DESIGN — Native reflection bridge (Phase 4 operational)

Status (2026-07-24): Phases 1, 2, and PHASE-3 INCREMENT 1 (the write-back half) are SHIPPED. Increment 1 landed via the §6.1 chip (session optimistic-bassi-496625; design + adversarial-review ledger in DESIGN-reflection-bridge-phase3-plan.md), validating errors (#34, 61/61 vs go test -json) as its demonstrated consumer. Companion to the Phase-4 operational campaign in ../Roadmap.md.

⚠ “Phase 1/2/3” below are THIS DOCUMENT’s scope phases (§ Scope), unrelated to the project’s Phase-3 / Phase-4 milestones.

Implemented — manualConversionFuncs in src/go2cs/manualTypeOperations.go is the authoritative list; derive it fresh rather than trusting this summary: the Kind classifier + type helpers (golib GoReflect: KindOf, GoTypeName, ElementType, IsComparable, TryAdapterWrappedType), internal/abi.TypeOf (type_impl.cs synthesizes the descriptor from the managed System.Type), reflect.{ValueOf, unpackEface, valueInterface}, the ~17 Value readers + MapIter.{Next,Key,Value}

Phase-3 increment 1 (SHIPPED 2026-07-24, the chip): the ONE new primitive — an addressable Value carries the ж<T> box it aliases (addrBox companion; reads go through the box lazily, Set writes through its slot ref via cached compiled accessors over ValueSlot) — plus reflect.{Value.Set, Zero, methodName} and the reflectlite errors.As surface (Value.{Elem,IsNil,Set}, rtype.{Elem,Implements,AssignableTo}, methodName), all over shared golib machinery (GoReflect.{TryMarshalAssignable, GoImplements, ReadPointerSlot/WritePointerSlot, CanonicalNilPointer, GoDynamicTypeOf}) so reflection and emitted _<T> asserts can never disagree about a method set. With it landed: canonical typed-nil pointer boxing (X2 — converter + golib ж<T>.NilBox + gen; see ConversionStrategies-Reference Canonical typed-nil pointer boxing), structural nil-pointer identity (INilPointer; heap-box-holding-nil ≠ nil pointer), the R10 adapter unwrap (GoDynamicTypeOf — descriptors classify the Go dynamic type, so adapter-held and raw-box values intern to ONE canonical Type), the golib assert-fallback unwrap + ж-overload closing (X1) with the non-generic TryTypeAssert(object, Type, out object), dyn-wrapper IInterfaceAdapter transparency + the [GoRecv] method-set fix (X3), Go-parity empty-name subtest numbering (X4), the oracle’s two-phase address-token pairing (X5), and %v-of- method-bearing-pointer fidelity (handleMethods wins, never &-prefixed). The getcallersp NotImplementedException chain is severed at methodName — the semantic boundary where a managed answer exists.

Phase-3 increment 3 (SHIPPED 2026-07-26) — the type-relation mirrors + conversion. The deferred “reflect-side Implements/AssignableTo mirrors” row landed with its demonstrated consumers (encoding/gob’s init via validUserType → implementsInterface, go/token’s TestSerialization, internal/fmtsort’s ct() table): hand-owned rtype.{Implements, AssignableTo, FieldByName}, PointerTo, Value.{Convert, Cap, SetLen} — each severing a descriptor-SPECIALIZATION read (Reinterpret<abi.Type, interfaceType/structType>, the ptrType prototype, the cvt/unsafe_New chain) that cannot exist behind a synthesized descriptor — plus the StructField.Index stamp (gob’s FieldByIndex walk), golib pointer order tokens (INilPointer.PointerOrderToken: equal pointers token equally, same-storage elements order by index — fmtsort’s map-key ordering), runtime.Pinner.Pin/Unpin as no-ops by construction, and -tests init-order relocation (the erasable initᴛᴛtests static-ctor hook; see ConversionStrategies-Reference *Package-Level Variable Initialization Order). Validated: internal/fmtsort 3/3, go/token 31/31; gob runs its Encoder/Decoder engines end-to-end (full gob validation remains open).

