rea/docs/managed-code-analysis.md

# Managed-code analysis

REA inspects .NET PE/CLI artifacts without loading or executing their code.
It can identify an assembly, inspect metadata and CIL, compare members across
builds, and connect declared native calls to supplied native-analysis evidence.
CIL is Common Intermediate Language, the instruction format used by managed
assemblies.

The seven shipped tools are:

| Task                                                                        | MCP tool                            |
| --------------------------------------------------------------------------- | ----------------------------------- |
| Identify an artifact and its managed deployment form                        | `inspect_managed_artifact`          |
| Inspect types, signatures, method bodies, calls, and field access           | `inspect_managed_members`           |
| List declared native calls and native implementation indicators             | `inspect_managed_native_boundaries` |
| Compare members across builds and map build-local tokens                    | `compare_managed_members`           |
| Check native call declarations against supplied export or function evidence | `verify_managed_native_boundaries`  |
| Import decompiled code against verified static member identities            | `import_managed_reconstruction`     |
| Add managed findings to an application graph                                | `project_managed_application_graph` |

Each has a matching CLI command: replace underscores with hyphens and prefix
the name with `rea`, for example `rea inspect-managed-artifact`.
Native-body bridge mapping and managed runtime execution are outside the
current tool set.

The canonical tool inventory is the
[build-generated catalog](mcp-contracts.md#generated-catalog).

## Inspect an assembly

```bash
rea inspect-managed-artifact /absolute/path/Assembly.dll --json
rea inspect-managed-members /absolute/path/Assembly.dll --json
```

Use the artifact result to identify the assembly and its metadata coverage.
Member results retain build-local tokens, signatures, CIL bodies, and decoded
relationships. Use an exact reported `full_name` with `--type` when a focused
inspection is needed; see [selected member inspection](#selected-member-inspection).
Compare separate builds with [member comparison](#member-comparison), preserving
their paths, digests, and selection scope.

### Metadata overview and retained rows

The existing CLI and omitted/complete MCP detail return complete metadata
inventories. For identity and classification questions, select
`inspect_managed_artifact` with
`{"path":"/absolute/path/Assembly.dll","detail":"summary"}` on the first call.
This keeps the complete source in the session and returns a distinct linked
`inspect_analysis_view` overview with complete PE/CLI, classification and its
evidence, metadata status, module/MVID, assembly, frameworks, returned-row counts,
coverage/issues and limitations. No assembly is loaded or executed.

Use `normalized_result.parent_evidence_id` to page original rows:

```json
{
  "source": {
    "kind": "retained-evidence",
    "evidence_id": "<parent_evidence_id>"
  },
  "view": {
    "kind": "managed",
    "collection": "attributes",
    "offset": 0,
    "limit": 16
  }
}
```

Collections are `references`, `resources` and `attributes`; rows retain their
original values, nulls, order, metadata tokens and offsets. Pages retain source
metadata status and coverage. Exhaustion covers the returned parent collection,
not all possible metadata, and empty rows under partial/unavailable metadata do
not establish absence. Tokens are build-local and require the recorded artifact
SHA-256 and known module MVID for cross-observation matching.

Retained views do not repeat the parser or require the source file to remain
available. Portable views use the authenticated original Evidence inline,
including the existing `inspect-analysis-view` CLI. Session record lifetime and
export/import semantics remain unchanged. A server without Evidence retention
rejects selected summary before analysis; complete delivery remains available.

### Member summary and retained member rows

Complete `inspect_managed_members` delivery always includes the assembly-wide
`member_refs` and `exported_types`, which `type_names` does not narrow, so even
one selected or unmatched type can return megabytes. Select
`{"path":"/absolute/path/Assembly.dll","type_names":["Namespace.Type"],"detail":"summary"}`
to keep the complete source in the session and receive identity, metadata
status, the selection with its unmatched names, coverage/issues, the separate
user-string coverage, limitations and per-collection counts. Then page original
rows with view kind `managed-members`:

```json
{
  "source": {
    "kind": "retained-evidence",
    "evidence_id": "<parent_evidence_id>"
  },
  "view": {
    "kind": "managed-members",
    "collection": "call_edges",
    "offset": 0,
    "limit": 64
  }
}
```

Collections are `types`, `exported_types`, `fields`, `methods`, `member_refs`,
`call_edges`, `field_accesses` and `user_strings` (the literals). Each page names
its collection, retains original rows, order, tokens, signatures and body facts,
and carries the source selection and coverage; `user_strings` pages carry the
literal pool's own coverage. Older Evidence that never examined exported types
or user strings reports a null count and refuses those pages rather than
returning an empty one.

