procedure_callees and procedure_callers returned a bare address list, so an
inferred no-return call that truncates Ghidra's function body silently dropped
real edges (fixes#2326). Wrap both views in the function-observation envelope
with noReturnCallLimitations, matching procedure_info, procedure_assembly,
procedure_pseudo_code, read_function_instructions, and analyze_function
(precedent: 69a8b3a95). Version the boundary-observation contract to
ghidra-terminal-call-limitations-v2 so older snapshots cannot replay
unqualified results.
Validation: red->green unit test for the envelope requirement, MCP boundary
evidence test extended for both views (limitations retained through Evidence
and snapshot replay), 8 harness test files + ghidra behavior tests green,
check:fast, docs:check, verify:test-discovery green. Real Ghidra host lane
unavailable here (no Ghidra installed); Java change mirrors the existing
procedureInfo call pattern.
* fix(native): refuse inspect_signature on non-Mach-O targets
This provider's signature inspection covers Mach-O code signatures only.
On a xar pkg, codesign's "code object is not signed" verdict cannot see
the archive-level signature layer, so a signed or invalidly-signed package
was reported as signed:false with "Artifact is not signed." Match the
existing inspect_macho/list_architectures gate and report
capability_unavailable instead of an unsigned claim for a layer that was
never evaluated.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(native): scope the signature-gate comment to this provider
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(native): align signature capability discovery with Mach-O support
Bind inspect_signature availability to Mach-O targets in capability discovery and align the public tool description.
Source inspection and git diff --check; tests/builds/typechecks/CI reruns skipped at explicit user request.
---------
Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* feat(ghidra): admit raw NMOS 6502 analysis
* test(ghidra): require verified POSIX verifier cleanup
* test(ghidra): verify 6502 low-memory raw imports
* docs(skill): document raw NMOS 6502 support
* fix(ghidra): preserve requested Windows raw source paths
* docs(ghidra): include raw 6502 in Windows installation scope
* test(ghidra): verify Windows raw junction rejection through CLI and MCP
* fix(validation): make fixtures portable and preserve explicit Codex home
* fix(ghidra): attest raw mapping state and address space
* fix(ghidra): align diagnostics and regression groups
* fix(ghidra): bind raw attestation to language address space
* chore(ghidra): avoid verifier insertion conflict
* docs(ghidra): align linked raw admission guides
* fix(ghidra): retain 6502 verifier workspace on failure
* test(ghidra): isolate Windows authority doctor fixture
* fix(binary): re-admit Windows raw paths on session reopen
* test(binary): require fresh provider execution after path change
* Constrain 6502 to raw lifecycle and snapshot targets
* Bypass saved reads for unadmitted Windows paths
* Guard direct CLI snapshot replay on Windows paths
* fix(snapshot): gate Windows Ghidra replay on native admission
* fix(snapshot): constrain nested 6502 subjects to raw images
* test: make dyld slice fixtures portable
* perf(native): add opt-in Mach-O summary delivery and retained pages
inspect_macho returns every import and export inline. For VS Code's
Electron Framework one MCP call carries 2.5 MB of text (about 828k o200k
tokens), and the same again as structuredContent; the MCP response
budget admits it. Agents asking for slice identity, UUID or dependency
counts paid that cost on every first call.
Add the established opt-in detail selection: omitted or complete keeps
the existing complete result and Evidence identity, while summary
retains the complete Evidence and returns slice identity with
per-collection returned/total/exhaustive counts (956 tokens).
inspect_analysis_view kind macho pages the original rows of any
collection or the command provenance, repeating the producer's own
total and exhaustive flag so page exhaustion is never read as absence.
Delivery detail never reaches the native tools or the source Evidence
ID: registerEvidenceTools gains a parametersForEvidence hook, so the
summary's parent is the same Evidence as a complete call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(mcp): refuse unretainable summary delivery before analysis
A server without Evidence retention (a fixed analysis source) ran the
complete inspect_macho or inspect_artifact analysis for detail summary
and only then refused it in summarizeRetainedAnalysis, paying the very
cost summary delivery exists to avoid. registerEvidenceTools gains a
preflight hook; both registrations check retention there, as the WASM,
EVM, managed, pcap and capture handlers already do before dispatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(native): keep whole rows in Mach-O architecture and segment pages
The page item union listed plain-object row arrays, and zod returns the
first union member that parses while dropping unknown keys. Entrypoint
rows declare only file_offset, so every architecture and segment row
(which also carry a numeric file_offset) matched it and lost its name,
CPU fields, permissions and sections. Rows now use exact key sets, and
imports/exports share one member. The MCP test pages every facet of a
real /bin/ls and compares it with the unparsed complete rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Group observations by exact normalized instant, retaining arbitrary fractional precision and offset equivalence. Coalesce equivalent assessments and all supporting Evidence; report conflicting latest assessments as unknown with verifier-result-ambiguous. Use the same instant comparison for freshness and future checks.
Replace first-input-wins coverage with permutation, provenance, offset, sub-millisecond ordering, and future-clock regressions, including the public MCP workflow. Validation: focused reconstruction/offset tests and MCP workflow passed; converted JSON Schema validation, typecheck/lint, and docs checks passed. Full suite and real-provider lanes were not run.
Summarize saved Wireshark4.x captures with enforced native resource limits, serialized admission, stable snapshot verification and retained fact/text pages. Preserve complete inventories within resource budgets and source uncertainty; report cleanup failure with partial Evidence and retained ownership.
Verified six recorded corpora /13026 packets with independent Scapy/Python oracles, real CLI/MCP/schema parity,60 SNI names,132 DNS addresses and1000 URIs. Real TShark CI now exercises summary/paging. Actual Windows/Bun Wireshark remains unverified; macOS capture requires a separate enforceable-bound decision.
Co-authored-by: 3QL <85309073+3QL@users.noreply.github.com>
Co-authored-by: N0zoM1z0 <161784452+N0zoM1z0@users.noreply.github.com>
Reuse the domain ZIP format list for session schemas, verify full Harmony
CLI/MCP lifecycles, and correct native view guidance and audited docs drift.