Phase-3 increment 3a (SHIPPED 2026-07-26) — Value.Addr, rtype.PkgPath, and the EMPTY interface. Found by running encoding/gob’s own 98-Test suite (68 → 79 passing). Three roots, each general:

✅ gob’s build blocker is CLOSED (2026-08-02, r37-gob) and gob is MEASURED — the “79 of 98” above is superseded by 86 of 106. It was build-blocked when increment 5 looked: package_info_internal_test.cs emitted [assembly: GoImplement<gob_internal_test_package.Point, Pythagoras>], but Pythagoras is declared in the EXTERNAL test package (example_interface_test.go, package gob_test) and gob declares a SECOND, unrelated Point there — so the record anchored the internal variant’s same-named type and left the interface unqualified in a file where it is not in scope: CS0246, no test host, all 106 verdicts empty. That was test-project-model record anchoring (the splitWhiteboxVariantRecords family), not reflection surface: a BARE record name resolves in the variant that RECORDED it, and the bridge’s declared-name set was being consulted across variants. Fixed and guarded — see ConversionStrategies-Reference.md, A BARE record name resolves in the variant that RECORDED it.

Measured, one run, zero empty verdicts: 86 of 106 match (C# 81 pass + 5 skip vs Go 101 pass + 5 skip; 19 capability-excluded, 0 disclosed). Full per-root census — seven roots, only one of them this arc’s descriptor surface — is the encoding/gob section of BOARD-next-validation-candidates.md.

Rooted gob residues owned by THIS arc (open, NOT disclosure candidates) — 3 of the 20. The managed array<T> type does not carry its LENGTH, so reflect.Type.Elem() of a *[N]T loses N and gob sees the wrong type where the wire says [7]int — TestSingletons and TestIndirectSliceMapArray (Value.Elem/abi.TypeOf recover dims from the LIVE value, but a type-only walk cannot). TestIgnoreDepthLimit is reflect.ArrayOf → the typelinks stub. A fourth is adjacent and worth this arc’s attention even though it is not the descriptor surface: the five remaining GobEncoder tests share ONE root that is not reflection at all — their GobDecode bodies write through a reinterpreted named-type pointer, fmt.Sscanf(s, "VALUE=%s", (*string)(v)), and the reinterpret is emitted as a value conversion into a fresh box (Ꮡ((@string)(v))) because reinterpretManagedEmission is gated on a deref context or a raw-address source, neither of which a pointer conversion used as an ARGUMENT satisfies. The direct-field-write case (ByteStruct, TestGobEncoderStructSingleton) passes, so Addr’s write-back path is sound. TestNetIP was never blocked on net’s package init (that claim is retracted — fd_windows is not on its stack): it dies in unique’s initializer, where r37-gob fixed a dead deref alias in internal/concurrent and thereby exposed the root behind it — abi.TypeOf(m).MapType() over a zero map yields a descriptor with no Hasher, so NewHashTrieMap throws Delegate to an instance method cannot have null 'this'. That one IS this arc’s surface. The rest are gob wire/typed-nil behaviours, all bucketed on the board.

Phase-3 increment 4 (SHIPPED 2026-07-31) — the getcallersp chain: runtime.Callers + Frames.Next managed. The chip’s io increment. The failing surface was never the flatten OPTIMIZATION (pure Go, converts fine) but the tests’ MEASUREMENT of it: runtime.Callers → callers() opens with getcallersp() (assembly) and Frames.Next reads linker funcInfo tables. Both contracts are answered natively in runtime/managed_impl.cs over System.Diagnostics.StackTrace with a GO-LOGICAL frame projection — only source-declared Go functions count; adapter shells (IGoAdapter) and go2cs-gen forwarders ([GeneratedCode]) are invisible, exactly as Go’s interface dispatch adds no frame — so relative depths match Go’s logical model (io’s readDepth == myDepth+2 holds bit-exactly); PCs are opaque interned process-lifetime tokens. getcallersp itself stays an honest stub — the chain is severed at the API boundary where a managed answer exists, the methodName rule. (⚠ Follow-up 2026-08-07, r43g-caller: this increment hand-owned the EXPORTED Callers only, and runtime.Caller calls the lower-case callers — so Caller itself stayed dead, which is what kept log and testing/slogtest on the stub. methodName was unaffected, being hand-owned in its own right. The lower-case funnel is hand-owned too now, one entry, and Caller works while staying auto-converted.) io: 45/54 → 47/54; the remaining seven verdicts keep their non-reflection owners (os arc, StringCheckCall, two alloc-profile disclosure rulings). Full design: ConversionStrategies-Reference runtime.Callers / Frames.Next walk the managed stack projected to GO-LOGICAL frames. (Same landing repairs the whitebox adapter-pair resolver — a bare cast of a variant-local type resolved to the first same-simple-name FOREIGN record (bytes_BufferжReader for io_test’s own Buffer, CS1503 ×20), which had silently re-walled the io build the board records as closed; anchor-local records now win, guarded by TestBareCastPrefersAnchorLocalRecordOverForeignSimpleNameMatch.)

