Skip to content

Netcap v0.16.11 — evaluator rerun

Choose a tag to compare

@dreadl0ck dreadl0ck released this 18 Nov 13:41
· 349 commits to master since this release

W87 evaluator rerun

v0.16.11 is the build for the full evaluator benchmark rerun.

Area Change
Capture and extraction Stream disk-backed HTTP artifacts and preserve independently completed output under local reassembly, decoder, or analyzer pressure. Report incomplete coverage explicitly.
Reassembly Avoid disjoint overlap walks, preserve TCP rollover behavior, reduce EOF sort allocations, and validate disk-backed connection indexes and identity.
Analysis Snapshot admitted rules, stabilize capture-time replay, and coordinate post-capture completion, cancellation, and live admission.
MCP Validate page counts, rows, generations and continuation; bound inventory scans and concurrent work; report schema reload failures and preserve the active registry on a failed reload.
Detection evidence Preserve TCP overlap bytes and severity. Qualify possible stale probes only with complete evidence; exclude truncated segment ends and FINs from keep-alive exceptions where their sequence evidence is incomplete.

The qualification ledger records
the synthetic 15,000-flow audit equivalence checks and official MCP client runs.
The original evaluator corpus and harness, the ≥64 GB tier, and native
Windows/Npcap live and offline capture remain unqualified. Keep capture options,
input hashes, platform, and artifact selection with the evaluator results.

MCP reliability and bounded capture state

v0.16.9 implements all 17 MCP findings from the v0.16.8 recheck and adds
decoder, capture-state and link-reconstruction fixes. Source and synthetic
qualification are recorded in
internal/collector/testdata/recheck-v0166.json.

Area Change
MCP identity and isolation Reuse the running CLI and bundled loader for capture helpers; allocate opaque upload IDs and separate fixed-root outputs; bind secret/device inventories to the requested capture
MCP input and lifecycle Stream bounded uploads to private files; reject special-file sources; publish completion after rules, finalized alerts and timeline indexing; return explicit failed/not-ready errors; delete only terminal owned uploads and scoped artifacts
MCP data access Add byte-bounded inventory pages, search, matched totals, generation checks, structured field/JSON Pointer projection and complete oversized JSON artifacts. Count-only queries scan all matches without transferring record bodies
MCP protocol and transport Publish output schemas and bounded structured results; correct read-only hints, resource URI encoding, PCAP download framing, bearer-scheme case handling, CORS grants and beyond-EOF stream pages. Order stream evidence by capture identity
Protobuf / HTTP Require multi-field protocol evidence for automatic Protobuf identification while preserving configured single-field decoding. Retain one qualified HTTP pair across malformed presentation metadata, including the supplied non-UTF8/NUL witness
TLS adapter Discover the installed tshark follower namespace: modern tls.stream, or legacy tcp.stream after an unsupported-field diagnostic. Preserve explicit zero and reject ambiguous decryptable selections
Bounded state Back TCP control/reassembly payloads, normalized Software observations, vulnerability/exploit deduplication and optional network-layer payloads with private bounded storage. Bound user-agent caches and materialization; backing/output failures prevent readiness
Spill performance Reuse prepared SQL statements, batch frozen UDP walks and stable packet exports, and cache bounded classifier bytes. Latest matched 10,000-flow backed median improves from 3.215 s to 2.811 s with identical audit counts/digests
Link reconstruction Add Darwin USB and bounded LinkFrame paths for AX.25, MS/TP, Bluetooth H4, I2C, MPEG-TS, Upper PDU, ALP and MDB. Reconstruct supported ALP/MS/TP/6LoBAC/ROHC payloads and retain source ranges and transformations
Upper profiles Add BACnet NPDU/APDU metadata, MPEG-TS program/PES and H.264/H.265/AAC headers, Bluetooth ACL/L2CAP/ATT/SMP metadata, USB descriptor correlation and explicit -link-profiles device mappings. Exact supported variants remain listed in docs/protocol-support.md
Windows packaging Sign production Windows CLI executables before archive/checksum generation; snapshot builds remain unsigned
Qualification Result or remaining gap
Local checks Focused package race/vet checks, native DPI comparisons, supplied-capture collector integration, fresh MCP stdio and Linux bundled-loader tests pass; the release qualification asset records packaged-runtime results
TCP memory witness A matched 200,000-SYN synthetic run reduced retained heap to 44,090,648 bytes and peak RSS to 280,322,048 bytes. A separate 32-MiB TCP payload fixture retained 82,128 additional heap bytes; these are synthetic measurements
DS-12/DS-14 / H-28 / V168-3 Still partial: the latest 10,000-flow backed run is 2.17× memory-only time, above the proposed 2× target. Encoded-cache limits are not hard RSS limits; decoded records and optional accumulators retain separate budgets
Original witnesses The original large-capture/GeekLounge corpus, 49 estimator-skipped captures and H-17's three residual frames remain unavailable. No new evaluator-wide MTA score is claimed
TLS and exotic profiles Real tshark 4.6.6 TLS 1.2/1.3 selection/decryption passes; legacy 4.4 namespace behavior has source/adapter coverage, not an installed-4.4 replay. Advanced/vendor-specific link profiles and independent real inventory still require qualification
Windows Signing/cross-compilation and any packaged startup checks do not establish Windows/Npcap capture-runtime qualification