Refresh locked dependency notices and gate inventory/provenance consistency,
including notice inputs in the cached documentation gate.
Fixes#2454
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Add explicit MCP Dart AOT summary delivery and exact retained string-pool pages while preserving complete defaults, CLI behavior, ABI selection, original Evidence identity, partial coverage and lower bounds.
Real LocalSend SDK replies shrink 92.05% with approximately unchanged SDK time. Two actual Codex pairs preserve all selected facts and reduce total input 48.75% / 42.38%, uncached input 33.63% / 42.52%, and task time 23.95% / 30.30%; all client errors/repeated calls remain counted. Preserve old portable Evidence and the real CLI complete result after rebasing onto SAZ/qlog main. Curated results keep measured and integration runtime pins separate; raw artifacts stay outside Git.
Refs #2016.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Inspect bounded recorded raw HTTP/XML SAZ archives and page retained capture Evidence. Preserve exact stored bytes and provenance, validate ZIP/XML integrity, and require an enforceable decoder memory bound.
Verified 356 real recorded sessions against independent byte and metadata oracles, public CLI/MCP parity, retained pages, and related interface checks. Current Fiddler Everywhere GUI and non-Linux producer workflows remain unverified.
Co-authored-by: zahhak <1792039+zahhak@users.noreply.github.com>
Co-authored-by: N0zoM1z0 <161784452+N0zoM1z0@users.noreply.github.com>
Add explicit package summary delivery and exact retained manifest pages while preserving complete MCP defaults, one-shot CLI output, provider requests, source identity and session lifetime. Return distinct linked Evidence and preserve partial/null facts; reject unavailable retention before engine work.
Validation: 27 relevant SDK/domain/Android boundary checks, focused test-only correction rerun, full supported compiler/lint and docs checks, real NewPipe SDK pairs, authentic old Evidence continuation, four actual Codex trials with failures/regressions retained, and all 39 exact-head CI checks. Package replies shrink 119,527 to 8,659 bytes; no model-token or time gain is established.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Keep complete default delivery and exact filters. Select stable ID pages at the ledger owner before cloning, return filtered totals and continuation offsets, and preserve detached records through the shared session port.
Validate public MCP paging, safe-integer edges, snapshot replay and converted JSON Schemas. Document the bounded Flameox workflow comparison and its token/CPU/latency measurement limits; no real model performance claim.
Add caller-selected managed metadata summaries and original collection pages while preserving complete CLI/MCP defaults and previous-runtime Evidence. Real SDK replies shrink substantially; actual Codex token/latency results remain mixed and are documented without a token-saving claim.
Validated focused SDK/domain tests, full static/docs checks, real two-assembly repeated workflows and portable old Evidence, real SQLite WASM, and all exact-head CI checks. Raw experimental artifacts remain private.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
* feat(server): frame analysis-target content as untrusted data
Strings, decompiled code, page text, scripts, captures and logs returned
by REA originate from the analysed target, which can embed text aimed at
the agent. Nothing marked that content, so a prompt injection inside a
binary or page could steer the agent toward tools that run code.
- Server instructions now state that target-derived tool output is data,
never instructions, and that any request in it to run, install,
download or send something is a prompt-injection attempt to report.
- Tool contracts declare returnsTargetContent. It is on by default; only
seven tools whose results hold REA session state or caller-supplied
values (binary_session, close_binary, export_evidence_bundle,
record_unknown, update_unknown, inspect_adb_client,
inspect_apktool_client) are audited exceptions, so new tools start
marked.
- Successful results of marked tools carry
_meta.provenance = { untrusted_content: true, source: "analysis_target" }.
content and structuredContent are unchanged, and the response budget
accounts for the fixed marker bytes.
Tests cover the encoding and budget accounting, the audited exception
list against the contract catalog, and an in-memory MCP client that sees
the instructions and the marker on a marked tool but not on
binary_session.
* fix(server): mark failed target-content results as untrusted too
The untrusted provenance marker was attached only on the success branch
of ToolResultDelivery, but failures can carry target-derived text as
well: error messages, captured output, and partial observations such as
the terminal frames a failed capture_process_scenario retains in
partial_observation. A host gating later process or device actions on the
marker would have treated an adversarial error payload as trusted.
Every error path in src/server now carries the failing tool's provenance:
- toErrorToolResult and toRecordedPartialErrorToolResult take the tool's
target-content flag explicitly, so each caller is audited by the type
checker; toCallToolResult, toEvidenceToolResult (recording failure), the
response-budget error that replaces an oversized success, and the
Electron partial-evidence path pass contract.returnsTargetContent.
- Admission failures in runAdmittedToolOperation resolve the flag from the
tool name through the new toolReturnsTargetContent, the same predicate
the contract metadata uses; unknown names stay marked.
- The transport's oversized-error replacement keeps the original's marker
and counts it in its budget check, matching the projected error.
- Errors the MCP SDK builds itself (input validation, thrown handlers) are
marked in the transport for target-content tools only.
Error encoding accounts for the marker bytes exactly as success encoding
does. docs/mcp-contracts.md now states that error results carry the
marker. Tests cover marked and unmarked error results, the
capture_process_scenario partial-observation errors in-process and over
the SDK transport, the recorded partial-evidence path, the budget
replacement, the oversized-error replacement, and SDK validation errors
for a marked and a content-free tool.
* fix(contracts): mark session status, close and client probes as target content
The content-free exception set was audited against successful results
only. Now that failures carry the marker too, every exception was
re-audited against its error paths with one question: can the result
include text produced by an external process, provider or device, or
stored content REA did not author?
Removed from TARGET_CONTENT_FREE_TOOLS:
- close_binary: a failed provider cleanup (GhidraClient cleanup error via
createGhidraDiagnostics) projects the launcher's captured stdout and
stderr, which can contain target-derived log text.
- binary_session: capability status reports runtime unavailability as the
provider error message (ProviderAdapterError and similar), and the
provider candidate and Android availability diagnostics include probe
output such as the apktool stderr excerpt.
- inspect_adb_client: the success result carries the version and
installed path parsed from `adb version` stdout, and a failed probe
keeps a stderr_excerpt.