Phase-3 increment 5 (SHIPPED 2026-08-02) — reflectlite’s rtype.String. The chip’s context increment, and the quietest descriptor read yet: String() is t.nameOff(t.Str).Name(), a name OFFSET into the linker-built name blob that a synthesized descriptor never populates — so every reflectlite.TypeOf(x).String() in the corpus answered "", and answered it without faulting, because "" is a legal name for an unnamed type. context’s stringify fallback printed WithValue(, c1k1) where Go prints WithValue(context_test.key1, c1k1). Bridged in internal/reflectlite/type_impl.cs over GoReflect.GoTypeName — the same answer reflect’s long-hand-owned rtype.String and %T give, so the full bridge and the mini bridge cannot disagree about a type’s name — with array dims threaded as on the reflect side. Blast radius is three call sites (the fix plus two panic messages); rtype.Name is untouched and still "" (it gates on the TFlagNamed bit synthesizeDescriptor never sets), and errors never reaches String() at all. context: 36/38 → 37/38, leaving only the measured TestAllocs alloc-count disclosure its banking commit owns. Full design: ConversionStrategies-Reference The type NAME is a descriptor read too. Recorded next gap of the same shape: rtype.Name — deliberately not fixed without a consumer that demonstrates it.

Phase-3 increment 6 (2026-08-02) — rtype.NumMethod, the method-set SIZE. The chip’s time increment, and the same silent-degradation class as increment 5: NumMethod counts uncommon() method tables a synthesized descriptor never populates, so it answered 0 for every concrete type — silently, because 0 is most types’ correct count. The consequence hid one hop away: encoding/json’s indirect() gates its Unmarshaler/TextUnmarshaler discovery on v.Type().NumMethod() > 0, so no custom UnmarshalJSON/UnmarshalText was EVER dispatched — every json.Unmarshal into time.Time fell through to the raw-struct path (TestTimeJSON), and {} decoded silently where Time.UnmarshalJSON rejects it (TestUnmarshalInvalidTimes). ONE root, no further layers: with the gate answering, the increment-1/3a machinery (field-alias Addr, Interface, the golib assert, write-back through the aliased box) carries the whole dispatch, struct fields included. Hand-owned in reflect/value_impl.cs over golib GoReflect.GoMethodCount → TypeExtensions.GoMethodSetCount, counted over GetGoMethodSetCandidates — the SAME candidate source StructurallyImplements and the shell binder resolve through, so the gate and the assert behind it cannot disagree about a method set; deduplicated by projected Go name (one pointer-receiver method has two emitted shapes), exported-only for concrete types, ALL methods for interfaces, 0 for the empty interface, adapter shells unwrapped per R10. Guard: JsonUnmarshalerDispatch (four dispatch shapes vs go run). Full design: ConversionStrategies-Reference The method COUNT is a descriptor read too. Recorded next gap of the same shape: rtype.Method(i) — still auto over the same absent tables, and a NumMethod() > 0 gate now lets method-enumeration loops get further than before; the first consumer that walks one demonstrates it.

Phase-3 increment 6 + 7 (SHIPPED 2026-08-03 as ONE increment) — the METHOD TABLE: NumMethod, Method(i), MethodByName, and the method VALUE. Increment 6 shipped the COUNT alone (rtype.NumMethod, gate-clean, time’s TestTimeJSON/TestUnmarshalInvalidTimes as its demonstrated consumers) and was reverted from master: the all-package sweep caught math/rand and math/rand/v2’s TestRegress failing with panic: reflect: Method index out of range — the successor gap increment 6’s own report had recorded, arriving one session later. The durable lesson, worth generalizing beyond this bridge: a truthful count is a PROMISE that the table behind it can be indexed. NumMethod reads uncommon() method tables a synthesized descriptor never populates, so it answered 0 for every concrete type; while it did, every method-ENUMERATION loop was unreachable and the auto Method(i) reading the same absent tables could not be observed. Making the count truthful is exactly what made them reachable. A descriptor read and the gate in front of it are one atomic increment.

