Fold 5 StringId accessors into get<Tag> with role tags

[typeless]

Summary

Extend get<Tag> to cover StringId-returning scalar properties by introducing a get_storage<Tag>::type trait (mirroring view<Tag>'s design). Add 4 role tags (Name, Display, SourceDir, InstructionPattern), specialize them to return StringId, and delete the 5 named accessors they replace.

Deleted:

• get_name, get_display_str, get_source_dir, get_instruction_pattern (all returned string_view via pool lookup)
• get_display_id (returned raw StringId)

Principle: view<Tag> is for arrays of semantic values; string-as-char-range doesn't fit. But storage-honest StringId access fits get<Tag> cleanly — the tag disambiguates by role, and callers that want string_view add an explicit global_pool().get(...).

Mechanism

get<T> was previously "T = return type" dispatch (worked because the 4 tags were unique scalar types). This PR generalizes to trait-based dispatch:

template <typename Tag> struct get_storage;
template <> struct get_storage<NodeType>    { using type = NodeType; };      // identity
template <> struct get_storage<Name>        { using type = StringId; };     // role tag
// ... 6 more specializations

template <typename Tag>
auto get(Graph const&, NodeId) -> typename get_storage<Tag>::type;

Existing 4 specializations (NodeType, NodeFlags, Hash256, OutputAction) are unchanged at call sites — their trait entries map the tag to itself. New 4 specializations use role tags mapping to StringId.

Free win: round-trip elimination

Four call sites in cmd_build.cpp were doing pool.intern(pool.get(stored_id)):

// before:
.name = pool.intern(graph::get_name(g, id)),
// after:
.name = graph::get<graph::Name>(g, id),

That's removing a pool lookup + re-intern per file-node and per command during serialization.

Size impact

• putup binary text: 435,278 → 435,014 bytes (-264 bytes)
• The round-trip elimination in serialization loops actually removes real code

Public API

Before: 39 named functions + 2 templates (4 + 3 specializations)
After: 34 named functions + 2 templates (8 + 3 specializations)

Original starting point was 57 named functions. This is -5 named, +4 specializations net.

Test plan

• make test — 23592 assertions pass
• make format clean
• make iwyu — no dead includes
• Binary size verified smaller (-264 bytes)
• CI: format-check, build (linux/macos/windows), test, tidy, build-gcc-bsp

🤖 Generated with Claude Code