fix(j1939): decode by source address (#13), split shared messages, stop 11-bit ID cross-matching

[dinacaran]

Fixes #13.

Problem

#13: when a DBC defines a J1939 PGN at several source addresses (DM1_239 at SA 0xEF, DM1_243 at SA 0xF3), the decoder still added every message with that PGN as a fallback candidate and used the first one. DM1 frames from every other ECU (VCU 0x13, OBC 0xE6, …) decoded into xxx_239 as well, so one node's DTC signal mixed in DTCs from other nodes. MF4 files had the same bug, because asammdf does its own PGN matching.

Found while testing #13: when a DBC defines a message once at a placeholder SA and several ECUs send it, those ECUs were merged into one series. The bulk loaders insert one frame ID at a time, so the merged series held one back-to-back sweep of the recording per sender, out of time order. In a log, xxx:Switch comes from four nodes: two send real switch states and two send 3 (not available). Zoomed out it drew a solid box between 0 and 3. Zoomed in, pyqtgraph's binary-search clipping picked one sweep and showed a flat 3.

Found in the same lookup: the decoder also indexed and looked up every ID masked to 11 bits, so standard and extended IDs cross-matched. Extended frame 0x18FECA13 got the candidates [Std 0x213, DM1], and the loader used the standard message. Standard frame 0x2EF decoded as the extended message 0x18FECAEF.

Changes

1. SA-specific PGN matching (DBCDecoder.candidates_for). If a PGN is defined at more than one SA, a frame only matches the message for its own SA. An SA the database doesn't define stays undecoded. A PGN defined at a single SA keeps the any-SA fallback, because that SA is usually a placeholder. The vectorized decoder and the debug inspector now call this one method instead of keeping their own copies of the rule.
2. One series per sender. If frames from more than one SA decode into the same message, each sender gets its own series, named like CxVS [SA 0x17]. A single sender keeps the database name.
3. Time order. SignalStore flags any series whose bulk insert starts before its last sample. sort_merged_series() reorders only those series after loading. All three bulk load paths call it, so one sender on two priorities still loads as one ordered series.
4. MF4. The reader reads the frame ID from each asammdf group comment. It drops groups the decoder would not match, which applies [BUG] J1939 Decoding ignores Source Address (SA), causing signal overlap on PGN FECA (DM1) #13 to MF4. It also applies the same per-SA split instead of merging groups back together.
5. No 11-bit cross-matching. Messages are indexed by ID type, so a standard frame only matches standard messages and an extended frame only extended ones. An ID that still carries the DBC's 0x80000000 extended flag keeps matching. The candidate cache key now includes the ID type. The bulk ASC/BLF loader still infers the type from the ID (anything above 0x7FF counts as extended). That covers J1939, but an extended frame with an ID of 0x7FF or below is still treated as standard there, as it was before this PR.

Testing

• tests/test_dbc_decoder.py: 8 decoder tests for [BUG] J1939 Decoding ignores Source Address (SA), causing signal overlap on PGN FECA (DM1) #13 (undefined SAs stay undecoded, a frame matches its own SA's message, priority is ignored, the single-SA fallback is kept, the vectorized decoder delegates).
• tests/test_j1939_senders.py (new): 8 end-to-end LoadWorker tests on generated ASC and MF4 files: split per sender, single sender keeps its name, [BUG] J1939 Decoding ignores Source Address (SA), causing signal overlap on PGN FECA (DM1) #13 on both formats, one sender on two priorities stays in time order, and the store sort.
• tests/test_dbc_decoder.py: 4 tests for ID-type matching. Three of them failed before the fix (extended matching standard on the low bits, standard matching extended, and one ID sent as both types). The fourth checks that each type still matches its own messages.
• Defect injection: reverting the decoder and loaders, dropping the loader sort, skipping the MF4 group handling, disabling the split, or disabling the store's out-of-order flag each fails the tests aimed at it.
• Full suite: 604 passed.
• Checked by hand on the J1939.MF4 with J1939.dbc: now loads as [SA 0x00], [SA 0x17], [SA 0x21] and [SA 0x31], each with 1958 samples in time order.

Note

In logs where a message now splits, saved configs or formulas that name the merged series will no longer find it. Nothing renames old references to the new names.

🤖 Generated with Claude Code