## Shipped scope

The canonical parser admits PE/CLI bytes and their metadata/CIL without loading
an assembly. It reports observed implementation markers and unavailable facts;
it does not unpack .NET single-file hosts, decode IL2CPP metadata, or infer a
NativeAOT identity from an ordinary native PE. Inputs without admitted CLI
metadata do not become managed assemblies through naming or routing guesses.
Inspect separately obtained components explicitly and preserve their identities.

A valid implementation marker establishes a candidate native boundary, not
recovered native semantics or runtime behavior.

An assembly targeting [.NET Standard](https://learn.microsoft.com/en-us/dotnet/standard/net-standard)
declares API compatibility across multiple .NET implementations. REA preserves
that target in `target_frameworks` and leaves `runtime_family` as `unknown`
unless independent markers identify a runtime family, such as a ReadyToRun
signature.

### Resource admission

The static provider admits regular input files up to 128 MiB before retaining
their bytes. Reads remain bounded if a file grows after admission. Metadata,
CIL decoding, and projected inspection records have a separate conservative
64 MiB representation budget; shared heap strings are cached, but repeated
output occurrences still consume that budget. These are inspection resource
policies, not PE format limits or a hard bound on total V8 heap use.

Path-based member comparison uses the same per-input snapshot policy and
inspects its two selected paths sequentially, retaining each inspection rather
than both full byte snapshots. The limit therefore remains a per-file resource
admission rule, not an aggregate comparison limit. Comparison Evidence keeps
the caller-selected paths as its parameters and artifact locations.

Input overflow returns a resource constraint. Decode or projection overflow
reports `resource-limit` with unavailable or partial coverage; omitted facts do
not become empty observations. Ambiguous PE section mappings are rejected
rather than interpreted according to section order.

### Selected member inspection

When whole-assembly member decoding reaches the representation budget, select
known declaring types before their signatures and CIL bodies are decoded:

```bash
rea inspect-managed-members /absolute/path/Assembly.dll --type Example.Parser
```

MCP uses `inspect_managed_members` with
`{"path":"/absolute/path/Assembly.dll","type_names":["Example.Parser"]}`.
Repeat `--type` or supply multiple unique `type_names` for several types.
Names match the reported `full_name` exactly. Top-level names are namespace
qualified; nested types currently report bare names, which can select several
types with that name.

The result preserves the requested names in `selection.type_names` and reports
unmatched names separately. If metadata names are unreadable, limitations retain
that uncertainty. With no selection, `selection` is null and whole-assembly
inspection keeps its existing resource policy. Selection restricts types,
fields, methods, call edges, field accesses and user strings; member references
remain an assembly-wide inventory, so output size is not guaranteed to be small.

Coverage describes the selected members. Comparisons leave members outside the
opposite observation's selection unknown, and application graphs report their
selected scope as partial. Neither makes an absence claim about the whole
assembly. Older member Evidence without `selection` remains readable.

### User strings

`ldstr` anchors keep their build-local `0x70` tokens. `user_strings.literals`
lists each distinct token loaded by an inspected method body once, in token
order, with its `#US` file offset and decoded value. Values keep their stored
UTF-16 code units, including unpaired surrogates. Token `0x70000000` names the
heap's empty entry at offset 0 and reads as `""`; any other item there is
malformed. The terminal flag is checked against ECMA-335 II.24.2.4, and the
length prefix against the width II.23.2 assigns. A difference is reported with
its byte offset as a user-string issue while preserving the readable literal.
Toolchain conventions can differ; this conformance observation does not
establish runtime rejection. A missing `#US` stream is one stream-level issue;
a malformed item is reported against its token. User strings carry their own
coverage: any of these makes `user_strings.state` partial and lists the issue
in `user_strings.issues`, while the member `coverage` and `metadata.status` are
unaffected, and readable literals remain. The state is also partial when an
inspected method's CIL was not fully decoded (a malformed exception section
alone does not count) or a method that may belong to the inspected types was
not inspected (its row could not be read, or its type's name could not be read
under a `type_names` selection), because `ldstr` instructions there were not
observed. Normalized CIL digests do not include decoded values.
`compare_managed_members` does not decode or compare literals: an unchanged
method does not establish unchanged strings, and the comparison says so in its
limitations. Older member Evidence without `user_strings` reads as `null`: not
examined.