Landed together: rtype.{NumMethod, Method, MethodByName} and Value.Method, all over ONE ordered list (TypeExtensions.GetGoMethodSetEntries) whose .Count IS NumMethod — a size and an order can no longer be derived separately and disagree — built on the same GetGoMethodSetCandidates the duck-typing assert and shell binder resolve through, deduplicated by projected Go name (keeping the delegate-bindable shape), exported-only for concrete types, and sorted ordinally by Go method name (Go’s own table order; a promoted embed sorts in place). A method value is an ordinary BOUND DELEGATE, which is the design’s economy: binding the receiver at Method(i) time makes the result a Kind-Func Value, so mv.Type(), NumIn/In/NumOut/Out and mv.Call(args) are all existing bridge surface unchanged — the receiver is already gone from the signature, exactly Go’s method-value contract — and Value.MethodByName needs no hand-own because it composes the other two. Binding is expression-compiled because Delegate.CreateDelegate cannot close over a value-type first argument (measured), which every Go value receiver is; one compile per MethodInfo, a closure per bind.

Found on the way and fixed in the table builder: a this object extension method is golib plumbing, never a Go method. The candidate source’s assignability safety net admits them for EVERY type, so TryCastAsInteger(this object, out ulong) was in every method table — nondeterministically, since a late assembly load re-runs the scan: the same binary reported NumMethod 4 or 6 for the same type depending only on load order. The shipped-then-reverted count was therefore also wrong in a way nothing could observe.

Consumers: math/rand 43/43 and math/rand/v2 36/36 at their exact banked counts, with TestRegress now genuinely walking (*rand.Rand NumMethod: 16 in Go’s order, Intn(1000000000) = 526058514 matching go run; before the pair it read 0 and the test passed VACUOUSLY — zero of its 320 golden comparisons ran). time: 146 → 148 pass of 159, the two increment-6 JSON rows re-landed; its 9 remaining failures are the timer-model item (TestChan ×8) and the TestUnmarshalTextAllocations disclosure ruling, neither this arc’s. Guard: tests/Behavioral/ReflectMethodTableWalk. Full design: ConversionStrategies-Reference …and the count and the WALK are ONE increment.

Phase-3 increment 8 (SHIPPED 2026-08-03) — the ZERO test, and a read that had degraded to a CONSTANT. The chip’s encoding/gob increment. Value.IsZero is three descriptor reads over flat memory (an Equal pointer against the shared zeroVal buffer, a TFlagRegularMemory all-bits-zero scan, and v.ptr == nil for a non-flagIndir value); a synthesized descriptor populates none, and the bridge populates neither v.ptr nor flagIndir, so the Array and Struct arms both fell to v.ptr == nil and answered true for every array and every struct whatever it held. Silent, like every member of this family — true is correct for the zero value of the same type. Measured against go run before the fix: [2]uint8{1,2}, NA{1,2}, inner{N:1}, outer{P:&n} all IsZero=true in C#, false in Go.

Three things landed together because each gates the next. (1) Len unwraps a named string — every other named container answers through the golib interface its wrapper implements, but a type NS string wrapper implements none, so Len fell to its 0 default and IsZero’s String arm (Len() == 0) called every non-empty named string zero. The increment-6 rule in a second form: the arm is a GATE on Len, so the gate and the read behind it are one increment. (2) IsZero becomes Go’s own recursive definition with the memory shortcuts removed — a composite is zero exactly when every element or field is, which is precisely the walk the shortcuts stand in for (Go itself falls back to it for a non-comparable, non-regular-memory type), so it needs only Index/Field/NumField, already answered. (3) Value.Grow reads a *unsafeheader.Slice off the same never-populated v.ptr and therefore nil-deref’d for every caller — reflect.ValueOf(&s).Elem().Grow(1) on []byte prints 4 8 in Go and panicked here; it is now a managed reallocation (golib GoReflect.GrowSlice) written back through the aliased box exactly as SetLen does. Growth within capacity writes nothing (Go reaches growslice only past the capacity, and a spurious write would detach another view sharing the backing store), and the landed capacity is deliberately unpinned — Go’s growslice rounds to a size class, so only len+n is guaranteed. Guard: tests/Behavioral/ReflectZeroAndGrow, byte-identical to go run.