Retest corrections

v0.16.8 implements the seven findings in netcap-full-recheck-v0.16.7.md.
The issue ledger distinguishes
local regression coverage from the unavailable original evaluator replay.

Area Change
V167-7 Replace per-byte Unicode lowercasing with bounded incremental ASCII prompt matching. Restore CLI harvester port filtering; Telnet prompt evidence still permits off-port recovery
V167-6 Return malformed first-section PCAPNG errors before configuring the reader; 24,125 full-opener fuzz executions passed without a crash
V167-3 Accept four-byte FCS metadata encoded as 4 or 32; unfamiliar optional FCS values no longer discard readable interfaces
V167-1 --quiet no longer authorizes overwrite; noninteractive replacement requires -y
V167-4 / CTF-13 Respect HTTP request-body framing, suppress duplicate banner form extraction when the parsed reader is active, add configured pswrd, and harvest bounded top-level JSON credentials. Repeated requests retain separate occurrences
V167-5 Query logins use actual user/password values. Tokens, API keys and sessions use additive Secret.Kind, Value and Parameter fields, including CSV and Web UI display
V167-2 Ship recursive Markdown documentation without .gitbook recordings or .DS_Store; validate actual archives against a 16 MiB documentation limit
Service evidence Retain ICMP refusal observations that workers process before the quoted UDP attempt. Keep port guesses in PortName; Service.Name uses observed application/probe evidence
TCP and TLS Track FIN sequence space for half-closed keepalives. Omitted TLS selection requires one decryptable stream; explicit zero remains explicit and ambiguous captures require selection. Community IDs resolve TLS follow indexes
Wireless and DNS Protected management bodies remain opaque. Add readable CNAME/NS/PTR companions while preserving raw bytes
D2-13 Add bounded PPI/DLT 192 decoding with nested Ethernet/raw IP, aligned metadata and wireless FCS provenance
Memory Spill Connection, Host, DeviceProfile, Service and retained UDP state through a shared bounded encoded cache with private SQLite backing. Stream UDP datagrams and directional evidence spans instead of retaining their complete payload/span sets
MCP Expand the local tool registry from 56 to 119 tools; analysis settings, live capture, protocol experiments, packet/artifact evidence and startup-bound operator workflows have backend/parity coverage
Qualification Result or remaining gap
Local checks Full portable Go suite, focused race/vet, native DPI/YARA, CLI/MCP E2E, frontend build and TLS 1.2/1.3 selection tests passed
Matched synthetic memory Linux/amd64, 2 CPUs, 200,000 SYN flows: enabling a 1 MiB encoded cache reduced retained heap from 701,117,288 to 422,654,640 bytes and peak RSS from 2,274,717,696 to 1,171,546,112 bytes; runtime increased from 10.61 to 28.25 seconds. This compares this revision's memory and spill modes, not the unavailable evaluator captures
DS-12/DS-14 / H-28 Still partial: TCP reassembly/factory state, optional accumulators and decoded output records remain outside the encoded-cache limit. Spilling is not a hard process-memory bound; the 49 estimator-skipped captures still need external qualification
D2-13 Exotic link types including 3/165/201/209/243/252/266/289/300 remain unsupported; PPI support does not imply support for every possible encapsulated DLT
Original witnesses The evaluator's MTA/CTF/Wireshark corpus, repro/v167, 146 crash inputs and residual trigger frames were unavailable. Local synthetic tests do not establish a new MTA score or closure of every reported residual
Windows Cross-compilation only; no Windows host with Npcap was available for runtime qualification