- inspect_apktool_client: the success result carries the first line of
the launcher's `--version` output, and a failed probe keeps a
stderr_excerpt.
- update_unknown: the new revision carries forward the stored question
and domain, which an imported evidence bundle can supply; list_unknowns
and get_evidence_bundle already return that data marked.
The set now holds only export_evidence_bundle (path, byte and record
counts; failures are filesystem, cancellation, or a fixed integrity
message) and record_unknown (the caller's input plus REA-computed ids
and digests; registry errors project only a reason code).
The conformance test pins the reduced set and asserts each removed tool
is marked. Tests that used binary_session as the unmarked example now use
record_unknown. docs/mcp-contracts.md lists the remaining exceptions and
why the removed tools stay marked.
* fix(contracts): include target-content policy in the catalog identity
createToolCatalog() and CATALOG_IDENTITY.tools projected every
client-visible contract field except returnsTargetContent. Because the
untrusted-content framing left schemas, effects, prompts, CLI names and
the package version unchanged, a same-version MCP process started before
the update digested equal to the new server and binary_session reported
it aligned although it never emits _meta.provenance.
Project the flag as returns_target_content in both the digested tool
catalog and CATALOG_IDENTITY.tools, and export createCatalogDigests so a
boundary test proves that toggling the flag on one tool changes the
tools and combined digests.
Treat missing, partial, unknown, and unresolved call coverage as insufficient for call-set equality. Combine IDA caller/callee coverage without losing either explanation and recheck source coverage when verifying historical comparisons.
Scope generated function-comparison unknowns with optional typed comparison/dimension facts, preserving legacy IDs and revision digests when absent. Keep unrelated dimensions conclusive, explicit dependencies applicable, and ambiguous function identity unknown. A confirmed change for matched functions remains a failure even if calls are incomplete.
Validation: function/reconstruction/residual-unknown and bundle tests, recorded IDA protocol regression, public MCP comparison/verification, and advertised SDK JSON Schema checks. Real IDA provider execution was not verified.
Fixes#2222
Evaluate stdout and stderr together, giving hard errors and non-ok wait statuses precedence over an attempted Starting line. Retain raw streams and typed partial launch facts when Activity Manager refuses or times out; retain successful raw output too.
Validation: parser/provider tests and public MCP app/activity starts through a fake executable at the real process boundary. Real Android device behavior was not verified; device-reported timeout remains distinct from a host deadline.
Fixes#2225
Inspect selected fields in existing captures through shared CLI/MCP Evidence.
Page matching frames without changing dissection or two-pass read coverage;
retain failed descriptor/process cleanup ownership and reject excess admission.
Verified with real TShark 4.0.17 and 4.2.2, public Wireshark recordings,
owned HTTP application traffic, real CLI/MCP workflows and schema checks.
Co-authored-by: Olorinm <59321404+Olorinm@users.noreply.github.com>
Assisted-by: Kogasa (gensokyo-skills:kogasa-surprise-testing)
Add opt-in EVM MCP summary delivery and retained/portable function and text views. Keep existing complete defaults, provider arguments and original full Evidence identities and records usable. Require retention before selected analysis; preserve inferred candidates, nulls, diagnostics, provenance and explicit unknowns.
Verify old schemas, historical Evidence migration, 64 focused feature checks, 30 affected MCP checks after latest-main rebase, 59 real EVM CLI/MCP cases and six shared-domain artifact workflows. Repeated SDK delivery is smaller; all real model trials and mixed token/time outcomes, limitation failures and resource qualification gaps remain documented without a uniform performance claim.
Refs #2016
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Add offline SQLite database/WAL inspection through the shared CLI/MCP.
Use compact inventory, selected-table schema, exact ordered columns and
bounded row samples while preserving source identity and lossless values.
Retain completed Evidence through late cancellation and private-root cleanup
failures, bound concurrent inspections, and recheck journal validity after
acquisition. Reuse prepared CI runtime artifacts across supported Node lines.
Co-authored-by: Y0UW1SH <29498139+Y0UW1SH@users.noreply.github.com>
Assisted-by: Kogasa (gensokyo-skills:kogasa-surprise-testing)
Bind existing EVM CLI/MCP metadata and the legacy service factory through one static descriptor, and register the same selected service that owns cleanup. Preserve original names, schemas, inputs/full defaults, configuration, Evidence and injection identity.
Verified exact SDK catalog/fact parity across six real EVM pairs, 59 actual EVM CLI/MCP cases, six real cross-domain workflows, six actual Codex CLI fact/reference/lifecycle trials and relevant local checks. CI run 38283026233 succeeds for exact head d999d9874b. No token reduction from wiring or fully qualified resource claim; final counters remain unknown. Continues #2016.
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
* feat(flutter): identify Dart snapshot builds by pure APK parsing
Add a flutter-provider family with one operation, identify_flutter_build,
that reads an APK's lib/<abi>/libapp.so and libflutter.so through the
hardened ZIP reader and reports build facts as observations:
- The Dart VM snapshot hash, read 20 bytes after every snapshot magic
(f5f5dcdc). The vm and isolate sections are cross-checked: agreeing
sections collapse to one snapshot_hash, disagreeing sections keep every
candidate with snapshot_hash null and partial coverage.
- Per-ABI library digests and byte sizes, plus the engine library's GNU
build-id note and any labeled 'Dart VM version' or Android clang/LLVM
toolchain strings, verbatim; release builds without a version string
report null rather than a guess.
- The APK is digested before parsing so Evidence binds the observation to
the exact input, and a target with no libapp.so reports not-detected
with complete coverage.
The scanner is streaming with explicit pending-magic and rescan
bookkeeping, so chunk boundaries cannot double-count or drop sections, and
magic-free streams never accumulate. No engine, no external lookups: the
hash-to-SDK mapping stays with external datasets by design.
Verified against Flutter Gallery 2.9.2 (112 MB, three ABIs): every ABI
reports the same agreeing hash 65817c30a78bb44c3dc3771876b6010a from two
sections, per-ABI build-ids, and the clang toolchain line; a plain
aapt2-built APK reports not-detected with complete coverage. Scanner
goldens come from the same libraries. READMEs carry the new 'Flutter apps'
row in every supported language.