The MapType().Hasher row is rooted and deliberately NOT landed — populating it would be worse than the failure. unique’s one-root wall (15 of 19) and net’s last cctor root is a map descriptor with Hasher/Key/Elem unpopulated. It looks like the same shape as every read above and is not: Hasher(unsafe.Pointer, uintptr) uintptr must hash the value at an address, and the address that call site produces cannot name a managed value. Measured three ways: two boxes holding equal @string values necessarily have different addresses, so an address-derived hash can never make unique.Make("hello") agree with itself; a box whose pointee contains a reference has no pinnable slot and its address moved across a forced GC; and the unsafe.Pointer handed to the delegate retains no link to its source box, its constructor taking a uintptr. Key/elem types are recoverable from the carried System.Type, but landing those alone is a regression: Key.Equal is the comparability SIGNAL (pointer identity, not value equality), so a half-populated descriptor converts a loud construction failure into a map that silently mislays every key — the increment-6 lesson inverted: a descriptor field whose read cannot be honored must not be populated to look truthful. The remedy is one layer down and outside this arc’s declared files: internal/concurrent.HashTrieMap is a managed-referent raw-metal case whose CONTRACT (a concurrent map over comparable K) the CLR answers natively while its MECHANISM (hash bytes at an address) it cannot — a hand-owned _impl.cs on the sync.Mutex precedent. Needs a coordinator ownership ruling before it is written.

The map READ pair landed 2026-08-09 (r57b), with go/ast as its first validator. Value.MapKeys and Value.MapIndex were the two map methods MapRange/MapIter.*/ SetMapIndex left behind, and both still ran their converted bodies: each opens with v.typ().Reinterpret<abi.Type, mapType>() to read the key/element types off the embedded abi.MapType. That is NOT the managed-box aliasing case toRType relies on — a synthesized descriptor is a bare abi.Type with no abi.MapType behind it, and the emitted mapType holds that embed as a promoted REFERENCE box (ᏑʗMapType), so the reinterpret reads the descriptor’s first word as an object. Both then index through mapaccess/hiter, which MapRange had already replaced. MapKeys is now MapRange collected; MapIndex is a golib comma-ok lookup (GoReflect.TryGetMapEntry, routed through IMap<K,V> rather than the non-generic IDictionary so the nil key in golib’s dedicated slot is found like any other). One key/element typing rule now serves the walk and the lookup alike, and a miss returns the invalid zero Value that text/template’s funcs.go already tests with IsValid().

Unnamed struct types render STRUCTURALLY (same arc). reflect.Type.String()/%T of a converter-LIFTED anonymous struct reported the lift’s C# name (ast_internal_test.typeᴛ1) where Go prints struct { X int; y int }, and golib’s EmptyStruct where Go prints struct {}. The [GoType("dyn")] stamp is what distinguishes a lift from an ordinary declared struct, and the arm in GoReflect.TypeNaming builds the text from GoFields — the same projection NumField/Field and the value side read, so a type’s reported NAME and the fields it hands out cannot disagree. Format verified against the toolchain (embedded field contributes its type alone; a tagged field appends the strconv.Quoted tag). This partially answers open question 3 on the READ side.

The two Value-layer gaps asn1 and edwards25519 measured, closed (L7, 2026-08-11). Both are descriptor reads that had degraded to a CONSTANT, and both were rooted against the real code rather than the board’s attribution — which was half right in the first case and pointed one layer off in the second.

NOT implemented — remaining Phase-3 surface: MakeFunc; variadic Call/CallSlice (text/template); SetMapIndex delete-on-invalid (encoding/json); the Go unnamed↔named directlyAssignable refinement beyond identity+wrapper (binary named-slice cases if they surface); FieldByName’s embedded-field depth search (a promoted name currently answers the not-found path); open question 3 (field-name/tag fidelity — [GoTag] is carried but not yet projected). Known limitation (recorded): named func types collapse to their structural Func<>/Action<> under canonType System.Type interning — carried named-func identity is needed if a consumer (gob’s type registry) lands on it.

The problem

Go’s reflect is built on reading an interface’s internal two-word layout through unsafe.Pointer. An any value is an eface = { *abi.Type type; unsafe.Pointer data }; reflect reinterprets the address of the interface as that struct, reads the type word (a pointer to the runtime type descriptor abi.Type), and reads/writes the value through the data word as flat memory at computed field/element offsets.