Gopacket dependency

Area Change
Packet decoding Upgrade from upstream v1.7.2 to v1.7.7, retaining the macOS dead-handle Close fix through dreadl0ck/gopacket@61002f918d93 (upstream PR #178)
Copied TCP reassembly Port sequence-wrap, retained-page accounting, cross-assembler page ownership, oldest-queued expiry, stale control-segment and queued-FIN fixes into internal/reassembly; retain netcap's post-close evidence observation and generation-safe pool lifecycle

Capture and local API corrections

v0.16.7 addresses the 53-row v0.16.6 issue report. The
issue ledger records 45 implemented,
3 implemented with explicit policy, 2 partial, 1 open and 2 by-design dispositions.
These describe source changes and local regressions; the evaluator's complete
MTA/CTF/Wireshark corpus and /workspace/hunt/repro/ were unavailable locally.

Area Change
Local HTTP and MCP Offline/live UI is opt-in; per-run authentication, strict Host/Origin checks and same-origin CORS protect the local API. Browser bootstrap uses an HttpOnly SameSite cookie; MCP authenticates its backend requests. Service-mode authentication behavior is unchanged
Output and CLI Finalize gzip CSV through EOF; stage owned outputs and restore them after failed initialization while preserving unrelated files. Preserve leaf-error exit codes, validate decoder selection and operands, parse decimal timestamps without float rounding, and explicitly reject obsolete generation flags
Capture formats Read patched/Kuznetzov PCAP and normalize timestamp carry. Preserve missing timestamps in PCAPNG SPBs; --pcapng-merge-sections explicitly enables compatible-interface continuity. Add USB link types 189, 220 and 249; identify unsupported 5View and damaged PCAPNG inputs
Packet evidence Normalize bounded GTP-U, Teredo and L2TPv2 inner traffic; retain tunnel scope. Persist fragment diagnostics and honor the defrag switch. Correct portless Community IDs, IGMPv1/v2 and capture-aware wireless FCS handling; export SSID, ordered IEs and WEP metadata
Protocol recovery Preserve long/non-UTF8 HTTP headers, bounded brotli artifacts, compressed Zabbix, KRB-SAFE, FTP navigation and directional Telnet credentials. Add attributed HTTP form aliases and bounded readable TLS HTTP/2 output. Tighten off-port Modbus recognition
DNS, identity and detections Retain malformed/raw DNS and exact name/TXT bytes; detect binary NULL tunnels. Correct reply-first UDP roles and ICMP-refused service claims, avoid keepalive overlap false positives, aggregate per-URI beacon incidents, and keep port guesses separate from observed protocol labels
Enrichment Verify snapshot authenticity independently of freshness; retain expiry diagnostics, digest pins and capture provenance
Performance Indexed service-flow membership reduces the 40k-UDP-flow regression from 3.21 s to 1.09 s on the isolated Linux runner. Lazy TCP payload readers reduce retained heap for 50k SYNs from 207,694,408 to 183,410,096 bytes
Packaging Enable YARA in Linux DPI/nodpi builds, stamp version/commit through Linux build arguments, include documentation trees and validate README-linked Markdown in actual archives
Remaining qualification or policy Disposition
DS-12/DS-14 Open: bounded aggregate-state spilling and qualification of the 49 large captures remain outstanding
H-28 Partial: lazy TCP readers reduce allocation; general per-flow state and full 200k-flow RSS qualification remain
D2-13 Partial: USB 189/220/249 are supported; exotic families including 165/192/209/289/300 remain unsupported
H-5 Conflicting fragment sets remain rejected; first-arrival reconstruction is not implemented
H-11 PCAPNG sections remain isolated unless --pcapng-merge-sections is selected
DS-2 / NEW-2 Preserve qualified aliases and LDAP directory matches without inventing IP ownership
External replay Full evaluator replay, including the original TLS capture/key log and large-capture corpus, remains required

Recheck fixes

v0.16.6 fixes the open items from the v0.16.4 full recheck that do not need a
large machine. DS-12 and DS-14 (memory on the 49 estimator-skipped captures) are not addressed.

