Correctness and robustness fixes from a deep audit against the C reference - #394

Merged

pavel-kirienko merged 35 commits into

Jul 18, 2026

Commits on Jul 18, 2026

Fix RX-path crash on crafted gossip names; guard CAN reader ingest

pavel-kirienko
and
claude
committed
b78e47f
Wrap the subject-ID hash+evictions² sum mod 2^64 to match the reference

pavel-kirienko
and
claude
committed
ac84d05
Make the UDP fragment-tree right-neighbor lookup inclusive per the reference

pavel-kirienko
and
claude
committed
d2102a6
Seed the UDP transfer-ID history with the unmasked 64-bit sentinel

pavel-kirienko
and
claude
committed
feef062
Destroy idle slot-free CAN RX sessions at the transfer-ID timeout

pavel-kirienko
and
claude
committed
43e57de
Fix CAN FD framing: the frame format follows the interface, never BRS

pavel-kirienko
and
claude
committed
af3af85
Classify wildcards by whole segment, not by character containment

pavel-kirienko
and
claude
committed
d5e3e5f
Allow CAN unicast to remote node-ID 0

pavel-kirienko
and
claude
committed
03f278e
Drop SLCAN standard-ID data frames instead of aliasing them as extended

pavel-kirienko
and
claude
committed
deee162
Restrict Linux multicast RX sockets to their own group membership

pavel-kirienko
and
claude
committed
fa40030
Validate subject_id_modulus with the reference predicate

pavel-kirienko
and
claude
committed
7aaae71
Wake a parked SocketCAN reader on close/fail

pavel-kirienko
and
claude
committed
bbab7ab
Send UDP redundant interfaces concurrently with per-socket serialization

pavel-kirienko
and
claude
committed
a5c4ef1
Use conforming subject-ID moduli in tests after the modulus validation

pavel-kirienko
and
claude
committed
b56395e
Bound subject_id_modulus validation against oversized values

pavel-kirienko
and
claude
committed
93833db
Harden the UDP concurrent send path

pavel-kirienko
and
claude
committed
8993ad0
Refine SocketCAN close/fail RX semantics

pavel-kirienko
and
claude
committed
7d015c1
Make node and transport setup paths transactional

pavel-kirienko
and
claude
committed
e414d6b
Keep response streams alive across node close and implicit GC

pavel-kirienko
and
claude
committed
4bcbf69
Bound per-remote state under untrusted traffic

pavel-kirienko
and
claude
committed
6c20570
Update docs and changelog for the correctness fixes

pavel-kirienko
and
claude
committed
b5b3551
Address review round 1: doc gaps, TX-socket rollback, comment cleanup

pavel-kirienko
and
claude
committed
7d4b57e
Simplify per review consensus: dedicated node housekeeping task

pavel-kirienko
and
claude
committed
af9f4c0
Fix Windows CI: avoid a real loopback multicast send in the lock test

pavel-kirienko
and
claude
committed
a0439dc
Bump version and clean the changelog
pavel-kirienko
committed
cf148b0
Update the review-loop skill
pavel-kirienko
committed
15daf8a
Set FDF and BRS on every CAN FD frame

pavel-kirienko
and
claude
committed
f5761e3
Stop acting on a closed node; guard the last constructor acquisition

pavel-kirienko
and
claude
committed
01e8c23
Count a cancelled interface as a failed send, not a delivery

pavel-kirienko
and
claude
committed
a1fbcd5
Keep the periodic housekeeping loops alive across a faulty sweep

pavel-kirienko
and
claude
committed
4b3bbb9
Fix media teardown: deregister before close, and don't misreport a clean close

pavel-kirienko
and
claude
committed
7970964
Record why three unbounded-looking structures are correct as written

pavel-kirienko
and
claude
committed
51c01ba
Run ruff in CI, test on Python 3.14, and tidy loose ends

pavel-kirienko
and
claude
committed
ba5316a
Tolerate event loops without reader/writer registration

pavel-kirienko
and
claude
committed
aca2eec
Downsize the comments and docs added by this branch

pavel-kirienko
and
claude
committed
6e3b369