None of that exists in the managed world. A go2cs any is a System.Object reference — one word, no adjacent type descriptor, no flat-memory value the GC will let you address by offset. Reinterpreting the object reference as {type,data} reads garbage and NREs. Concretely, the first operational hit is:

color.New(FgGreen).Println("…")
  → fmt.Sprint(a…) → fmt.doPrint → reflect.TypeOf(arg).Kind()   // spacing decision
  → internal/abi.TypeOf(any a):  ~(ж<EmptyInterface>)(uintptr)(Ꮡ(a))   // reinterpret → NRE

(Plain fmt.Println sidesteps this: doPrintln never calls reflect.TypeOf; only doPrint — used by Print/Sprint/Sprintf fallbacks — does. That is why the fmt.Println("hi") milestone runs.)

Key insight — the entire unsafe chain enters at THREE constructors

A full read of the reflect/internal/abi/fmt surface (see the surface map below) shows the whole model bottoms out on one primitive — the eface {type,data} reinterpret — reached at exactly three points:

# Function File What it does today
1 internal/abi.TypeOf(any a) internal/abi/type.cs:125 &a → *EmptyInterface → read .Type word
2 reflect.unpackEface(any i) (→ ValueOf) reflect/value.cs:156 &i → *EmptyInterface → read .Type + .Data
3 reflect.toType(*abi.Type) reflect/type.cs:3040 wraps a descriptor pointer as an rtype

Every downstream Type/Value method then reads from those words — Type.Kind() masks abi.Type.Kind_; Value.Int() does ~(ж<int64>)(uintptr)(v.ptr); Value.Field(i) does add(v.ptr, offset); etc.

So the bridge is: replace those three constructors so they carry a System.Type (+ the boxed object for a Value) instead of two raw words. The ~30 downstream methods fmt needs then become ordinary managed-reflection wrappers (obj.GetType(), FieldInfo.GetValue, array indexing, Convert.ToInt64, IDictionary enumeration) with no unsafe anywhere.

Architecture — a native reflect/abi shim over System.Type + System.Object

Hand-own the reflection entry points and the exercised methods (whole-file [module: GoManualConversion], the established pattern — cf. native sync, atomic.Value), so:

The one genuinely new primitive is the Kind classifier — System.Type → reflect.Kind — the root of every method. It reads go2cs’s own representations and attributes:

Go Kind go2cs C# representation detect
Bool / Int* / Uint* / Float* bool,nint,int,long,byte,double,… typeof
Uintptr uintptr (golib struct) typeof(uintptr)
Complex128 System.Numerics.Complex typeof
String @string typeof(@string)
Slice / Array / Map / Chan / Pointer slice<T>/array<T>/map<K,V>/channel<T>/ж<T> open generic typedef
Func GoFunc / delegate IsSubclassOf(Delegate)
Struct [GoType] value struct (no num:) [GoType] + IsValueType
Interface Go interface → C# interface IsInterface
UnsafePointer @unsafe.Pointer (: ж<uintptr>) typeof
named type Celsius float64 [GoType("num:float64")] struct Celsius Kind = underlying; Name/String from the type

The metadata the shim recovers Go type info from is already emitted: [GoType(def)] (struct/ interface marker + num:<kind> for named numerics), [GoTag] (= DescriptionAttribute, the raw Go struct-field tag), [GoRecv] extension methods (the method set), and the golib generic types. C# FieldInfo enumeration returns fields in declared (= Go source) order.

Scope — three phases, each independently useful

Phase 1 — TypeOf().Kind() / .String() (unblocks the color sample + scalar Print/Sprint). doPrint calls reflect.TypeOf(arg).Kind() for every arg; the value itself formats via the fast path or the Stringer/error/Formatter C# interface assertions in handleMethods — which use no reflection at all. So a Type shim exposing Kind(), String(), and Elem() (byte-slice check), plus the Kind classifier, makes fmt.Print/Sprint/Sprintf work for every scalar and every type with a String() method. ~1 shim type, ~4 methods, +1 classifier. This is the minimal color-sample unblock.

Phase 2 — full printValue walk (composites without a String() method format correctly). Add the Value shim (~21 methods actually exercised: Kind, IsValid, CanInterface, Interface, Type, Bool, Int, Uint, Float, Complex, String, IsNil, NumField, Field, Elem, Index, Len, Bytes, CanAddr, UnsafePointer, Pointer), Type.{Elem, Field(i).Name}, a MapIter (Next/Key/Value over IDictionaryEnumerator, for fmtsort), and StructField.{Name,Tag,Type}. ~2 core shim types + MapIter + StructField, ~30 methods, all managed reflection. Makes %v/%+v/%#v of structs, slices, maps, and pointers correct.

