fix(learning-harness): round zero writes down every pair it has paid for

The 176-pair vfc baseline hit its four-hour timeout and wrote nothing. Scores
were accumulated in memory and persisted once at the end, so roughly four
hours of paid API calls were lost. I fixed the exception path this morning --
a provider hiccup costs one pair -- and missed that a timeout, a kill or a
machine going away cost all of them.

Each pair is now appended to `scores.jsonl` and flushed as it is measured, and
`--resume` keeps what a previous run of that exact baseline already bought.
This also gives the per-pair outcomes the round-3 review asked for, which make
the noise floor auditable after the fact instead of only in aggregate.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG

src/openultrasast/cli.py

@@ -223,6 +223,12 @@ def main(argv: list[str] | None = None) -> int:
             command.add_argument("--catalog", type=Path, default=DEFAULT_CATALOG)
             command.add_argument("--slice", choices=SLICE_NAMES, default="all")
             command.add_argument("--json", action="store_true")
+        if name == "baseline":
+            command.add_argument(
+                "--resume",
+                action="store_true",
+                help="keep the pairs a previous run of this baseline already measured and paid for",
+            )
         if name == "candidates":
             command.add_argument("--generator", choices=("ir", "ruleset"), default="ir", help="which pass enumerates the candidate sites")
         if name in {"baseline", "round", "score"}:
@@ -1352,6 +1358,7 @@ def _learning_run(args: argparse.Namespace, cases: Sequence[PairCase], taxonomy:
             slices=(args.slice,) if args.slice != "all" else DEFAULT_SLICES,
             spent_usd=spent,
             thinking=config.learning.thinking,
+            resume=args.resume,
         )
         print(f"learning baseline cost: ${report.cost_usd:.4f}")
         print(f"learning baseline {model}: {len(report.floors)} families, k={report.k_runs} -> {out / 'baseline'}")

src/openultrasast/learning/rounds.py

@@ -103,6 +103,7 @@ def run_baseline(
     spent_usd: Callable[[], float] | None = None,
     cutoff: str = "",
     thinking: bool = False,
+    resume: bool = False,
 ) -> BaselineReport:
     """Clone one detector into every family, measure each family **with its own configuration**, write the artifacts.
 
@@ -120,28 +121,40 @@ def run_baseline(
         write_default_configs(configs_dir, taxonomy, prompt=prompt or _SYSTEM_PROMPT, version="0")
     configs = load_family_configs(configs_dir, taxonomy)
     selected = [case for case in cases if case.slice in set(slices)]
+    # Every pair is written down as it is measured. Round zero over the vfc slice is 1760 detector runs and hours of
+    # paid calls; accumulating in memory and persisting once at the end means a timeout, a kill or a machine going
+    # away costs all of it. An exception already cost only one pair; nothing else did.
+    checkpoint = out_dir / "baseline" / _artifact_key(model, slices, thinking) / "scores.jsonl"
+    done = _checkpointed(checkpoint) if resume else {}
+    if not resume and checkpoint.exists():
+        checkpoint.unlink()
     scores = []
     for case in selected:
         name = _family_of(case, taxonomy)
         config = configs.get(name)
         if config is None:
             continue  # a family with no configuration is not measured rather than measured with someone else's
+        if case.name in done:
+            scores.append(done[case.name])
+            continue
         try:
-            scores.append(score_case(case, scan_factory(config), taxonomy=taxonomy, runs=k_runs, family=name))
+            score = score_case(case, scan_factory(config), taxonomy=taxonomy, runs=k_runs, family=name)
+            scores.append(score)
+            _append_score(checkpoint, score)
         except Exception:  # noqa: BLE001 — a provider hiccup costs one pair, not the whole run
             # Round zero is the most expensive thing this harness does. Raising here threw away everything already
             # measured and paid for, so the only honest response to a flaky endpoint was to run it all again.
-            scores.append(
-                PairFamilyScore(
-                    pair=case.name,
-                    family=name,
-                    slice=case.slice,
-                    fix_date=case.fix_date,
-                    runs=(),
-                    outcome="unscorable",
-                    unscorable_reason="detector_unreachable",
-                )
+            unreachable = PairFamilyScore(
+                pair=case.name,
+                family=name,
+                slice=case.slice,
+                fix_date=case.fix_date,
+                runs=(),
+                outcome="unscorable",
+                unscorable_reason="detector_unreachable",
             )
+            scores.append(unreachable)
+            _append_score(checkpoint, unreachable)
     metrics = aggregate(scores, taxonomy=taxonomy, cutoff=cutoff)
     floors = {name: _floor(name, block, [item for item in scores if item.family == name], k_runs) for name, block in metrics.items()}
     report = BaselineReport(
@@ -160,6 +173,59 @@ def run_baseline(
     return report
 