* feat(flutter): inspect Dart AOT snapshots through the ELF symbol table
Add inspect_dart_aot to the flutter family: for one ABI's libapp.so it
locates the four _kDart*Snapshot* symbols through the ELF dynamic symbol
table (ELF32 and ELF64 little-endian, bounds-checked reads; a payload
without exported symbols is refused), reports each section's offset and
size, validates data-section headers (magic, kind, length; instructions
sections honestly report nulls for the fields their layout does not
carry), and projects a bounded string pool from the isolate data section:
distinct package: source URIs (the app's dependency inventory), dart: SDK
URIs, and identifier-shaped tokens, each with counts, bounds, and partial
coverage when truncated.
The string pool is labeled as a byte-pattern projection, not deserialized
cluster semantics: typed class and function recovery remains an
SDK-specific parser, which the snapshot hash now unlocks. The ABI is
caller-selected with a deterministic default (arm64-v8a, x86_64, then
first available).
Verified against Flutter Gallery 2.9.2: all four symbols on both ELF64
(arm64-v8a) and ELF32 (armeabi-v7a) images, header validation, hash
agreement with identify_flutter_build, and a 756 package: / 201 dart:
URI pool from the isolate data section; goldens in DartElf.test.ts come
from the same libraries.
* fix(flutter): own snapshots, bound image reads, and preserve cancellation
Mirror the ownership bar the maintainers applied across the merged JEB,
ADB, and Apktool PRs:
- Copy the admitted APK into a provider-owned, read-only snapshot (EXCL
create, chmod 0400) before digesting and parsing; every later read uses
the snapshot, so a concurrently modified original cannot mix into one
observation. The result reports the snapshot's byte size and digest.
- Retain failed snapshot-root cleanup in a provider set and retry it on
close(), throwing an aggregate error when a root is still owned; root
removal is injectable so the retention is testable.
- Allocate the AOT image buffer at exactly the declared library size
before reading (bounded before allocation, not concat-after-stream),
keep the 256 MiB ceiling, and check cancellation between digest and
stream chunks through a captured abort reason.
- Accept APK paths containing spaces (covered by a fixture test) and
escape the ABI pattern's dash for v-mode regex compatibility.
Also add tests for cleanup retention with close retry and for
cancellation before any snapshot root is created.
* docs(flutter): cover the Italian README with the Flutter row
The Italian translation landed between the merged feature PRs; its table
now carries every merged row plus this one, keeping its documentation-link
set identical to README.md.
* fix(flutter): make the ABI pattern portable for v-mode validators
Use a hex-escaped hyphen in the ABI character class, matching the form
the ADB portability fix uses, so clients compiling advertised schemas in
unicodeSets mode accept the inspect_dart_aot input schema.
* refactor(flutter): reuse shared harnesses and deduplicate the vertical
The duplicate-detection gate (jscpd, 2.5% threshold) tipped to 2.51% with
the flutter files added. Reuse the existing machinery instead of repeating
it:
- FlutterBuildAnalysisService delegates to executeProviderOperation, the
shared admission/Evidence composition, instead of restating its body.
- The CLI registers both commands from one table-driven spec.
- Boundary tests use the shared serveCli and assertAdvertisedMcpContracts
harnesses rather than hand-rolled equivalents.
- The provider test consolidates the four payload-shape cases into one
table-driven case; the service test uses one admission table and
distinct phrasings.
Duplicate lines fall from 2.51% to 2.48%, back under the gate.
* fix(flutter): address review findings from real-app verification
CassieFox23 verified the PR against three real F-Droid Flutter apps and
a control app, and found five gaps. All are fixed and re-verified against
cone (info.tangential.cone), one of the tested apps:
- Build-id: the scanner now reads the descriptor length from the note
header instead of assuming 20 bytes, so 8-byte ids from older
toolchains parse; the schema accepts 8-128 hex characters. Verified
against cone's readelf-reported ids on all three ABIs.
- Compiler line: the toolchain capture bound rises from 200 to 500
characters, keeping cone's 262-character clang line verbatim.
- Zero-byte libapp.so: the extraction target is created before streaming,
so an empty entry yields a clean unsupported_target refusal for
inspect_dart_aot and elf_header_valid: false with partial coverage for
identify_flutter_build, instead of a raw ENOENT with code UNKNOWN.
- Truncated library: identify_flutter_build marks fewer than two snapshot
sections as partial coverage with a limitation naming the count, and a
missing ELF magic gets its own limitation.
- Docs: the default ABI preference now matches the code exactly
(arm64-v8a, x86_64, x86, armeabi-v7a, then the rest alphabetically).
Also: every test that read the author's /tmp/opencode goldens now builds
a self-contained synthetic ELF64 image with dynsym, four snapshot
symbols, two snapshot sections, and a string pool, so the suite runs
anywhere; the verify lane now requires a present hex build-id per ABI
instead of accepting null.
* fix(flutter): address the Codex review findings
Eight findings, all applied:
- P1: SnapshotHashScanner keeps candidates in a Set (no quadratic
Array.includes) with a 64-distinct-hash cap, so a hostile payload of
magic+hash records cannot stall the scan or bloat the response.
- P1: projectStringPool stops admitting new distinct entries once each
category reaches twice its projection bound and reports
collectionCapped; allocation and sorting now track the response, not
the payload.
- P2: ZIP reader rejections during entry iteration or entry streaming
map into the typed error algebra (cancellation vs input refusal)
instead of escaping as unhandled exceptions in both operations.
- P2: magic_valid now marks coverage only for the *Data symbols; a
complete four-symbol image with valid data magics reports complete
coverage (the instruction symbols intentionally lack the magic).
- P2: the MCP server closes the Flutter provider during shutdown, so a
retained snapshot root is retried and reported, not leaked.
- P2: LabeledStringScanner commits matches only when a printable run is
delimiter-terminated; the carry grows to 600 bytes so a 500-character
toolchain line survives any chunk boundary untruncated.
