netsleuth/ARCHITECTURE.md

Architecture

Русский

netsleuth collects network facts from many independent, mostly keyless sources, merges them into typed structures, interprets them with pure functions, and renders two artifacts per run. Collection, interpretation and rendering never mix: every module either gathers data or reasons about already-gathered data, never both.

Directory layout

src/netsleuth/
  cli.py          Typer app: flags, phase orchestration, Rich progress. No business logic.
  config.py       pydantic-settings loader. Precedence: env > .env > config.yaml > default.
  models.py       Every shared dataclass, plus to_jsonable() for serialization.
  orchestration.py  run_module(): timing, timeout, exception -> ProbeError classification.
  netinfo.py      Capability detection and local interface/gateway/MTU/resolver facts.
  iface.py        --interface resolution: adapter name/IP -> BindTarget, pure and tested.
  ip_geo.py       Six-provider identity chain, per-provider normalizers, field-wise merge.
  bgp.py          RIPEstat, CAIDA ASRank, Team Cymru DNS, PeeringDB (disk-cached).
  reputation.py   FireHOL netset cache + prefix lookup, Shodan InternetDB, DNSBL decode.
  stats.py        Pure RTT statistics: loss, min/avg/max, mdev, jitter.
  traceparse.py   Pure text -> TraceHop parsers for Windows, Linux and BSD/macOS output.
  interpret.py    Pure verdict engine: thresholds -> Finding[], VPN scoring, bufferbloat grade.
  speed.py        Cascading speedtest, throughput math, cfL4 header parsing, bufferbloat probe.
  exporter.py     JSON and Markdown rendering, atomic writes into ./logs/. Which
                  artifacts get written is decided by the caller (cli.py), not here.
  compare.py      --compare: diff two saved JSON reports; render_diff_brief() feeds the automatic vs-previous-run summary.
  history.py      Same-ASN report lookup by filename convention: report_key(), find_previous(), latest_key().
  trend.py        --trend N: aggregates N saved JSON reports into per-label sparkline series.
  metrics.py      collect_metrics() extracts a flat Metric list from a report; render_prometheus()/render_csv() serialize it.
  watch.py        --watch: periodic re-run loop with a live Rich dashboard.
  probes/
    latency.py         Ping fan-out and jitter/loss/mdev/p95/p99 statistics.
    traceroute.py      Cascade: mtr --json -> icmp_win -> icmplib -> system binary text, plus the opt-in tcp_trace tier.
    icmp_win.py        ctypes IcmpSendEcho2 / Icmp6SendEcho2 engine (Windows only).
    dns_leak.py        Per-adapter resolver enumeration, echo probes, ECS detection.
    tls_rtt.py         --tls: TCP RTT + TLS 1.3 handshake + TTFB via two sequential connections.
    dns_advanced.py    --dns-advanced: system DNS vs DoH, transparent-proxy probe on a bogon resolver IP.
    bgp_path.py        --path-diversity: CF-RAY edge colo lookup, international-routing-loop detection.
    prefix_benchmark.py --prefix-bench: ping the first host of a capped sample of the AS's announced prefixes.
    dpi_check.py       --my-server: single-host TCP RST/filtering self-diagnostic. Not a scanner — see Key decisions.
    nat64.py           One RFC 7050 AAAA query (ipv4only.arpa) to detect NAT64/464XLAT translation.
    hop_asn.py         Per-hop ASN/AS-name/country/PTR enrichment via Team Cymru DNS, in-run deduped.
    captive_portal.py  Phase-0 gate: three generate_204-style probes for a public-wifi login wall.

File-size exceptions. The ~200-line guideline (Global Constraints) held for most modules, but these grew past it during implementation and were kept as single files rather than split, because each one is a single cohesive responsibility that splitting would only scatter across more files with a thinner reason to draw the line anywhere in particular:

