fix(models): carry failure_class through the DB read path

[usjoh]

--attributable was reporting the opposite of the evidence, silently.

The bug

db_attempt_rows never SELECTed the failure_class column, so its row dicts reached the classifier without it. DB rows also carry no notes, so model_log_row_failure_class had no fallback and returned None for every failure. --attributable keeps a failure only when it is classed model — so it rejected all of them.

Measured on this machine's log, before:

total attempts : 371
non-PASS       : 137
--attributable : "dropped 137 row(s)"   ← every failure in the log
result         : all 33 model rows read 100% first-try

A filter advertised as removing failures no model caused was removing the model's failures too, and printed a reassuring count while doing it.

The column was populated correctly in both runs.jsonl and ringer.db the whole time. Only the reader dropped it — which is why no db rebuild ever fixed it, and why the JSONL fallback path was right while the default path was wrong.

After

--attributable : dropped 46   (2 engine-error + 44 unknown)

Matches the JSONL path exactly, and the scoreboard shows a real distribution again — 0%, 20%, 33%, 44%, 57%, 67%, 74%, 75%, 77%, 100% — instead of a uniform 100%.

opus-5's OpenRouter outage rows are classed unknown and are now genuinely excluded, so the 100% it shows over its one completed task finally means what it claims. (That claim was made in #35 on the strength of the broken flag, corrected in #37, and is only now actually true.)

What this does not do

The four ohalloran-demonstrator rows are stored as model and are correctly retained by the now-working filter. That classification is itself wrong — the workers exited cleanly and could not write to a denied path (#36, #37) — but re-classifying them is a separate problem from the reader dropping the column. claude-opus-4.8 and claude-sonnet-5 still read 0% on site-build in both views.

Verification

• Full suite: 356 pass
• Prove-fail: with ringer.py reverted and the tests kept, all four fail
• Tests pin the classes surviving the round trip, that a judged failure is kept while engine-error and unknown are dropped, that the DB and JSONL paths agree on what counts as evidence, and a guard that any attempts column the classifier consumes actually reaches the row dicts

🤖 Generated with Claude Code