- P2: both Flutter tool effects declare writesFilesystem: true; the
owned APK snapshot and extracted libapp.so are temporary writes.
- P2: the verify lane asserts every ABI agrees on exactly one snapshot
hash instead of discarding the computed set.
* fix(flutter): address the second Codex review round
Nine findings on the first-round fixes:
- P1: identify_flutter_build enforces a cumulative extraction budget
(768 MiB) across every library entry, not just per library, so an APK
with many ABIs cannot stream an unbounded total.
- P1: a single NUL-free printable run stops materializing at 1024 bytes
and marks the pool capped, so an adversarial one-run section cannot
allocate unbounded strings.
- P1: ELF cString bounds the NUL-terminator scan to the name maximum
plus one, so malformed symbol tables cannot scan the whole image per
entry.
- P2: mapReaderFailure distinguishes ArtifactReaderFailure texts —
budget/limit rejections map to resource constraints, integrity
failures (CRC/overlap/signature) to their own input refusal — instead
of collapsing all reader failures into one message.
- P2: the labeled-string carry is cleared when a delimiter commits a
run, so consecutive toolchain lines are never concatenated.
- P2: string-pool counts are renamed *_count_at_least and documented as
lower bounds, because capped collection discards values beyond twice
the projection bound.
- P2: inspect_dart_aot marks hash disagreements, non-pair source counts,
and candidate-cap saturation as partial coverage.
- P2: only valid lowercase-hex hash records count as snapshot sources;
complete magic records with non-hex fields do not inflate agreement.
- P2: the per-entry size refusal in inspect_dart_aot moves after ABI
selection, so an oversized unselected ABI no longer refuses a small
request.
* fix(flutter): address the third Codex review round
Five findings:
- P2: the owned snapshot revalidates the source after the copy (size,
mtime, inode), so an in-place replacement or swap mid-copy refuses
the observation instead of publishing mixed bytes.
- P2: a failed ZipArtifactReader close retains the snapshot root for
close() retry instead of silently swallowing the descriptor leak,
in both operations.
- P2: snapshot headers are parsed within their symbol bounds and a
symbol smaller than the 16-byte header reports magic_valid false,
so trailing-image bytes can no longer fake a valid data section.
- P2: the verify lane counts provider.close() cleanup failures as
lane failures instead of certifying a run with retained roots.
- P2: LabeledStringScanner reports toolchain_capped when further
distinct matching lines are observed past the 4-line projection;
identify marks such results partial with a naming limitation.
* fix(flutter): address rounds three and four of the Codex review
Round three (one missed finding) plus round four (five findings):
- P2: the GNU build-id note descriptor size is decoded from all four
bytes of the little-endian field, rejecting nonzero high bytes
instead of masking them to 16 bits.
- P1: the Flutter MCP handlers route through createContractToolHandler
and the session's analysisAdmission, so close_binary and snapshot
operations drain in-flight Flutter calls instead of racing them.
- P2: the owned-snapshot copy is a streamed, abort-checked write (no
copyFile), so cancellation stops the write promptly while the
source-identity revalidation (size, mtime, inode) still guards
against mid-copy replacement.
- P2: identify marks any snapshot_hash_sources value other than exactly
two as partial, matching the inspection path's pair requirement.
- P2: the verify lane refuses to run without --apk (exit 2) instead of
printing a false-green pass, and the engine check requires every
libapp-carrying ABI to have a present libflutter with a valid hex
build-id rather than filtering missing engines away vacuously.
* fix(flutter): address the fifth Codex review round
Six findings plus the upstream duplication baseline:
- P2: the flutter-provider kind reports available in the capability
inventory, so binary_session no longer lists always-registered
Flutter tools as provider_missing.
- P2: printable runs that exceed the 1024-byte cap are discarded at
their delimiter instead of emitting their truncated prefix as a
fabricated observation.
- P2: ELF symbols must fit inside the file-backed portion of their
PT_LOAD segment (p_filesz), so addresses in gaps or BSS are refused
instead of accepting unrelated bytes.
- P2: the build-id carry spans the complete 80-byte maximum note
(16-byte header + 64-byte descriptor) so chunk-boundary splits
cannot drop the first byte.
- P2: the labeled-string carry is sized for the toolchain pattern's
true 605-byte maximum, not the previous 600.
- P2: the AOT completeness predicate requires exactly two sources and
one candidate unconditionally, so zero-source images with valid data
magics report partial instead of complete.
Also: the shared ELF fixture builder and snapshot-section helper moved
to a dedicated FlutterTestFixtures module, eliminating cross-file test
duplication; jscpd now passes at 2.50% while main sits at 2.52%.
* fix(flutter): address the sixth Codex review round
Five findings:
- P2: snapshot-hash scanning is bounded to the exported data symbols'
byte ranges, so decoy or stale magic/hash sequences elsewhere in the
image cannot inflate source counts or contribute candidates.
- P2: labeled-string runs longer than the carry bound are rejected
instead of truncated, preventing a regex prefix from being published
as a verbatim source string.
- P2: the AOT image reread checks cancellation between file reads and
before the synchronous parsing stage, so a cancelled caller cannot
receive a success from a completed 256 MiB read.
- P2: the verify lane requires exactly two snapshot sources per ABI,
matching the provider's pair requirement, instead of accepting any
count >= 2.
- P2: ELF section and program header entry sizes are validated against
the class-defined minimum widths before walking either table, so
undersized entries cannot cause field reads to cross into
neighboring bytes.
* fix(flutter): address the seventh Codex review round
Six findings:
- P2: data-symbol ranges are deduplicated by offset and size, so aliased
symbols mapping to the same bytes are scanned once and the result
marks partial coverage with a limitation naming the shared range.
- P2: openEntrySafely forwards return/throw to the underlying Readable's
iterator, so an early for-await exit closes the ZIP entry stream and
releases its producer instead of leaking a backpressured pipe.
- P2: inspect_dart_aot now reports snapshot_hash_candidates and
snapshot_hash_sources alongside snapshot_hash, so a null hash from
conflicting evidence is distinguishable from missing hash data.