• exporter.py (~910 lines) — JSON and both Markdown languages are renderers of the same report shape; keeping them together is what guarantees the English and Russian reports can never drift on which sections or rows exist. render_markdown(report, lang="en"|"ru") and a handful of inline _t(lang, en, ru) calls at each string site render both languages from one code path, deliberately, instead of a second renderer file that could silently fall out of sync. This grew by five sections (TLS, DNS advanced, path diversity, prefix bench, DPI check) without changing that argument.
• cli.py (~805 lines) — the Typer app, every flag, and the full diagnose() phase pipeline; splitting flag parsing from what the flags toggle would just add an import hop.
• models.py (~490 lines) — every shared dataclass plus to_jsonable(); this is deliberately the one place that defines the report's vocabulary.
• interpret.py (~600 lines) — threshold bands, the latency/path/speed/TLS/DNS/DPI/anycast/prefix finding generators, VPN scoring and bufferbloat grading all reason over the same Finding[] shape.
• bgp.py (~305 lines) — four independent providers (RIPEstat, CAIDA ASRank, Team Cymru, PeeringDB) plus the disk cache they share.
• ip_geo.py (~280 lines) — six independent provider normalizers plus the field-wise merge that reconciles them.
• netinfo.py, speed.py (~260 lines each) — OS-specific capability probing and the four-tier speedtest cascade both fan out into several backends that don't split cleanly by single responsibility.
• reputation.py (~250 lines) — FireHOL netsets, Shodan InternetDB and DNSBL decoding are three unrelated reputation sources sharing one result shape.
• probes/traceroute.py (~225 lines) — the four-tier cascade plus the opt-in tcp_trace tier.

Data flow

                +-- netinfo (local facts, capabilities)
                |
  cli.py -------+-- ip_geo provider chain  ---> egress IP + ASN   [blocking: everything downstream needs the ASN]
                |
                +-- bgp || reputation || dns_leak || dns_advanced || path_diversity   (parallel)
                |
                +-- latency || traceroute || tls fan-out        (parallel, semaphore-bounded)
                |
                +-- prefix_benchmark  (opt-in, sequential — isolated from the latency fan-out above)
                |
                +-- dpi_check         (opt-in, sequential — single-host, rate-limited)
                |
                +-- speed                                 (exclusive phase lock)
                |
                +-- interpret  ---> Finding[] + overall verdict
                |
                +-- exporter   ---> logs/report_<ASN>_<ts>.{md,ru.md,json}

Three hard concurrency constraints:

1. Measurement isolation. The speedtest never overlaps traceroute, ping or API calls — it holds an exclusive phase lock. The single intentional exception is the bufferbloat probe, which overlaps saturation by design, inside speed.py.
2. Bounded traceroute fan-out. Parallel traceroutes to different targets share early hops and inflate each other's latency, so concurrent traces are capped by a semaphore (probing.trace_concurrency, default 2).
3. The AS prefix benchmark never shares a phase with the reference-host latency fan-out. Pinging up to prefix_bench.max_prefixes (capped at 256, default 32) hosts at once would both skew its own results and pollute the latency numbers measured for the fixed reference hosts in the same window, so it runs as its own sequential, opt-in step (--prefix-bench) after phase 3 finishes, never inside --full.

Key decisions

Failure is a value, not control flow. Every module call goes through run_module(), which times it, enforces a per-module timeout, catches everything except CancelledError, and returns a ModuleResult with status in ok | partial | failed | skipped. asyncio.gather(..., return_exceptions=True) is a second net at the orchestration layer. No single provider outage can abort a run, and both output formats show every failure explicitly — a report missing its speed section must never look identical to a report where speed was fine.

Never require elevation. Capability detection is attempt-based, never OS-assumption-based: try unprivileged datagram ICMP, then raw ICMP, then (on Windows) the IcmpSendEcho2 API, then fall back to TCP-connect timing and the Cloudflare Server-Timing: cfL4 header. Worst case, the tool still produces a full report and states the degradation once, plainly, with the one-line remedy.

Windows traceroute uses the Win32 ICMP API, not tracert.exe text. IcmpSendEcho2 from Iphlpapi.dll is the same unprivileged API tracert.exe uses internally. Driving it through ctypes gives per-hop RTT, TTL-expiry detection and real timeout handling with no Administrator, no Npcap and no third-party dependency — strictly better than parsing a localized three-probe text table.

Text parsing is a last resort, and parses by line shape. When the system binary is the only option, traceparse.py matches on structure (leading hop integer, RTT tokens, IP regex) and never on header or footer wording, because Windows tracert output is code-page dependent — cp866 on Russian Windows, not UTF-8. BSD/macOS additionally emits multiple IPs per hop and !H/!N/!X annotations that Linux does not, so the three parsers are genuinely separate.