### Edge target declaring types

`member_refs[].parent_name`, `call_edges[].target_declaring_type`, and
`field_accesses[].field_declaring_type` name the type that declares each
target. Referenced spec signatures validate encoded type coordinates against
the metadata tables, including nested type arguments and custom modifiers.
TypeRef resolution scopes are checked transitively with cycle detection;
readable names survive malformed scopes as partial evidence.
TypeDef and TypeRef names use the namespace-qualified `full_name`
form, and nested types keep bare names. A generic-instantiation TypeSpec
parent names its generic type, such as ``System.Collections.Generic.List`1``.
Other well-formed TypeSpec forms and ModuleRef or MethodDef parents stay
`null`. A `method-spec` call edge reports the name and declaring type of the
generic method it instantiates. A field access is named only by a field or a
field-signature MemberRef. TypeSpec and MethodSpec blobs are decoded with the
member-signature grammar, and a MethodSpec must instantiate a generic method
with as many type arguments as it declares. Malformed or unsupported rows and
blobs, and arity mismatches, make coverage partial. A readable MethodSpec
method column that names a method, or a TypeRef whose resolution scope is
invalid, keeps its name.
Older member Evidence without these fields reads them as `null`: unknown.

## Analysis objective

Start by identifying the artifact with `inspect_managed_artifact`, then inspect
its types and CIL with `inspect_managed_members`. Use [member comparison](#member-comparison)
for another build, and [managed/native boundary rules](#managed-native-boundaries)
when matching declared native calls to supplied native Evidence. Static inspection
does not load or execute the assembly. Reconstructed structure and unresolved
native/runtime questions retain their own authority and coverage.

## Member comparison

Names and semantic labels do not establish durable member identity. Comparison
preserves observed differences, structural inferences, and unknowns separately.

Cross-version matching first pairs members of the same declared type and name:
methods by exact CIL/signature identity, then by raw signature, then by
structural body shape; fields by raw signature. Exact CIL/signature identity
requires equal raw CIL digests and decoded fingerprints. Equal decoded
fingerprints alone cannot establish an exact match: they omit switch targets
and collapse distinct NaN encodings. Only then may a remap pair a
renamed or moved member across declared types or names: a method by exact
CIL/signature identity or structural body shape, a field by exact decoded
signature. A remap never takes a member whose own declared type and name remain
unpaired on the other side, so a member is never taken from its own changed
counterpart. A member whose declaring type is unresolved could belong to any
type, so it pairs by name only when no other member shares its exact
CIL/signature key, never by structural shape alone, and any remap group it
joins stays ambiguous. The exact-signature key uses names only as part of that
full tuple; names alone never select a pair. Duplicate tuples remain ambiguous
rather than being paired by token order. The raw signature is exact whether or
not REA decoded it, so undecoded signatures pair by that tuple too, but they
never enter structural rounds. An unmatched member is therefore reported as
`unknown`, not added or removed, when its own signature was not decoded or when
the other side has an unpaired member, one-sided or ambiguous, with the same
declared type and name whose signature was not decoded.

For a matched method, an unavailable or partial body makes body-shape facets
unknown while preserving observed signature differences. For complete bodies,
a difference in either the raw CIL digest or the decoded fingerprint is reported
in the `cil` dimension, including reordered switch targets with equal decoded
fingerprints. Structural identity compares normalized signatures, the limited
decoded-CIL tuple fingerprint,
constants, and bounded shape context. It reports all candidates at the winning
score and remains ambiguous when the evidence does not distinguish them. The
digest does not itself remap metadata tokens.

Tokens are always remapped through observed structure. A caller cannot carry
`0x06001234` into a new MVID and assume it names the same method.

## NativeAOT metadata recovery

NativeAOT removes the ordinary CIL method bodies, so a CIL decompiler cannot
reconstruct those bodies as C# source. The executable remains a native target
for Hopper or Ghidra, where decompilation produces pseudocode and analyst
inference. This is a different route from ordinary managed member inspection.