- P2: overlong printable runs only set toolchain_capped when they begin
with a recognized label prefix (Dart VM version: or Android (); other
long strings such as license text are skipped without marking partial.
- P2: readElfSymbol falls back to PT_DYNAMIC/DT_SYMTAB when section
headers are absent (e_shnum 0, ET_DYN), so stripped Flutter libraries
with exported snapshot symbols are no longer refused; the dynamic
table is bounded by DT_STRSZ, DT_SYMENT, and the image end.
- P2: the verify lane exercises inspect_dart_aot for every ABI carrying
libapp.so, not just the preferred default.
* fix(flutter): address the eighth Codex review round
Five findings:
- P2: each data-symbol range gets a fresh SnapshotHashScanner so the
carry from one noncontiguous slice cannot fabricate a record spanning
into the next; results are aggregated after scanning.
- P2: dynsym name lookups are bounded to the linked string table —
st_name offsets outside the table and terminators beyond it are
rejected instead of reading unrelated payload bytes as symbol names,
in both the section-header and PT_DYNAMIC paths.
- P2: identify_flutter_build routes its reader.entries and reader.open
calls through the same collectEntriesSafely and openEntrySafely
wrappers as inspect_dart_aot, so malformed ZIPs surface as input
refusals instead of AnalysisProtocolError.
- P2: mapReaderFailure recognizes ArtifactReaderFailure texts carrying
a cancelled reason, mapping them to AnalysisCancelledError instead
of AnalysisInputError.
- P2: the cumulative extraction cap is enforced inside the chunk loop,
not after each library completes, so refusal happens before the
remaining bytes are streamed and hashed.
* fix(flutter): harden streaming, sizes, and cancellation after self-review
Four issues found in a systematic self-audit across all bug classes
from eight Codex rounds:
- createReadStream in both the digest function and the owned-snapshot
copy now receives the AbortSignal directly, so cancellation aborts
the filesystem read at the source instead of waiting for the next
chunk to arrive on a slow filesystem.
- The owned snapshot enforces a 2 GiB APK-level size cap before any
copy I/O, so a path pointing at an enormous file is refused before
disk usage grows.
- identify_flutter_build preflights the sum of declared entry sizes
against the 768 MiB cumulative budget, refusing before any ZIP
extraction when the declared total already exceeds it.
- The LabeledStringScanner triple result() call is collapsed to a
single invocation (the previous code called result() three times in
the emission, relying on end() being idempotent).
No behavior change for valid inputs; all 36 tests, check:fast, and
the cone verify lane pass unchanged.
* fix(flutter): require structurally parseable ELF headers for complete coverage
CassieFox23's third corrupt-header case: an ELF with valid magic but
zeroed section/program header offsets passes the 4-byte magic check
and the byte scanner still finds hash patterns in the data sections,
so identify_flutter_build reports complete coverage for a structurally
broken file.
The elf_header_valid check now also verifies beyond the magic:
e_type must be a known value (ET_REL/ET_EXEC/ET_DYN), at least one
of e_shoff/e_phoff must be non-zero, and at least one of
e_shnum/e_phnum must be non-zero. A file with valid magic but zeroed
structure reports partial coverage with the existing ELF-magic
limitation.
* fix(flutter): validate ELF and snapshot coverage
Share bounded ELF header parsing between identification and inspection. Accumulate headers across ZIP chunks and validate table entry widths and extents before claiming complete identification coverage.
Bind AOT hashes to the exported VM and isolate headers, validate stored snapshot lengths including the magic bytes, and preserve partial evidence for missing, malformed, aliased, or overlapping data headers.
Preserve native stream cancellations and APK resource constraints across both operations. Bound snapshot copies while streaming and retain useful input-failure diagnostics and lower-bound count wording.
Validation: 72 focused tests; check:fast; check including formatting; docs:check; test discovery; jscpd; real cone APK and non-Flutter control on Linux. Verified CLI/MCP parity for valid and corrupted ELF headers and AOT inspection across three ABIs. Native Windows and macOS were not run.
* fix(flutter): unify artifact ownership and bound ELF evidence
* fix(flutter): preserve ELF provenance and partial coverage
Read GNU build IDs only from bounded ELF note ranges and retain conflicting candidates. Preserve source versus owned storage failures and honor cancellation through parsing and final cleanup.
Separate snapshot structure from string-pool coverage. Verify every Gallery ABI with digest-bound engine labels and URI counts, and document the supported snapshot-kind layout.
---------
Co-authored-by: REA Contributor <contributor@rea.tools>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Add opt-in retained inventory and WASM Evidence views with exact nested output contracts and a small static WASM descriptor. Preserve old complete CLI/MCP workflows, producer choices, original authenticated Evidence and cleanup ownership.
Validated with real public applications, genuine old Evidence migration, 123 focused assertions, 11 synchronized description checks and 42 installed-package WABT cases. Detailed generated experiment JSON stays outside the final tree; concise findings and reproduction inputs remain. Main-existing completion/package CI failures are separately attributed and left unchanged.
Refs #2016
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Distinguish an unresolved goal from an exhausted call graph and preserve provider failures through cancellation. Retain completed IDA replies and post-operation identity observations as partial Evidence without claiming a verified final binding. Remove source-wide bundler cache-name heuristics and stale expectations.
Use Android 8-compatible batched find/stat with NUL-framed names and raw symlink targets. Preserve malformed or invalid UTF-8 records as partial evidence with exact source bytes. List hidden direct children without synthetic navigation entries.
Share process byte capture with UTF-8 callers and screenshots, removing duplicate timeout and cancellation supervision. Verified 35 focused process, ADB, CLI and MCP tests; compatibility checked against AOSP Oreo sources. Live Android device/emulator verification was not run.
* feat(go): inspect embedded binary build metadata
* fix(go): preserve byte metadata and validate container mappings
* fix(go): support moduleless legacy build metadata
Recognize optional module string headers in declared zero-filled memory while rejecting ambiguous and unmapped pointers. Preserve the original module-pointer field as the zero-byte source anchor.
Classify initial symlink-loop and overlong-path failures as invalid input, with CLI and MCP regressions.