Latency never parses ping text. That would mean three more locale-dependent parsers for no benefit. Every PingResult is tagged with the method used, because TCP "loss" (connection failure rate) and ICMP loss are different metrics and must never be silently conflated.

Reputation defaults to sources that do not learn what you are checking. FireHOL netsets are downloaded once, cached, and matched locally — no runtime query, no rate limit, no disclosure. Shodan InternetDB tells a home user something actionable (an exposed router, a forgotten port-forward) far more often than a mail blocklist does. Classic DNSBLs are opt-in behind --dnsbl.

DNSBL responses in 127.255.255.0/24 are errors, not listings. Spamhaus returns 127.255.255.254 for "you queried through a public resolver" — which is the case for anyone on 1.1.1.1 or 8.8.8.8 — and 127.255.255.255 for "rate limited". A naive "any 127.x.x.x means listed" implementation red-flags most users. Both are surfaced as result unavailable, never as a finding.

--interface binds where the code structurally can, drops where it structurally can't, never runs unbound and silent. Every httpx-based provider client can bind an exact local address (local_address= on the transport) — cheap and solid. The Ookla binary tier passes --interface/--ip through to the external process. icmp_win uses IcmpSendEcho2Ex (which adds a SourceAddress) instead of IcmpSendEcho2 only when a bind is forced, so the unforced path is byte-for-byte unchanged. The one structural dead end is Windows tracert.exe, which exposes a source-address override for IPv6 only, never IPv4 (-S is documented IPv6-only in its own usage text) — filter_trace_tiers() drops that tier rather than let it silently measure the OS-default path under a report claiming a forced adapter. netinfo.collect_local_net()'s own observation of the OS's real default route is deliberately left untouched by the flag: the report needs to show both the forced adapter and what the OS would have chosen on its own, and overwriting the observation would turn the VPN gateway/tunnel-interface signals into a tautology (force the VPN adapter, get "VPN confirmed" for free).

An interface with no route to the internet is a diagnostic answer, not an error. Resolving an unknown/down/addressless adapter aborts immediately (typer.BadParameter, every adapter listed) — that is bad input. But an adapter that resolves fine and turns out to carry nothing is exactly the kind of thing this tool exists to surface, so the run proceeds, every probe fails into the normal unavailable/timeout classification, and the report says so plainly rather than pretending to be a default run.

Bulk RIPEstat endpoints are always bounded. routing-history and friends return 10 MB+ payloads for large ISPs, so every such call carries max_rows and a timeframe from config.yaml.

PeeringDB is cached to disk. Anonymous limits are 20 requests/minute with large responses throttled to one per hour, so a response is never re-fetched within a run and persists in .cache/ between runs for providers.peeringdb_cache_hours.

The JSON dump carries a raw section. Every provider response is stored verbatim under its source key. The typed layer above it necessarily drops fields; raw is what actually satisfies "100% of collected data, no truncation".

Why the DPI check is not a port scanner. probes/dpi_check.py is, mechanically, code that opens TCP connections to an IP on several ports — the same primitive a port scanner uses. The difference between a self-diagnostic and a scanner is the constraints wrapped around that primitive, and they are load-bearing, not cosmetic:

1. The target is only ever the string the user typed after --my-server. check_dpi(target, resolved_ip, ports, ...) takes both as required positional-or-keyword arguments with no default — there is no code path that fills them in from IpGeo.ip, CfTrace.ip, --target, or BgpIntel.announced_prefixes. tests/test_dpi_check.py pins this via inspect.signature.
2. The port list is fixed and short. dpi_check.ports in config.yaml defaults to six ports and is capped at 16 by a pydantic validator in config.py — there is no --ports/range flag, and raising the cap requires editing and redeploying the config, not passing a CLI argument.
3. One host per run. --my-server takes a single value; CIDR input is rejected outright in cli.py before Options is even built ("this is a single-host self-diagnostic, not a network scanner").
4. Rate-limited by construction. asyncio.Semaphore(dpi_check.concurrency) (capped at 4) plus asyncio.sleep(delay_between_ports_seconds) between probe starts; one attempt per port, no retries.
5. Consent is printed, not just documented. Before probing, cli.py prints which host and how many ports are about to be touched and that this must be a host the user owns or is authorised to test — the same pattern as the existing --tcp-trace warning, not the file-marker consent --ndt7 uses, because there's no third-party data disclosure here, only a question of authorization.
6. The target is audited. meta["dpi_target"] in every JSON report records exactly what was checked.