NativeAOT is not limited to Windows `.exe` files. The native image can be PE,
ELF, or Mach-O, and NativeAOT can also publish shared libraries (PE DLL, ELF
`.so`, or Mach-O `.dylib`) with explicitly exported entry points. Identify the
image from its bytes and loader metadata, not its filename extension. When the
image is inside an application bundle, inventory or extract the bundle first
and carry the component identity into native analysis. Matching PDB, ELF debug,
or dSYM sidecars can add symbol evidence when the selected provider supports
them; verify their identity and keep them separate from observations made from
the executable itself. [Microsoft's NativeAOT overview](https://learn.microsoft.com/en-us/dotnet/core/deploying/native-aot/)
lists supported OS/architecture targets and deployment limitations, while its
[native library guide](https://learn.microsoft.com/en-us/dotnet/core/deploying/native-aot/libraries)
describes exported shared-library entry points.

REA offers a built-in read-only snapshot parser for x64 PE/RTR 9.1 metadata and
an optional Linux headless Ghidra adapter for annotations and broader recovery.
The adapter requires a locally built `REA_GHIDRA_NATIVEAOT_JAR`; REA does not
install it or its toolchain. The [NativeAOT guide](ghidra-nativeaot.md) defines
these routes, their layout/host restrictions, derived-memory provenance, and
verification coverage.

REA's experimental Windows Ghidra boundary admits native x86 and x86-64 PE
applications and DLLs for static analysis. For x64 PE NativeAOT, the read-only
parser also runs there; the [NativeAOT guide](ghidra-nativeaot.md) lists what a
Windows host covers.

For NativeAOT analysis, preserve the selected native artifact's path and digest
and report recovered type data with its returned provenance. Linking those types
to native function addresses requires verified
provider evidence; names or vtable similarity alone do not prove that a native
function implements a specific managed method.

<div id="managed-native-boundaries"></div>

## Managed/native boundary rules

`verify_managed_native_boundaries` compares declared P/Invoke module and entry
point identities with supplied native export or function evidence. Name/module
matches establish declaration links; they do not establish that a call executed.

Metadata tokens are not native addresses. Native body mapping for C++/CLI,
ReadyToRun, NativeAOT, and IL2CPP requires separately verified format and provider
support. The managed static tools do not recover those mappings. The optional
[Ghidra NativeAOT extension](ghidra-nativeaot.md) has its own supported boundary.

Declared import inventory supports a bounded positive claim. Its absence does
not exclude dynamic resolution, generated code, protected code, native helpers,
or server-side behavior.

## Evidence record shape

Member comparison, application graph projection, and managed/native verification
authenticate supplied inspection Evidence and check that any subject SHA-256
matches the normalized artifact SHA-256. A valid Evidence ID alone does not
establish that those two identities agree. Evidence without a subject remains
usable, with no subject digest available to cross-check.

Artifact inspection returns the selected path, byte length and SHA-256, observed
PE/CLI classification, assembly and module identities, and metadata coverage.
Member inspection binds build-local tokens, signatures, row offsets, method
bodies, and decoded relationships to that artifact and MVID. Missing or malformed
facts remain explicit in coverage and per-member status.

Metadata table streams may carry the CLR-compatible `ExtraData` extension
(`HeapSizes` flag `0x40`), which Microsoft readers honor although it is not
specified by ECMA-335. REA skips its fixed four-byte word after the row counts
before decoding table rows; the word's value is not a length. An incomplete
word is reported as a metadata coverage failure at its file offset, without
reading beyond the declared table stream.

For unoptimized `#-` metadata, nonempty `FieldPtr` and `MethodPtr` tables
translate TypeDef list positions to physical member definitions. REA resolves
these indirections for member ownership and local attribute-constructor types;
field and method tokens still identify the original definition rows.
`field_list` and `method_list` retain the metadata's list coordinates, which
refer to the pointer table when present rather than necessarily contiguous
definition rows. Malformed or ambiguous pointers leave the owner unknown with
partial coverage while preserving readable member and attribute facts.
The retained inverse indexes are admitted against the existing representation
budget before allocation.

`exported_types` reports ExportedType rows as declared. Facade assemblies such
as .NET `System.Runtime.dll` provide their types this way: `forwarder: true`
marks a type forwarder, and `implementation_token` names the AssemblyRef, File,
or enclosing ExportedType row that implements it. REA does not load the
implementing assembly to resolve the forwarded definition, and
`compare_managed_members` does not compare exported types. Member Evidence
recorded before this field existed reads back with `exported_types: null`,
which means the table was not examined.