Area Change
policy.http-on-443 Fires only for an answered request with a well-formed HTTP/1.x or HTTP/2.0 version and no reassembly qualification. Unanswered probes, version-less request lines and records qualified by capture loss no longer alert. The h2c preface (PRI * HTTP/2.0) is no longer also reported as policy.non-tls-on-443
Overwrite prompt With a non-empty output directory and no -y, a non-terminal stdin now fails with pass -y to overwrite and keeps the previous output. It no longer prints Overwrite? [Y/n]: and exits on EOF
NBNS Node-status (type 0x21) responses are decoded. They previously failed in the DNS parser and were dropped, so captures with only NBSTAT traffic had no NBNS.ncap.gz
Link types Adds DLT 108 (OpenBSD loopback), 50 (PPP in HDLC), 117 (PFLog) and 119 (Prism). PPP frames with the 0xff03 prefix, including DLT 9, now reach the transport decoders. DLT labels in capture-input.json come from the tcpdump registry; several ranges were off by one (289 and 300 printed no or wrong names)
Host identity Hosts.json hosts gain directoryMatches, the LDAP computer objects whose name equals a host's observed name. No IP is bound from LDAP. A bare name and its FQDN (web01, web01.corp.example) are one name instead of conflicting-observations; two different FQDNs still conflict
c2.beacon The TCP-timing and HTTP-timing detectors still emit separate rows. In offline runs each row now carries corroborated-by:<RecordType> when the other detector reported the same source, destination and port
Tests only V-8: a rejected fragment set keeps its fragment-rejected Connection with no ports, by design. V-7: a malformed 802.11 information element is a decode error and leaves the input complete; snaplen truncation is what makes it partial

Not changed: unsupported link types such as 165, 209, 249, 289 and 300 still end as unsupported.
The missing 10.3.14.2 fact on 2017-02-11 is probably an ICMP-quoted or ARP-target address,
which netcap deliberately does not report as a peer; the capture was not available to confirm it.

Analyst web UI

v0.16.5 ships the shared web UI @dreadl0ck/netcap-ui 0.13.3 on the v0.16.4
analysis engine. Capture, decoder and audit-record behavior is unchanged.

Area Change
Charts Categorical charts show the 20 most frequent values, with ties broken by name. /api/chart/data answers 204 No Content when no plottable values exist in scope and 200 with an empty series for format=json; read and configuration failures stay 500
Dashboards Charts without data in the selected scope are hidden and the grid closes the gap. The grid follows its container width, keeps rows aligned and stacks below 768 px; pie legends scroll and axis labels truncate
Connections and services Behavioral context is collapsed by default and lists observations as a table. Directional byte bars and an aligned packet table replace stacked statistics; service details show endpoint, traffic split, identity and an expandable banner
Offline rules Pack, validation and evaluation are three steps with status. Limits are shown in readable units; results summarise alerts, records, matched rules and elapsed time
Copy Removes repeated disclaimers, the HTTP reassembly completeness note and raw nanosecond timestamps; actionable statuses remain

Capture evidence and resource fixes

v0.16.4 includes the capture-evidence and resource fixes following v0.16.3.

Area Change
Raw IP Accept historical on-disk link value 12 as well as LINKTYPE_RAW 101; retain original capture link values during filtered evidence export
HTTP and identity Export bounded multipart fields (1 MiB total, 1,024 parts, no temporary files); retain LDAP computer names as directory subjects without inventing an IP binding; distinguish a unique scoped hostname claim from untimed aliases
ICMP Export quoted IPv4 source/destination/protocol in separate fields, including JSON/protobuf/CSV; quotations do not create observed peers
443 policy Avoid duplicate generic observations for recognizable HTTP requests and exclude client-source-port-only matches
Memory Size out-of-order payload buffers to captured bytes instead of reserving 1,900 bytes for every small segment; release burst allocations and retain at most 1,024 free pages per unbudgeted assembler; avoid oversized empty/small HTTP body allocations and repeated request-body copies
Scan performance Index large host port sets and classify retained TCP prefixes only when their bytes grow, preserving detector match counts

The 148-capture MTA replay retains all baseline carved-file hashes and has no
missing tshark HTTP requests. Both Raw-IP injection witnesses decode; the FKIE
DSB witness retains all 10 tshark-visible TLS HTTP requests. CTU-13 scenario 3
completes 495,056 packets with no SSH role-swap alerts; the ACM forwarder produces
two DNS tunnel incidents with destination port 53.