The AS prefix benchmark (probes/prefix_benchmark.py) gets the same treatment for the same reason at a different scale: pinging the first host of every prefix a large ISP announces is mass probing of a network the user does not necessarily own. prefix_bench.max_prefixes is capped at 256 in a validator (default 32), the probe is ICMP-only (never TCP-connect), and it is never enabled by --full — only by explicit --prefix-bench.

TLS handshake timing uses two sequential connections, not ssl.SSLObject.start_tls() on one. probes/tls_rtt.py opens a plain TCP connection first (tcp_connect_rtt(), already used by latency.py) to get a TCP-only baseline, then a second asyncio.open_connection(..., ssl=...) to the same host and subtracts the baseline from the second connection's total time to get tls_handshake_ms. This slightly overstates the true handshake time (a second TCP handshake happens too) but avoids the complexity of upgrading one live connection mid-flight, and the report says explicitly, in a footnote under the table, that the number is derived by subtraction — not presented as a lab-grade measurement.

dns_advanced.py and dns_leak.py answer different questions and must not merge. dns_leak.py asks "where do my DNS queries physically exit to the internet" (echo probes, resolver-ASN-vs-egress-ASN comparison) and feeds the VPN-confidence scoring. dns_advanced.py asks "how fast and how honest are the answers" (system vs DoH timing and content) and feeds its own findings. They share nothing but dnspython; a shared helper (resolver_for-style) was deliberately not extracted, because the one place they'd overlap (resolver construction) is three lines each and not worth a coupling.

A resolver answering differently from DoH is not, by itself, evidence of anything. Any CDN behind geo-DNS or ECS legitimately hands different resolvers different IPs for the same name. dns_advanced.compare_answers() records the difference as plain text; detect_poisoning() only escalates to a finding when the system resolver's answer is itself bogus (0.0.0.0, loopback, private, or RFC 5737) while DoH's answer is a normal public address — that specific asymmetry, not mere disagreement, is what a poisoned/hijacked response looks like.

The Cloudflare colo→city/country table is a curated ~30-entry dictionary, not a full IATA database. Pulling in a full airport-code package for this one lookup would be a new dependency for a feature that only needs to answer "is the edge PoP in a different country than the client" for a modest, CIS/Europe-weighted set of colos. An unrecognized code renders as "unknown" rather than guessing.

The international-routing-loop check compares the client's own geolocated country against the edge PoP's country — nothing else. An earlier version of detect_international_loop() also required the target's resolved-IP geolocation to match the client's country before it would fire, on the theory that this ruled out an unrelated confound. In practice nothing ever populated that field, so the check could never fire at all — caught by manually running --path-diversity end-to-end and inspecting the JSON output, not by any unit test, since the unit tests exercised the function directly and simply supplied the value the buggy production code path never provided. The lesson generalized: two inputs beat three when the third can't reliably be filled in, and a smoke run of new opt-in flags belongs in the same pass as the unit tests, not after them.

CGNAT detection uses three evidence paths, not one. A user behind a home router sees a private 192.168.x.x address locally while carrier-grade NAT (RFC 6598, 100.64.0.0/10) lives one hop further up — checking only LocalNet.local_ipv4 would miss the common case entirely. cgnat_findings() checks the local address, then the default gateway, then every traceroute hop, in that order, and records which one fired as the finding's evidence string.

NAT64/464XLAT is detected the RFC 7050 way, not by checking the client's own address. The well-known synthesis prefix 64:ff9b::/96 is where a NAT64 resolver embeds an IPv4 destination, never something the client's own egress address falls into — an earlier design that checked the egress IPv6 against that prefix would never fire. The correct probe is a single AAAA query for ipv4only.arpa: its A records are two fixed addresses (192.0.0.170/.171), and a DNS64 resolver synthesizes an AAAA answer embedding one of them in the network's actual (often ISP-specific) /96 prefix; a real dual-stack network returns no AAAA for that name at all.

The first traceroute hop is diagnosed separately from the rest of the path. Loss or latency at hop 1 means the user's own router or Wi-Fi, not the ISP or anything downstream — path_findings() takes an optional LocalNet and, when supplied, compares the first hop's IP against the OS-observed default gateway before running the existing mid-path loss-jump logic, so the two questions ("is my LAN fine" vs "where does the wider path degrade") stay structurally separate rather than one drowning out the other.