CLI `#GUID` heap values retain their exact 16-byte GUID text. REA does not impose
RFC 4122 UUID version or variant bits on MVID, EncId, or EncBaseId because
ECMA-335 metadata does not require those bit patterns. Tokens and RVAs are
build-local coordinates, not durable cross-build identities or native addresses.

## Decoded-CIL fingerprint

`inspect_managed_members` exposes two method CIL hashes with different byte
boundaries:

| Field                  | Input                                                                                                     | Excluded from the digest                                                                                                                                      |
| ---------------------- | --------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `il_sha256`            | The exact `il_size` raw CIL bytes after the tiny or fat method header                                     | Method header, locals signature blob, alignment, and extra sections including exception clauses                                                               |
| `normalized_il_sha256` | UTF-8 `JSON.stringify` output for the instruction-order array of `[opcode, operand_kind, operand]` tuples | Instruction offsets except projected scalar branch targets, switch targets, resolved metadata/string identities, method header, locals, and exception regions |

The normalized field is available only when the entire CIL stream decodes with
status `present`. A malformed or absent body reports `null`. The tuple format
is defined here.

Tuple operands use these exact projections:

- operand-free instructions: `none` and `null`;
- metadata and user-string operands: the build-local token as lowercase
  `0x`-prefixed eight-digit hexadecimal;
- scalar branches: the target CIL byte offset as a decimal string;
- `switch`: the case count as a decimal string, without target offsets;
- signed integer and variable operands: decimal strings; floating-point
  operands: JavaScript's `String` representation of the decoded IEEE value.

Opcode names retain short/long forms and prefix opcodes. This makes the value a
reproducible decoded-tuple fingerprint, not a canonical semantic CIL identity:
it does not resolve tokens across MVIDs, does not commit full control flow or
exception semantics, and does not replace the raw CIL digest. Cross-build
comparison requires equal raw CIL digests as well for exact CIL/signature
matches and reports ambiguity when structural evidence is insufficient.

For example, a decoded `nop; ret` body serializes exactly as
`[["nop","none",null],["ret","none",null]]` and has digest
`5e5fad7741cb44bca3a4f045546b7449990da343f612f5f34c0ca30e9eee0636`.
This vector specifies the tuple order, `null` handling, absence of whitespace,
UTF-8 encoding, and lowercase hexadecimal output.

## Tool and packaging boundary

The production parser is REA-owned TypeScript and ships with the existing Node
application. Deterministic conformance uses source-owned byte-built PE/CLI
fixtures and expected semantic facts. An optional real ILSpy lane runs only when
its executable is explicitly supplied; the default lane does not establish
independent pinned-oracle parity. Production inspection needs none of those tools.

- `ICSharpCode.Decompiler`/`ilspycmd` can supply reconstruction inference.
  A BYO reconstruction import records the supplied version
  and remains non-canonical. When `REA_ILSPY_CMD_PATH` points to an absolute
  runnable `ilspycmd`, `verify:managed` runs a source-owned real ILSpy oracle
  and imports its C# output as reconstruction inference against exact static
  member Evidence.
- The package contains no .NET runtime, SDK, ILSpy installation, proprietary
  assembly, or compiled conformance fixture.
- Setup never installs or upgrades those tools. Doctor may inspect an explicit
  optional path without changing it.

Real-tool downloads used in development or CI require exact package coordinates
and SHA-256 lock entries. They are cached outside the source tree and restored
into isolated directories. Updating any coordinate requires review of license,
runtime, supported host, output contract, and conformance results.

Target platform is not host support. The canonical parser must inspect
Windows-produced PE/CLI bytes on REA's supported Linux and macOS hosts with the
same result. Pinned Windows oracle jobs may provide additional conformance, but
they do not claim that the REA application itself supports Windows. Native
ReadyToRun, C++/CLI, NativeAOT, or IL2CPP coverage is additionally constrained
by the selected Hopper/Ghidra host and format matrix.

## Conformance verification

The current lane uses source-owned byte-built fixtures and malformed-input
regressions, with optional operator-supplied applications and ILSpy verification.

Run the deterministic lane with `npm run verify:managed`. See
[testing](testing.md#real-toolchain-verification-lanes) for prerequisites and
coverage. Generated outputs belong outside tracked fixture directories and must
not remain after verification or packaging.