With GOMEMLIMIT=4GiB, the full 1.7 GB ACM capture completes 4,156,081 packets
at 5.48 GB peak RSS on macOS, replacing the earlier 11 GB OOM on a smaller slice.
GeekLounge 151022 retains all 2,274,747 packets at 5.75 GB peak RSS under the same
soft limit. With GOMEMLIMIT=8GiB, its peak is 8.88 GB; a soft GC target is not
a hard process-memory guarantee. The matched four-vCPU Linux 2018-12-18 benchmark
improves from a v0.15.2 median of 4.34 s to 3.56 s, retaining every baseline artifact
hash.

The historical 5,510-unit inventory includes 415 failed HTML downloads under
sc-html-failed; these are not packet captures. Unsupported formats, damaged
inputs and unavailable archive sources retain separate dispositions.
The qualification receipt is internal/collector/testdata/recheck-v0163.json.

Recheck reliability and portable runtimes

v0.16.3 includes the full recheck-fix branch following v0.16.2.

Area Change
TCP and output finalization Fix queued-overlap page ownership, preserve readable audits on cleanup errors, remove drain-only reader goroutines, and close independent writers concurrently
Linux archives and memory Relocate absolute ELF dependencies, resolve launcher symlinks, avoid unused host/connection maps, and apply cgroup-aware Go GC headroom while respecting explicit GOMEMLIMIT
Protocol recovery Restore HTTP/SMTP artifacts, support larger FTP transfers, preserve responder-originated HTTP/TLS and post-close commands, and recover additional SMB-carried DCE/RPC
Identity and links Add NBNS and IEEE 802.15.4 audit records, retain scoped hostname assertions, reject domain/PTR hostname pollution, and handle SLL2 RARP
TLS secrets Decode supported authenticated TLS 1.2/1.3 HTTP from embedded pcapng NSS key logs; mark decrypted records explicitly
Detection Retain conflicting-overlap evidence after FIN/RST, aggregate beacon incidents, detect skipped POST polling, cover additional service-control/Netlogon operations, and reduce Redis/SMTP, SSH-role and Apache CVE false positives
Input and replay Identify captures by content, distinguish partial/unsupported exits, report effective gopacket replacements, and add checksum-pinned resumable corpus replay

Pre-release checks completed 143/143 MTA captures with no missing tshark HTTP
requests or baseline artifact hashes, 67/67 FKIE captures, 26/26 PCAP-ATTACK
captures and 14/14 injection captures. The default IoT-23 1-1 replay completes
under a 4 GiB Linux container limit. The full 4,695-file Phase B inventory and
original fact-score corpus remain unverified; aggregate MTA capture time remains
18.2% above v0.15.2. Follow-up testing is planned.

Full changes: v0.16.2...v0.16.3

Stream cleanup and failed-capture diagnostics

v0.16.2 fixes two regressions found during the v0.16 try/desktop rollout.

Area Fix
TCP reassembly Closing a direction at its stream byte limit now releases queued KeepFrom pages after scatter/gather cleanup; the Brazil replay no longer triggers the assembler panic
Failed captures Admitted source identity and bounded failure metadata remain exportable when processing fails before a capture manifest exists; stale manifests and replaced source aliases are rejected
Desktop downloads Desktop hosts can save reproduction metadata through native dialogs and their local-output mapping instead of fetching a missing capture route
Native attribution Global Ubuntu notices now match the qualified libtrace 4.0.34-1 and wandio 4.2.7-1 suppliers; per-archive inventories and source hashes remain authoritative

Malformed captures remain partial/error with their observed packet/link/decode
counts. A downloadable diagnostic does not turn a failed capture into a complete
analysis or establish evidence the capture did not contain.

Archives include the seven CLI variants, standalone Linux amd64/arm64 firewall
profiles, exact supplier sources, checksums and qualification receipts. Both
macOS CLI variants are Developer ID signed and notarized. Hosted Windows and
macOS smoke checks do not claim Windows Npcap or full clean-macOS-15 BPF/YARA
certification.