Per-hop ASN enrichment is a separate phase after the trace completes, not folded into traceroute itself. TraceHop.asn/as_name existed in the model since the original design but were never populated by any writer — probes/hop_asn.py::enrich_hops() fills them (plus a new country field and a PTR fallback for backends that don't resolve hostnames themselves) by collecting the unique set of hop IPs across every trace and firing one Team Cymru DNS query pair per IP, deduped in-run. This runs as its own run_module("path_enrich", ...) step in cli.py after the latency/path/TLS fan-out and before speed, and mutates the already-collected TraceHop objects in place; a lookup failure folds into that step's warnings only; it can never downgrade the trace's own ModuleResult.status. Team Cymru's DNS service has no rate limit (unlike the cached_json()-wrapped RIPEstat/CAIDA calls), so no disk cache is needed — only in-run dedup of repeated hop IPs.

Captive portal detection tags, never short-circuits. ModuleResult.status is informational everywhere else in the pipeline, and nothing in diagnose() branches on it — introducing the codebase's first status-driven control flow for this one feature would contradict "failure is a value, not control flow" above. A captive portal is also rarely total: DNS and ICMP often pass while only HTTP is intercepted, and a short-circuited run would produce a report indistinguishable from a dead link. So probes/captive_portal.py runs as Phase 0, before everything else, and on detection adds a crit finding plus a banner at the top of the report — every other phase still runs and is honestly labeled as measured from behind the portal, which is more information than refusing to measure at all.

The three default captive-portal check URLs must share one success shape, and picking them is not interchangeable. check_captive_portal() expects every URL to behave like a generate_204 endpoint: HTTP 204 with an empty body on success. Mozilla's detectportal.firefox.com/success.txt was tried as a third default and immediately produced a false positive on a real network — its correct, non-portal response is HTTP 200 with the body success, which is indistinguishable from classify_portal_response()'s "200 with a body" portal signal. Caught by the same live-smoke-test discipline as the international-loop bug above, not by a unit test (the unit tests fed the function literal response shapes and never modeled a real endpoint's actual behavior). Fixed by swapping in a second genuine generate_204 endpoint (clients3.google.com) instead of teaching the classifier a second success shape — one uniform contract across all check URLs is simpler than a per-URL expected-body table for three entries.

Speedtest server geo-sanity compares countries, not coordinates. The original design called for a haversine distance in km between the client and the speedtest server, which needs both ends' lat/lon. Live inspection of the bundled Ookla CLI's own JSON output (speedtest.exe --format=json) showed its server object carries location (a city name) and country, never lat/lon — that's the documented schema, not a version quirk. speed_findings() compares SpeedResult.server_country against the client's own geolocated country instead: coarser than a distance, but real, always available when a tier reports a country, and it still catches the case that matters (the test hit a server in another country entirely).

TLS certificate handshake tries verified first, only degrades to CERT_NONE on a validation failure. probes/tls_rtt.py::tls_context(verify: bool) defaults to a real, unmodified create_default_context() — a change from the tool's original always-CERT_NONE posture. measure_tls() attempts a verified connection first; only on ssl.SSLCertVerificationError does it retry unverified (re-timing the retry, since the failed attempt's duration isn't representative of a real connection). This is the only way to get subject/issuer/notAfter at all: Python's getpeercert() returns {} for an unvalidated chain, populating the parsed dict only when the certificate actually verified — getpeercert(binary_form=True) (the raw DER, hashed into cert_sha256) is the one thing available either way. The fingerprint is the SHA-256 of the full DER certificate, not the SPKI — extracting just the subject-public-key-info needs the cryptography package, a new dependency this feature doesn't need; the report footnote tells a user comparing against an externally-pinned value to compute it the same way (openssl x509 -outform DER | sha256sum).

Pinning is configured, not flagged. tls.pinned_fingerprints (host → expected SHA-256, capped at 32 entries, each normalized and hex-validated by a pydantic validator) follows the same posture as dpi_check.ports: raising the list requires editing and redeploying the config, not a --pin CLI argument on a diagnostic tool that anyone could point at anything.

