the first source may be a different variable than the var_name the error is being attributed to, so the offsets can point into another variable's equation. If per-variable inference diagnostics are emitted (item 1), the formatter should select the loc of the source matching var_name instead of first().

Why it matters

Diagnostic correctness and user experience: today the per-variable inference path is latent (nothing emits per-variable inference diagnostics), so this is low urgency. But it blocks ever giving inference errors a correct per-field underline in the diagram, and item 2 is a landmine that would produce wrong-equation offsets the moment item 1 lands.

Components affected

• src/simlin-engine/src/db/units.rs (umbrella diagnostic emission)
• src/simlin-engine/src/errors.rs (format_unit_error InferenceError arm)
• src/simlin-engine/src/common.rs (UnitError::InferenceError.sources -- already carries the needed data)
• Downstream: src/diagram per-field underline placement (consumer, once per-variable diagnostics exist)

Context

Identified during the fix for #836 (July 2026), where the diagram-side placement bug was resolved by giving inference errors no per-field underline ("Option A"). Related: #614 (partial inference results, closed -- this builds on that work), #613 (residual C-LEARN unit-diagnostic strictness).