Full changes: v0.16.1...v0.16.2

Capture analysis and application firewall

v0.16.1 includes the qualified changes since v0.15.4. The net archives
include net firewall; standalone firewall binaries and deployment examples
are distributed separately with this release.

v0.16.0 remains a qualification prerelease. v0.16.1 restores the tracked
frontend bootstrap required by immutable Go module consumers; module-zip tests
now reject an archive that omits it even when the ZIP itself is valid.
Linux runtime attribution also verifies the bundled YARA-X bytes, version,
source pin and build recipes against the selected builder before packaging.

Area Change
Analyst workflows Source-bound catalogue, occurrence navigation, host identities, software applicability evidence, endpoint/log review, reproduction downloads and explicit chart/artifact completeness status in CLI, WebUI and MCP
Capture inspection Bounded PCAP/PCAPNG inspection and net split packet/time subsets and time shards; time-only exports work without a native BPF runtime
Rules and delivery Capture-time scoped ordered rules, credential occurrence context, readiness/finalization checks and a durable alert outbox with CLI delivery controls
Streaming Bounded incremental HTTP/1 transactions, capture-time expiry, shared TCP page limits and live application-detection checkpoints
Protocol recovery Correct BGP directions, bounded Syslog framing, CIP context, RADIUS reachability, SMB compounds, legacy PCAP versions and retained round-2 transport/fragment/VoIP qualifications
Application firewall Routed HTTP/HTTPS and API inspection, scoped identities/approvals, Coraza/HTTP3 build profile, SMB/FTP gateways, typed response actions, transactional control/recovery, embedded policy console and Helm deployment examples
Firewall qualification Owned multi-protocol soak, image/chart contracts and digest-bound candidate/review tooling; inclusion does not assert an independently approved production appliance or promote OCI images
Dependencies Matched YARA-X Go/C API 1.22.0, Alpine 3.24.4 and digest-pinned Linux builders; qualified dependency updates and refreshed Go/Rust/native notices
Gopacket Upstream v1.7.2 replaced by dreadl0ck/gopacket at f2eeabc1ff6e; live-only close wakeups, bounded readers and portable descriptor lifecycle checks; fork attribution is in legal/THIRD_PARTY_NOTICES.txt
Performance Shared single-pass entropy calculation; ARM64 entropy and SIMD text-classification experiments remain bounded by their measured replay coverage
Enrichment Signed optional catalogues remain opt-in; NetAPI popularity filtering defaults on with PSL-aware matching, while no popularity corpus is bundled

Runtime and source delivery

Artifact Scope
Linux amd64 glibc/musl, DPI and no-DPI Bundled native library closure, inventory, supplier notices and exact source links; matched YARA-X 1.22.0
macOS arm64 DPI Minimum OS 26.0; Developer ID signed and notarized; execution qualified on the current macOS 26.6 host
macOS arm64 no-DPI Minimum Mach-O target 15.0; Developer ID signed and notarized; execution qualified on the current macOS 26.6 host, not on a clean macOS 15 machine
Windows amd64 no-DPI Driver-free offline client build; native Npcap interface/live-capture qualification remains unavailable; YARA-X is excluded
Source and provenance Netcap tag source, third-party source archive, macOS supplier source/recipe archive, checksums and qualification receipts accompany the binaries

Keep lib/ beside the macOS executable and the full Linux launcher/runtime
directory intact. CLI notarization uses Apple's online ticket lookup; raw
Mach-O/ZIP files do not support stapling. JA4+ remains source-only and opt-in.

Install signed data with net util -download-dbs, refresh it with
net util -update-dbs, and request optional enrichment with
net util --download-enrichment.

Full changes: v0.15.4...v0.16.1

Earlier module consumption and catalogue integration fixes

v0.15.4 repairs module consumption and optional enrichment integration from
v0.15.3. Install the reviewed optional channel with net util --download-enrichment.

