Struct / Encoding coverage audit

Generated 2026-05-23 — the Phase 5 artifact of the post-c9a535a rigor-type-coverage-uplift line. Phase 5 was scoped as “the remaining low-priority carriers”. The audit’s conclusion for both types is a decision, not a dispatch slice: one is deferred as an ADR-worthy feature, the other is recommended for permanent exclusion.

---

1. Struct — deferred (needs a struct-shape carrier; ADR-worthy)

Current state

Struct.new(*members) returns a fresh anonymous subclass. STRUCT_CATALOG is defensive (see struct_catalog.rb): it recognises Struct as a receiver class, classifies Struct.new :block_dependent (the optional class-body block), and blocklists :[] / :hash / :initialize_copy against a “hypothetical future Constant<Struct> carrier”.

spec/integration/fixtures/struct_catalog.rb fixes the status quo:

klass = Struct.new(:foo, :bar)   # => Nominal[Struct]
inst  = Struct.new(:foo).new(1)  # => Dynamic[top]

inst.foo cannot fold because nothing models the per-subclass member layout.

Why this is not a coverage-uplift slice

Precise Struct member access (Point = Struct.new(:x, :y); p = Point.new(1, 2); p.x # Constant[1]) needs two new carriers, not a dispatch-tier entry:

1. A struct-class carrier — an anonymous Struct subclass parameterised by its ordered member-name list (and keyword_init: flag).
2. A struct-instance carrier — member-name → value-type, essentially a HashShape with a class tag, so .x / .x= / [] / to_h / deconstruct project precisely.

That is a genuine feature with real design surface, none of which the coverage-uplift corpus has decided:

• the Struct.new class-body block (do … end) may define extra methods — the carrier must degrade gracefully when it does;
• keyword_init: true structs, positional structs, and the Ruby 3.2 hybrid;
• subclass-of-a-struct (class Point3D < Point);
• the immutable sibling Data.define (Ruby 3.2+) — arguably a better first target than Struct because Data instances are frozen, so the carrier soundness story is simpler.

Recommendation

Treat Struct / Data value folding as a dedicated feature with its own ADR (carrier zoo addition + the Struct.new / Data.define block-degradation contract), not an autonomous coverage slice. The HashShape carrier is the obvious structural prior art. No implementation is authorised by this audit.

---

2. Encoding — recommended for permanent exclusion

Current state

ENCODING_CATALOG is defensive in the same shape as STRUCT_CATALOG: Encoding is a recognised receiver, the RBS tier answers Nominal[Encoding], and the registry-walking singletons (Encoding.find / .list / .aliases / .name_list) plus the mutating setters (default_external= / default_internal=) are blocklisted as process-registry-dependent.

Why a carrier would not pay for itself

A Constant[Encoding] carrier could only fold a vanishingly small surface:

• Encoding::UTF_8.name → Constant["UTF-8"], .dummy? → Constant[false], .ascii_compatible? → Constant[true].

Real programs do not perform static computation on Encoding objects — the type is used as an opaque tag passed to String#encode / #force_encoding. The blocklisted registry methods (the only ones with interesting return values) stay blocklisted because their answer depends on the analyzer process. The precision gain over the existing Nominal[Encoding] answer is negligible, and every new SCALAR_CLASSES member widens the carrier zoo (Ractor shareability, describe, erase_to_rbs, the == / hash contract) for no user benefit.

Recommendation

Do not add an Encoding constant carrier. The current Nominal[Encoding] answer is correct and sufficient; ENCODING_CATALOG’s defensive blocklist already covers the unsound paths. This is a deliberate permanent exclusion, recorded so the question is not re-litigated.

---

3. Phase 5 outcome

型	判断	理由
Struct / Data	延期（ADR 化推奨）	struct-shape キャリア 2 種が前提。実機能であり coverage スライスではない。
Encoding	恒久的に対象外	キャリア新設のコストに見合う精度向上がない。Nominal[Encoding] で十分。

Phases 1–4 of the coverage-uplift line landed as dispatch / carrier slices; Phase 5 closes the line by recording that the two remaining candidates are a deferred feature and a permanent exclusion respectively.