Verification compares semantic facts, not decompiler text. It checks resource
ceilings, deterministic ordering, complete cleanup, CLI/MCP parity, Evidence
commitments, packed-package behavior, and target-free MCP startup. A parser and
an oracle disagreeing on malformed input is investigated explicitly; one tool's
acceptance does not automatically make the bytes valid.

## Operator-local osu! benchmark

The optional osu! benchmark accepts only paths and expectations supplied by the
operator. Run it with:

```sh
REA_MANAGED_APP_MANIFEST_PATH=/absolute/managed-app-manifest.json npm run verify:managed
```

`verify:managed` always runs a source-owned manifest-verifier self-test first.
If `REA_ILSPY_CMD_PATH` is set to an absolute `ilspycmd` path, it also runs a
source-owned ILSpy oracle: version discovery, class listing, bounded C# output
for a pinned fixture type, and import through `import_managed_reconstruction`.
The verifier prints only command/version, executable and output digests,
Evidence IDs, method locks, and compact fixture identity; it does not print
decompiled C# text. A failing ILSpy oracle fails the verifier because it is an
explicit real-tool claim, but leaving `REA_ILSPY_CMD_PATH` unset keeps the
oracle disabled.

When `REA_MANAGED_APP_MANIFEST_PATH` is set, the manifest's target path may be
absolute or relative to the manifest file. A compact manifest contains:

```json
{
  "label": "operator-local osu!stable semantic slice",
  "target": {
    "path": "./osu!.exe",
    "sha256": "<exact target digest>",
    "mvid": "<exact module MVID>",
    "assembly_name": "<expected simple name>",
    "runtime_family": "dotnet-framework",
    "managed_architecture": "x86"
  },
  "methods": [
    {
      "label": "operator-local semantic label",
      "token": "0x06000000",
      "signature_sha256": "<raw signature blob digest>",
      "il_size": 0,
      "il_sha256": "<raw CIL byte digest>",
      "normalized_il_sha256": "<digest>"
    }
  ],
  "application_graph": {
    "expected_node_kinds": [
      "artifact",
      "managed-assembly",
      "managed-module",
      "managed-type",
      "managed-method"
    ],
    "feature_traces": [
      {
        "label": "operator-local feature label",
        "method_token": "0x06000000",
        "seed": "<method name, API name, string, or digest to trace>",
        "match": "exact",
        "case_sensitive": true,
        "min_matched_seeds": 1
      }
    ]
  }
}
```

Before evaluating methods, the verifier fails closed on target SHA-256, MVID,
assembly name, runtime-family, and managed-architecture mismatches whenever the
manifest supplies those fields. It then pages each declared MethodDef token
directly by row, so selected methods do not need to appear in the first member
page of a large application. Optional `il_sha256` locks the exact raw CIL bytes
in addition to REA's decoded-instruction-tuple digest. That normalized field
has the byte boundaries documented above and is not a complete semantic CIL
identity.

The optional `application_graph` block reuses those exact-build method
commitments. For each referenced MethodDef token, the verifier builds a bounded
managed application-graph projection from the authenticated artifact and the
single selected member page, then checks requested node kinds and feature-trace
seed matches. This proves only that the selected static facts enter REA's graph
and tracing vocabulary for that exact local build; it does not execute the
application, inspect arbitrary unlisted methods, or infer runtime behavior.

Output contains assertion status and compact identities only: file name,
target SHA-256, MVID, assembly name, runtime-family, managed architecture,
method tokens, method names, signature digests, IL sizes, normalized IL
digests, graph Evidence IDs, node-kind summaries, and trace seed hit counts. It
does not print method bodies, reconstructed C#, application data, runtime
telemetry, account/service details, full filesystem inventories, or the
target's absolute path. The manifest and target remain outside git.

The benchmark may prove that REA reproduces selected facts for that exact local
build. It cannot establish general support on its own; source-built conformance
remains the admission requirement.

## Runtime behavior boundary

Managed inspection does not execute assemblies or observe CLR internals.
Questions about actual behavior remain unknown until supported runtime
evidence is collected. A direct process capture can retain declared inputs,
outputs, filesystem and protocol activity for a target run; it does not prove
which managed methods executed or how the CLR resolved them.

## Method metadata during build comparison

Managed member comparisons report a `metadata` dimension when an observed
MethodDef flags or implementation flags value differs, including accessibility
or synchronization changes. This is separate from the CIL/signature matching
tiers: identical instructions do not establish unchanged method metadata. The
reported difference is static metadata evidence, not proof of runtime behavior.