* fix(go): reject undersized PE optional headers
* fix(native): stabilize ELF memory checks and clean up exec wrappers
Use existing parent/group/run-token ownership for caller-selected Python launchers, which can exec through a shell or prlimit. Add cancellation regressions for the valid wrapper and wrong-parent/token cases.
Replace the timing-dependent 30,000-section memory workload with a valid 32 MiB padded ELF under a 64 MiB address-space limit. Check the same image successfully through CLI and MCP before requiring a typed memory failure, preserving the production timeout and all error assertions.
Validation: npm run check and check:fast passed; 61 focused tests passed including Go CLI/MCP and session snapshots. Real decoder probes with pinned packages passed on Python 3.12.3 and 3.12.14. The full local ELF verifier is blocked by this container's procps/process-ownership boundary; the exact pushed head requires GitHub's real-binary-layout lane.
* test: repair Electron fixtures and stabilize catalog startup checks
Declare Electron imports in option, preload override, and path-resolution fixtures after upstream began requiring proven API provenance. Retain all option and preload assertions and adjust exact source ranges for the added import line.
Use the existing bounded acceptance timeout for the cold catalog lifecycle test instead of a separate 15-second ceiling. This retains the coverage allowance and all catalog, error, and cleanup assertions without retries.
Validation: npm run check, final check:fast, formatting and diff checks passed. Source path regression and Go boundary/MCP/service checks passed. All 20 Go CLI cases and both supported-host runtime cases passed on rerun; two macOS-only cases skipped on Linux. A prior contended local run timed out once in a Go CLI diagnostic; that case passed on rerun. Electron process-boundary verification is blocked locally by procps namespace inspection and requires fresh CI.
* test(mcp): isolate nested schema references between tool contracts
Clear Ajv's non-meta schema registry after each root, including compilation failures. Disabling root registration alone leaves nested IDs available to later tools and can let a dangling reference pass the catalog gate.
Add regressions for leakage after successful and failed compilation while retaining shared meta-schema validators. The regression failed before the fix; all 11 schema cases and full npm run check pass. Reviewed independently.
* test(ci): align Go effects and source-map verification contracts
Audit the Go inspector wording as static metadata inspection. Assert retained source declaration indexes in BOM reconstruction and the compiler-derived nested-map oracle. Keep a 45,000-map admission control below the 32 MiB output budget and the 50,000-map overflow case above it with the expanded representation.
Validation: reproduced all four failing assertions, then passed 120 focused tests and the final admission suite. npm run check and real Chrome 154 / esbuild 0.25.10 source-map verification passed, including all four nested-map scenarios and verified process lineage.
---------
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* feat(mcp): derive an API surface view from retained web network captures
* fix(mcp): fail closed on sensitive parents and budget api-surface path expansion
* fix(mcp): keep recorded api-surface paths and budget only templated segments
* fix(mcp): read api-surface paths only from canonical URLs and charge new values once
* fix(mcp): reject authority-less api-surface URLs and key groups by rule positions
* fix(mcp): label api-surface confidence from all groups and cap transient path memory
* fix(mcp): budget api-surface dedup sets and narrow detail outputs
Charge retained deduplication sets before allocation, advertise only
each analysis operation's summary from its detail output, and document
that any templated group makes every API-surface page inferred.
* fix(mcp): bound api-surface path text and record limitations
Charge retained literal path copies before allocation, publish only
valid media-type essences, and deduplicate examined record limitations
onto the view.
* fix(mcp): keep api-surface refusals and path splits inside the reply budget
Sensitive-parent refusals now use the token sensitive_values_unretained so declared literals are not echoed. A later path split is charged against bytes already retained, and the published example includes the HAR record limitation with its matching digest.
* fix(mcp): budget api-surface method text before group keys
Charge method and origin text against the reply budget before the group key is serialized. Document that samples merge only when method, origin, and exact segment rules match.
* fix(mcp): bound api-surface copies and control-character authorities
Reject authorities containing a tab or newline so WHATWG cannot move userinfo into the recorded path. Charge four path copies, two method and origin copies, and each distinct content type against bytes already retained.
* fix(mcp): preflight combined api-surface group charges
Total path, method, origin, and segment charges before retaining a new group. Track content types in a set so distinct MIME essences are not compared quadratically.
* fix(mcp): charge api-surface content-type sets on first use
Create the per-group content-type set when the first essence is retained, and charge it before allocation. Groups with no content type no longer keep an empty set.
* refactor(analysis): centralize API-surface ownership and admission
Own grouping, deduplication, and combined retained/temporary resource admission in one accumulator. Charge fixed group storage before allocation and materialize only the requested page.
Share recorded URL parsing between capture redaction and projection, preserving source paths and excluding ambiguous credential fields. Cover resource admission invariants and the compiled HAR-to-view workflow.
* fix(analysis): admit capture schema clone before parsing
* fix(javascript): share Electron mutation proof with wrapped calls
* test(mcp): align graph and catalog checks with current contracts
* fix(analysis): admit and validate portable capture views
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Return deduplicated execution caveats once and preserve provider-owned relationship diagnostics and completeness on graph nodes. Keep complete relationship reads exhaustive despite contextual caveats, including portable call-path reconstruction.
Preserve omitted diagnostic fields in historical Evidence and bind reported diagnostics into semantic identity. Revise graph workflow and Ghidra adapter profiles so old cached representations cannot replay under the new contracts.
Remove stale duplicate-lookup test claims and redundant fixture validation. Prune million-entry duplicate cases while retaining argument-stack and deep-chain regressions; bounded Flameox captures identified validation and canonical hashing costs.
Validation: static and architecture checks, generated docs, 364 focused tests before pruning, 21 call-path tests after pruning, public MCP schema validation and targeted profile tests. Real engine behavior was not locally exercised.
BREAKING CHANGE: get_call_graph normalized_result now contains { graph, limitations }; traversal-depth groups are under graph.
Keep FAT container format distinct from selected slice architecture through target admission, Evidence, snapshots, and CLI/MCP lifecycle results. Preserve native, Hopper, Ghidra, extraction, and dylib routing for the Mach-O family.
