test: round-trip coverage for every protocol code, dumper and lossless path

Goal
Make every round trip the library promises a tested invariant, so that a mismatch fails CI instead of turning up in review. This follows the owner's ruling that every round trip must pass and match exactly.

Scope
Each new path below sits alongside the registry-driven construct → parse → construct cycle that already exists in tests/protocols/test_option_roundtrip_unit.py.

Capture-driven parse → rebuild: for every layer of every frame in examples/captures/*, Cls.from_data(layer.info).data == layer bytes.
Protocol edge cases: minimum and maximum lengths, reserved bits set, zero-length values, padding variants, unassigned codes.
Dumpers: format='pcap' output is byte-identical to its input, and json/plist/tree dumps are complete and deterministic.
Foundation: reassembly recovers the original payload, traceflow output re-reads to exactly its frames, every offline engine agrees, and registry register/unregister is symmetric.
corekit: Field and Schema pack/unpack are symmetric, and Info ↔ dict round-trips.
const/vendor: each crawler regenerates its const module byte-identically from the offline snapshots.

Every case that still fails is listed with the issue that tracks it, in the same way as EXPECTED_FAILURES. A case not on that list must pass.

Depends on
The ten round-trip audits now running, one per protocol family plus dumpkit, foundation, corekit and const/vendor/CLI. Their defects get filed first, so the known-failure tables can cite real issue numbers.

Blocked until the ten round-trip audits have reported and every defect they found is filed. I will post the issue list here when that is done.

Unblocked. All ten round-trip audits reported on 2026-10-07, between 17:19Z and 19:16Z, and every defect they found is now filed.

Audit	Issues
link	#1204–#1210
transport	#1211–#1223
internet-v6	#1228–#1237
internet-v4	#1312–#1329, #1349
application	#1238–#1252, #1302, #1303, #1333, #1443 (new)
misc + pcapng	#1267–#1286
dumpkit	#1255–#1266
corekit	#1274, #1288–#1296
const / vendor / CLI	#1297–#1300
foundation	#1354–#1364

#1443 was the one gap: unassigned FTP reply codes and HTTP status codes raise EnumValueError, so those replies cannot be parsed. I re-ran it on 78991b7d5 and filed it today.

Three other findings I checked on main are already fixed:

the IPv6 trailer, by fix(ipv4,ipv6)!: keep the octets past Total Length or Payload Length as captured (#1209) #1431;
the MH StringField encoding, by fix(corekit): StringField is not lossless: guessed charset, U+FFFD substitution, and a character-count length #1336, which made StringField lossless;
the PCAP-NG IDB reserved octets, by fix(pcapng): keep captured padding octets on rebuild (#1437) #1441.

Every other issue in the table is closed. This issue is now pending, and #1443 is the only known failure it should list.

A worker is on it now, on branch test/1202-roundtrip-coverage. This is test-only work: any new mismatch it finds will be filed as its own issue.

PR #1445 covers the first two parts of the scope: capture-driven parse→rebuild (20,501 subtests) and the dumper checks (1,725 subtests). It found three new mismatches, now filed as #1446, #1447 and #1448. The PR says Refs, so this issue stays open for the parts still to do: protocol edge cases, foundation, corekit pack/unpack, and const/vendor. That last part first needs a decision on where offline crawler snapshots should live.

test: capture-driven parse→rebuild and dumper round trips (#1202)

A second worker has started on the protocol edge-case slice, on branch test/1202-protocol-edge-cases. It is test-only and its PR will say Refs #1202.

test: capture-driven parse→rebuild and dumper round trips (#1202) (#1445

test(protocols): edge-case round trips for every layer family (#1202)

Recorded: the owner does not want offline crawler snapshots committed, so the const/vendor part will not use them.

That leaves one choice for that part:

Drop it from this issue (recommended). cron-vendor.yml already runs every crawler against the live sources each Saturday and commits any regenerated const module. The per-crawler tests in tests/vendor/ already cover the parsing. A byte-identical check would need either snapshots or the network.
Add a network-gated test. It would regenerate each module from the live sources and compare it with the committed file, but run only when opted in, so ordinary CI never depends on the network (the test(integration): pcapng end-to-end counts fail when many_interfaces.pcapng falls back to the synthesised stand-in #1478 failure mode). It would fail whenever a registry changes upstream, so it would really be a drift detector.

The foundation and corekit slices are already being worked; neither depends on this.

take option 1

Recorded: the owner chose option 1, so the const/vendor part is out of this issue's scope. The owner also ruled that when a crawler fails, its const enum is simply left as it is. cron-vendor.yml already works that way: a target that exits non-zero leaves its const file unchanged, and only the targets that succeeded are committed.

That leaves two parts here, both being worked, and each will be its own Refs #1202 PR:

foundation: reassembly, traceflow re-read, engine agreement, and registry symmetry;
corekit: Field and Schema pack/unpack, and Info ↔ dict.

This issue closes once both have merged and any defect they find has its own issue.

test(corekit): field, schema and Info round trips (#1202)

test(corekit): field, schema and Info round trips (#1202) (#1492)

test(foundation): reassembly, traceflow, engine and registry round tr…

Both parts have merged, so this issue is complete:

corekit: Field, Schema and Info ↔ dict round trips, in test(corekit): field, schema and Info round trips (#1202) #1492.
foundation: reassembly, traceflow re-read, engine agreement and registry symmetry, in test(foundation): reassembly, traceflow, engine and registry round trips (#1202) #1508.

The const/vendor part was dropped from scope by the owner's earlier ruling.

Every defect these PRs found has its own issue, recorded as a known-failure row that a fix PR must delete:

Navigation Menu

Repository navigation

test: round-trip coverage for every protocol code, dumper and lossless path #1202

5 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

test: round-trip coverage for every protocol code, dumper and lossless path #1202

Description

Activity

JarryShaw commented on Oct 7, 2026

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

5 remaining items

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

JarryShaw commented on Oct 9, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions