Mir dedupe

[sbillig]

• Structural layout dedup
  Drop source_ty from struct/array/enum layouts and name from enum variants, so structurally-identical layouts collapse to a single LayoutId. Variant selection is by index, so the dropped names were cosmetic. This is the bulk of the snapshot churn (layouts/variants now render as struct/N, variant_N, etc.).
• Canonical resolved-trait-method instances
  When a trait method resolves to a concrete impl, build a canonical ImplEnv (impl scope, single resolving witness, no caller assumptions) and cap method args at the impl's generic-param count, so the same resolved method dedupes across callers regardless of call-site context.

Also includes a small refactor (shared ImplEnv::for_resolved_trait_method constructor, simplified const_lowering_ty) and a regression fixture confirming trait default bodies reached through a generic param receiver still resolve under the canonical env.

Practical impact is faster compile times for large projects:
(fe test total runtime)
bountiful: 13.38s -> 4.61s
zk-kit: 58.08s -> 45.40s
chz: 17.37s -> 16.05s

[sbillig]

Will merge this after 26.2. I think this will also require some inliner tweaks in sonatina to avoid gas regression.

[micahscopes]

agent:

Tiny one: const_ref.rs stores the raw call-site callable.trait_inst() as the
canonical witness while classify.rs uses the normalized concrete_inst, so the
two paths can produce spelling-dependent key splits — missed dedupe only, not a
correctness issue.

[sbillig]

agent:

Tiny one: const_ref.rs stores the raw call-site callable.trait_inst() as the
canonical witness while classify.rs uses the normalized concrete_inst, so the
two paths can produce spelling-dependent key splits — missed dedupe only, not a
correctness issue.

I think this is fine, and changing it would maybe be less correct.

[sbillig]

Will merge this after 26.2. I think this will also require some inliner tweaks in sonatina to avoid gas regression.

No inliner tweaks required. This branch provides a small gas benefit on the rosetta suite relative to master, with no regressions.

  master:      6,805,266
  mir-dedupe:  6,794,563
  delta:         -10,703 (-0.16%)