Phase 3 — write-back & call (Value.Set*, Value.Call, MakeFunc, addressability) — OPEN; the chip’s scope. Needed by encoding/binary, encoding/gob·json·xml, testing/quick, text/template, and (transitively) math/big; not by fmt. Larger, and best designed against a concrete consumer — which is exactly why the charter (§6.1) spawns this chip only once a package’s differential actually lands on this surface, and requires designing WITH the user + adversarial design review before implementation. Carry the getcallersp stub and the adapter-type Kind/Elem unwrap in the same chip.

Open design questions (for review)

Resolved by what shipped (2026-07-22): Q1 — the native shim was confirmed and built; the descriptor-synthesis alternative was not pursued. Q2 — both Phase 1 and Phase 2 were built. Q4 — the answer was both, deliberately: the entry points and exercised methods are whole-file hand-owned *_impl.cs, while the reusable managed logic (Kind classification, Go type naming, element types, comparability, adapter unwrap) lives in golib GoReflect so reflect, internal/abi, internal/reflectlite, and golib’s own builtin formatting all share one implementation. Q3 is still open and belongs to the Phase-3 chip. Q1–Q4 are kept below as the record of the decision.

  1. Approach — confirm the native shim. Replace the 3 constructors + reimplement the exercised methods over System.Type/System.Object (recommended). The alternative — synthesize faithful abi.Type descriptors from System.Type and keep Go’s converted reflect reading them via unsafe offsets — is judged infeasible (the offset reads themselves don’t work in managed memory; it is strictly more surface than the shim).

  2. Scope to build now — Phase 1 only, or Phase 1 + 2? Phase 1 unblocks the color sample and all scalar/Stringer formatting for the least work; Phase 2 makes composite formatting correct. Phase 3 is deferred regardless.

  3. Field-name / tag fidelity. fmt’s %+v/%#v (and later json) need the exact Go field name + tag. C# field names are the Go names except where escaped (@string) or Δ-collision-renamed; [GoTag] already carries the tag. Options: reverse-map the escapes at the shim, or have the converter emit a per-struct Go-name table (a small attribute) the shim reads. (Phase 2 concern.)

  4. Where the shim lives. Hand-own reflect + internal/abi.TypeOf as whole-file [module: GoManualConversion] (consistent with native sync/atomic.Value), or put the managed logic in a golib GoReflect helper that the converted reflect delegates into (smaller hand-owned footprint, but a converted↔golib seam through the shim). Recommend whole-file hand-own of the entry points, since the value/type model is pervasively unsafe and not worth converting.

Surface map (reference)

abi.Type fields: Size_, Kind_ (the field everything keys off), TFlag, Str/PtrToThis (offset-encoded), PtrBytes/Hash/Equal/GCData (unused by fmt). Kind enum (values 0–26) is defined identically in internal/abi/type.cs:40 and reflect/type.cs:245. fmt calls only Type.{String, Kind, Elem, Field(i).Name} and the ~21 Value methods tabulated in Phase 2. handleMethods (fmt/print.cs:795) formats Stringer/error/GoStringer/Formatter via C# interface assertions — no reflection. Fast-path printArg types (no reflection): bool, int/8/16/32/64, uint/8/16/32/64, uintptr, float32/64, complex64/128, string, []byte, reflect.Value.

Fix — reflect.Type must be canonical (map-key ordering)

Go’s reflect.Type is a canonical interned descriptor: TypeOf(x) == TypeOf(y) exactly when x and y share a dynamic type, so aType == bType is a pointer compare. internal/fmtsort.compare (the map-key ordering used by fmt’s %v) relies on it: if aType != bType { return -1 }.

The bridge minted a fresh wrapper per access — abi.TypeOf allocates a new abi.Type box, and both Value.Type() and toType then build a fresh rtypeжΔType (an IжAdapter compared by box identity via golib AreEqual). So two Types describing the same type never compared equal → compare returned -1 for every pair → the stable sort reversed the keys (map[b:2 a:1] instead of map[a:1 b:2]).

Fix: hand-own Value.Type and toType in reflect/value_impl.cs (registered in go2cs/manualTypeOperations.go manualConversionFuncs["reflect"]) so both route through canonType, which interns the ΔType wrapper in a ConcurrentDictionary<System.Type, ΔType> keyed on the abi.Type.sysType the Phase-1 synthType stamped. Identity-equality then matches Go. The cache is process-lifetime (type descriptors are permanent, like Go’s). Interning by System.Type preserves Go’s named-type distinctness for free: a type Celsius float64 is a distinct [GoType("num:float64")] struct whose System.Type differs from float64, so TypeOf(Celsius) != TypeOf(float64). typeSlow (method-value Types) stays auto — not exercised by fmt.