Area Fix
Go modules Eight case-colliding GitBook asset pairs retain both original blobs under distinct filenames; module-zip checks guard release and documentation CI
DHCP CLI and service collector configurations enable option-55 lookups through admitted catalogues; stale legacy JSON remains inactive
Service catalogue API /api/service-signatures and capture initialization use the same verified optional installation
Database summary Device catalogue availability is read from the installation without mutating capture globals; scope: installation distinguishes availability from a capture's loaded state

Optional signed enrichment

v0.15.3 adds an independent, opt-in signed enrichment channel. Core database
download/update commands and their defaults remain compatible with v0.15.2.

net util --download-enrichment
Area Contract
Installation NC_ENRICHMENT_ROOT, default $NC_CONFIG_ROOT/enrichment; separate signed scope, bounded member verification and monotonic sequence checks
Service signatures Four independently authored SSH/HTTP/SMTP/FTP rules; verified installed catalogue is ready, absent/invalid data uses the four builtins with degraded status; no Nmap corpus or coverage-parity claim
Passive web technology Bounded origin-transaction advertisements; WordPress/Drupal builtins and reviewed optional Hugo/Joomla descriptors; no active probes or inferred exact versions from major-only headers
Device catalogues Optional historical FingerBank option-55 candidates under ODbL/DbCL, and IEEE vendor prefixes with the recorded secondary nonassertion acceptance; absent data retains the bounded MAC fallback
CISA KEV Attributed CVE enrichment and source links; federal deadlines are identified as federal deadlines and Unknown ransomware labels are not negative assertions
Attribution and source offer /api/enrichment/attribution, device database summaries, full packaged notices and digest-addressed public /enrichment/signed/v1/ archive/member downloads; complete rebuild inputs in the release source archives
Popularity filtering Explicit external CSV only, off by default; PSL-aware filtered Maltego transforms; no popularity dataset bundled, mirrored or approved
Publisher recovery Authenticate and recount the previous signed SQLite before refresh; missing caches cannot reset the count-drop baseline

The historical resolver ZIP is removed from current source delivery. Historical
Git revisions remain available; removal does not establish retrospective rights
clearance. Device counts and synthetic fixtures do not establish classification
accuracy or parity with retired corpora.

Signed core updates and portable runtimes

v0.15.0 and v0.15.1 were runtime-qualification candidates; v0.15.2 is the
binary release with the completed macOS closure and Alpine supplier notices.

net util -download-dbs installs the signed community release without extra
configuration. net util -update-dbs refreshes an existing installation. The
publisher can now refresh reviewed sources without shipping a new client for
each exact snapshot. v0.14.0 retains its frozen compatibility archive; legacy
v1/v2 endpoints remain retired with HTTP 410.

Area Change
Authentication Embedded Ed25519 root verifies the root-authorized publisher key and signed opaque release payload
Freshness and rollback Expiring updates, future timestamps, lower key epochs/sequences and same-sequence equivocation are rejected; verified installed data remains usable offline
Publication policy Reviewed source origins, transformations, static revision pins, full notices and required filenames; complete nonempty annual NVD feeds, per-year counts and count-drop gates
Community contents NVD schema 2, DB-IP City/ASN, IANA services, UA-parser and restored HASSH from a pinned BSD-licensed source
Installation Cross-process update lock, bounded verification, Windows activation without symlink privileges and preservation of previous data on failure
Operations Dedicated non-root publisher image, separate host signer, daily refresh timer and readiness/freshness monitoring

Private publishers retain explicit NC_DBS_APPROVALS and
NC_DBS_APPROVALS_SHA256 overrides. Retired CMS/DHCP corpora and Exploit-DB
payloads remain excluded from core delivery; optional catalogues use the separate
channel above.

Packaging and dependency fixes

Area Change
Dependencies Upgrade klauspost/compress, Vite, Vitest/mocker, shell-quote and source-map-js to patched versions
Linux WebUI Pass the actual built WebUI as an explicit Docker build context instead of embedding only its placeholder
Runtime provenance Record every bundled loader/library hash, supplying package/image and corresponding-source link; retain package copyright and licence notices
macOS runtime Bundle and relocate native libraries instead of depending on old Homebrew paths; include exact supplier sources and build recipes. DPI requires macOS 26+, nodpi macOS 15+
Source bundle Include the qualified libtrace 4.0.23, wandio 4.2.6 and libwandder 2.0.8 sources; correct the prior runtime-version notices
Geolocation Missing optional GeoLite2 fallback data logs at debug level; explicitly selected or corrupt providers retain actionable errors
Windows On-demand qualification runs the actual client, signed download/update and structured CVE lookup on a Windows runner

Packaged verification

Linux amd64 builds report the release version and source revision through the bundled loader. Both macOS arm64 builds are Developer ID signed and passed runtime smoke tests; notarization was not verified. Windows amd64 Authenticode signing and timestamp verification passed; native Npcap capture remains unqualified. See qualification.json and checksums.txt.