Reject selected FAT ranges overlapping their table or too short for a Mach-O header. Replace empty-slice fixtures with in-bounds executable headers and cover thin, universal, single-slice, FAT64, and swapped containers.
Validation: static and architecture checks, generated docs, targeted producer and public MCP tests. Provider routing uses controlled fixtures; no live Hopper, IDA, JADX, or Ghidra analysis was performed for this change.
Select exact declaring types before managed signatures and CIL bodies are decoded, preserving caller choices in CLI/MCP Evidence. Keep unmatched names and selected coverage explicit; comparisons and graphs retain unknowns outside their observed scope. Reuse existing metadata ownership/accounting and advertise unique selections through the actual SDK Schema.
Validated with 229 focused tests across 21 files, 11 managed conformance controls, fullstatic/docs checks and real Linux CLI and fresh MCP inspection of two unchanged DLLs. All 118 selected CIL hashes matched original PE-mapped bytes; result/caller-argument parity and unmatched controls passed. Full artifact/native-boundary metadata is complete; whole-assembly members retain the unchanged 64 MiB resource constraint. No target code executes.
Refs #1878
Co-authored-by: skared <190618464+skoodster@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Ysopwnable <161784452+N0zoM1z0@users.noreply.github.com>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Group repeated session tool availability and provider capability facts, preserve complete named contracts and ordered limitations, and remove duplicate catalog descriptors. Update consumers and migration guidance.
Validated by all 31 CI jobs, focused source checks, and actual Claude Code/Linux Hopper/native Windows Ghidra small-target workflows on the current integration tree.
Refs #723.
BREAKING CHANGE: binary_session no longer includes server_identity.catalog.tools; tool_availability and provider capabilities use grouped facts documented in docs/mcp-contracts.md.
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
Expose the selected Mach-O slice and its container offset, derive that identity once, and map entrypoint and nonempty segment Evidence to container coordinates. Reject relocation sections whose byte length is not divisible by the entry size instead of silently accepting a floored count.
Validation: native parser and selected-architecture boundaries, actual public CLI and stdio MCP on a compiled universal Mach-O, and a pyelftools 0.33 malformed-table probe. Full Linux pwntools bridge verification was unavailable on this host.
Refs #1973
Parse Apktool YAML and XML at the producer boundary, preserving scalar spelling, XML text and partial string coverage. Retain repeated JADX inventory rows with explicit duplicate-name counts and lookup ambiguity. Make the Apktool verifier report cleanup failures.
Validation: parser, provider and JADX fixture boundary tests; typecheck, lint and documentation checks. Real Apktool and JADX engines were not exercised.
Refs #1978
Separate observed relationship coverage from general provider caveats. Authenticate and replay coverage without changing legacy Evidence IDs; expose unknown and capped relationship sets through CLI/MCP graphs and traces.
Validated focused domain, snapshot, IDA/Ghidra boundary, MCP, schema, package and static checks. Provider fixtures establish contract behavior; licensed engine behavior was not reverified.
* test(ida): expect callee edges from the call graph on non-x86 targets
* fix(ida): read callee edges from the call graph on non-x86 targets
* test(ida): pin address order of call graph additions
* docs(ida): describe the call graph source for non-x86 callees
* test(ida): pin the x86 baseline, the call graph reader and its failure modes
* fix(ida): check the call graph root, scope truncation to the root and degrade without callgraph
* test(ida): expect managed PE off the x86 path, possible-truncation wording and a complete ARM64 traversal
* fix(ida): keep managed PE and x86 reads off the call graph path and qualify truncation
* test(ida): answer only expected leaf roots and assert the requested call graph roots
* test(ida): expect a call graph the producer cannot give consistently to fail the callee read
* fix(ida): fail the callee read when the call graph cannot be obtained consistently
* fix(ida): preserve call edge authority and workflow uncertainty
* fix(ida): preserve explicit root call graph caps
---------
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
Analyze exact selected sources in first-request order before unrelated
static scopes, preserving full artifact identity and byte verification.
Retain selection in CLI/MCP Evidence and summaries, with explicit unknowns
and partial coverage for unselected sources. Keep default full analysis.
Fixes#1848
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>
MCP annotations derived destructiveHint only from mayDiscardData, so tools
that run the analysis target or a caller-selected executable
(capture_process_scenario, observe_native_calls) and tools that change an
attached Android device (install/uninstall/push/start/stop) advertised
destructiveHint: false.
Add explicit executesForeignCode and mutatesDevice effect flags, audit them
across all 170 tool effect records, and derive
destructiveHint = mayDiscardData || executesForeignCode || mutatesDevice.
readOnlyHint semantics are unchanged.
Hopper/Ghidra metadata writers (set_address_name, set_addresses_names,
set_bookmark, set_comment, set_inline_comment, annotate_native_function)
can replace or clear existing analyst names and comments, so they now
declare mayDiscardData as well.
A contract test pins the destructive set and requires every tool whose
description mentions running, launching, executing, starting, installing or
uninstalling to declare one of the new flags or carry a reasoned exception,
so wording and effects cannot silently drift. The binary_session effects
output schema carries the new flags.
* feat(native): admit raw binaries with caller-supplied architecture and entry
* fix(native): honor raw-binary preparation and entry identity
* fix(ghidra): stream raw snapshot attestation with cancellation
* fix(native): preserve and validate raw lifecycle entries
---------
Co-authored-by: morluto <76467478+morluto@users.noreply.github.com>
* fix(managed): report ExportedType rows in managed member inspection
Facade assemblies such as System.Runtime.dll provide their types only
through ExportedType rows (type forwarders), which member inspection
never read, so they reported complete coverage with no types. Report
each row as declared with its flags, forwarder bit, TypeDefId hint, and
implementation token, and record invalid or null implementation coded
indexes as partial coverage. Comparisons now state that exported types
are not compared when either side has them.
Fixes#1812
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): read earlier member Evidence without exported types
Member Evidence recorded before ExportedType rows were read omits the
field. Parse it as null (not examined) instead of rejecting it, and
state in comparisons that such an observation's exported types are
unknown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(managed): leave exported types unknown when metadata is unavailable
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