Fix — the bridge’s per-type marker reads are memoized (2026-07-26)

The bridge recovers Go type identity from managed metadata, so it reads the converter’s per-type markers — [GoType] for a type’s kind and underlying, [GoLocalName] for a lifted local type’s original Go name — from its hottest entry points: KindOf (under every ValueOf, Value.Field, Value.Elem, MakeSlice, rtype.FieldByName, and abi.TypeOf), GoTypeName (under every %T and reflect.Type.String()/Name()), and TryConvertTo/TryUnwrapWrapperValue (under the whole Value.Set/SetMapIndex/Call/Convert marshalling surface). Custom-attribute retrieval materializes fresh attribute instances on every call and none of those callers caches its own result, so the cost was paid per VALUE rather than per type — measured at 165–558 ns and 72–448 bytes per call across those entry points, of which the attribute read was the bulk. All four reads now route through two per-type ConcurrentDictionary memos (goTypeMarkerOf / goLocalNameOf), which is sound for the same reason canonType’s intern above is: type descriptors are permanent, and a loaded type’s own attributes cannot change. Full measurement and the two remaining non-attribute residuals are in DESIGN-iface-shell-caching.md §11.2, which is where the audit that found them lives.

Fix — a type descriptor’s order token is its NAME, not its box identity (2026-08-10)

The section above made two reflect.Types for one Go type compare equal. This one is about what happens next, when they compare unequal: internal/fmtsort.compare’s reflect.Interface arm orders interface-kinded map keys by their dynamic types, and it does that by comparing the two type descriptors as POINTERS —

c := compare(reflect.ValueOf(aVal.Elem().Type()), reflect.ValueOf(bVal.Elem().Type()))

reflect.ValueOf(aType) is a *reflect.rtype, so compare recurses into its reflect.Pointer arm and returns cmp.Compare(aVal.Pointer(), bVal.Pointer()). That token is therefore not an internal detail: it is the printed order of fmt.Println(map[I]int{…}).

Go answers with the descriptor’s machine address — the linker’s type-section layout. That ordering is deliberately unspecified: fmtsort’s own TestInterface says “the relative ordering of types is unspecified” and asserts only that same-type keys form contiguous sorted subgroups. It is also not a function of anything the managed side can observe; measured against Go 1.23.1, one program lays main.Mango before main.Zebra before main.Apple while another lays the same three names in alphabetical order.

The managed bridge has no addresses, so reflect.Value.Pointer answers ж<T>.PointerOrderToken, which for a descriptor box composes RuntimeHelpers.GetHashCode of the underlying abi.Type. That is worse than unspecified. CoreCLR draws an object’s identity hash from a per-thread PRNG, so the token is fixed for a given build but bears no relation to the type it describes, and the printed order flips whenever an unrelated edit changes how many hashes are drawn before these two. That is a coin flip re-tossed on every commit, and it landed tails: tests/Behavioral/InterfaceInheritance (a map[I]int keyed by main.T1/main.T2, both rendering as "", so only the values reveal the order) was graduated to output comparison green at c87a7f145 and printed map[:2 :1] against Go’s map[:1 :2] by c7b297c87. No commit in that window broke it — bisecting names a bystander. The demonstration is direct: adding a diagnostic that merely touches reflect BEFORE the Println moves both tokens and restores the correct order.

Fix: reflectPointerToken (reflect/value_impl.cs) now routes a pointer to a type descriptor — ж<rtype> or ж<abi.Type> — through typeDescriptorOrderToken, which packs the leading IntPtr.Size bytes of the type’s Go name big-endian, so comparing tokens arithmetically compares the names lexically. The name is the one Type.String() prints (GoReflect.GoTypeName over the descriptor’s sysType + arrayDims), which gives the invariant worth stating: types that print alike token alike, and types that print differently order by that printed name. Same-type grouping is therefore exact, and the order is stable across builds, runs and unrelated edits — the one property the address model cannot offer and the PRNG model actively destroys.

Two bounded residuals, both deliberate:

Only two call sites in the corpus read reflect.Value.Pointer: internal/fmtsort (this ordering, validated 3/3 including TestInterface’s grouping assertion) and vendor/golang.org/x/crypto/internal/alias, which compares slice-element pointers and never reaches the descriptor branch.