Prometheus and CSV are one extraction, two serializations — not two renderers. metrics.py::collect_metrics(report) -> list[Metric] does the per-section special-casing a flat walk of the report dict can't (latency needs a label/host/method triple, TLS a label/host pair, path hops a target/ttl/ip/asn quadruple) as a single pure dict -> list[Metric] function; render_prometheus() and render_csv() are then ~15 lines each, formatting the same list two ways. FORMAT_EXTENSIONS (config.py) is still the single source of truth both cli.py::parse_formats() and config.py::Output._normalize_formats() key off of — adding "prom"/"csv" there was the only wiring point.

Same-ASN auto-diff runs before write_report(), or the current run would diff against itself. history.find_previous() globs ./logs/ for report_{key}_*.json using the exact same sanitize_name()/timestamp convention exporter.report_filename() already writes, so the lookup key can never drift from what's actually on disk. It only fires when this run is itself writing JSON (there'd be nothing to diff against on a future run otherwise) and only prints a one-line note — never a warning — when JSON isn't in this run's formats, since silence is correct on every other skip path (first run ever, or a prior run that didn't write JSON). compare.py::render_diff_brief() exists because the full render_diff() table is too long to print unbidden after every run; it surfaces only the ASN/IP change, the single worst latency delta, the download delta, and new/resolved finding counts.

--trend needs no live network call to find "this network's" history. Rather than re-running identity detection just to know which ASN's reports to load, history.latest_key() reads it back off the newest report_*.json filename already on disk — the same file-naming convention already encodes it, and a read-only historical view has no reason to touch the network at all.

Python 3.14 is the floor, not 3.10. The project moved off >=3.10 when this batch of probes was added; none of the new code actually depends on a 3.11+-only construct (ssl.SSLObject.start_tls(), asyncio.timeout(), datetime.UTC) — the two-sequential-connection TLS design above was chosen independently of what the interpreter allows. The version floor moved because there was no longer a reason to hold it back, not because a specific new feature required it.

Storage

• ./logs/report_<ASN>_<YYYYMMDDTHHMMSSZ>.{md,ru.md,json,prom,csv} — the artifacts selected by --format/output.formats (default: English Markdown only; all writes all three: English Markdown, a full Russian translation with identical structure, and the JSON dump — English only, no .ru.json). When both Markdown languages are written they cross-link each other at the top, the same way README.md/README.ru.md do; writing only one skips the cross-link rather than pointing at a file that doesn't exist. <ASN> is the target's subject in target mode (an ASN, IP or domain) and the local egress ASN in auto mode. Compact ISO timestamps because Windows forbids : in filenames. Falls back to report_unknown_… when the ASN lookup fails entirely. Written temp-file-plus-os.replace(), always atomic. Gitignored: reports contain the external IP, city, coordinates and ISP name.
• ./logs/watch_<ASN>_<YYYYMMDDTHHMMSSZ>.json — one time-series artifact per --watch session, not one report per tick.
• ./.cache/firehol/*.netset — downloaded blocklists, refreshed per providers.firehol_refresh_hours.
• ./.cache/pdb-net-<asn>.json, ./.cache/pdb-netixlan-<net_id>.json — PeeringDB responses, valid for providers.peeringdb_cache_hours.
• ./config.yaml — all non-secret settings. ./.env — the three optional API keys and nothing else.

Configuration

config.py builds a pydantic-settings Settings object whose sources are ordered init > environment > .env > config.yaml > field default. The three API keys are SecretStr | None = None at the top level: a diagnostics tool must run with zero configuration, so a missing key downgrades the enrichment that needs it to skipped with a warning and never fails the run.

Five sections back the L4-L7 probes, all opt-in at the CLI level and all with a validator-enforced ceiling where the feature could otherwise become mass probing: tls, dns_advanced, path_diversity, prefix_bench (max_prefixes capped at 256), dpi_check (ports capped at 16 entries, concurrency capped at 4).

Tests

uv run pytest -q. Roughly 807 tests covering the parsers, normalizers, verdict engine, DNSBL decoding, throughput math, ping statistics, jitter percentiles, serialization and the report diff. Glue is deliberately untested: live HTTP, real ICMP, Typer wiring, psutil/ctypes calls and the --watch timing loop. Traceroute fixtures live in tests/fixtures/traceroute/{windows,linux,darwin}/ and include a cp866-encoded Russian Windows sample, because that encoding is exactly what breaks naive parsers. DoH response fixtures (RFC 8484 JSON form) live in tests/fixtures/api/doh_*.json alongside the other provider fixtures.