 
+def _append_score(path: Path, score: PairFamilyScore) -> None:
+    """One line per pair, flushed, so an interrupted run keeps what it measured."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    payload = {
+        "pair": score.pair,
+        "family": score.family,
+        "slice": score.slice,
+        "fix_date": score.fix_date,
+        "outcome": score.outcome,
+        "runs": list(score.runs),
+        "flips": score.flips,
+        "other_findings_vuln": score.other_findings_vuln,
+        "other_findings_fixed": score.other_findings_fixed,
+        "fabricated_vuln": score.fabricated_vuln,
+        "fabricated_fixed": score.fabricated_fixed,
+        "rung": score.rung,
+        "unscorable_reason": score.unscorable_reason,
+    }
+    with path.open("a", encoding="utf-8") as handle:
+        handle.write(json.dumps(payload, sort_keys=True) + "\n")
+        handle.flush()
+
+
+def _checkpointed(path: Path) -> dict[str, PairFamilyScore]:
+    """What a previous run of this exact baseline already measured and paid for."""
+    if not path.is_file():
+        return {}
+    out: dict[str, PairFamilyScore] = {}
+    for line in path.read_text(encoding="utf-8").splitlines():
+        if not line.strip():
+            continue
+        try:
+            row = json.loads(line)
+        except json.JSONDecodeError:
+            continue  # a line half-written when the run died is not a measurement
+        out[str(row["pair"])] = PairFamilyScore(
+            pair=str(row["pair"]),
+            family=str(row["family"]),
+            slice=str(row.get("slice", "")),
+            fix_date=str(row.get("fix_date", "")),
+            runs=tuple(row.get("runs", ())),
+            outcome=row["outcome"],
+            flips=int(row.get("flips", 0)),
+            other_findings_vuln=int(row.get("other_findings_vuln", 0)),
+            other_findings_fixed=int(row.get("other_findings_fixed", 0)),
+            fabricated_vuln=int(row.get("fabricated_vuln", 0)),
+            fabricated_fixed=int(row.get("fabricated_fixed", 0)),
+            rung=row.get("rung", "suspicion"),
+            unscorable_reason=row.get("unscorable_reason"),
+        )
+    return out
+
+
 def score_case(case: PairCase, scan: Scan, *, taxonomy: FamilyTaxonomy, runs: int, family: str | None = None) -> PairFamilyScore:
     """Run one detector over both sides of one pair ``runs`` times and score it for its family.
 

tests/test_learning_rounds.py

@@ -5,6 +5,8 @@ from __future__ import annotations
 import json
 from pathlib import Path
 
+import pytest
+
 from openultrasast.benchmark import ExpectedFinding
 from openultrasast.findings import StaticFinding
 from openultrasast.learning.families import load_families
@@ -295,3 +297,81 @@ def test_two_slices_on_one_model_are_two_baselines_not_one_overwriting_the_other
     table = compare_baselines(out)
     assert set(table) == {"deepseek-v4-flash", "deepseek-v4-flash (vfc)"}
     assert "memory" in table["deepseek-v4-flash (vfc)"] and "injection" in table["deepseek-v4-flash"]
+
+
+def test_round_zero_keeps_every_pair_it_has_already_paid_for(tmp_path: Path) -> None:
+    """A four-hour run that is interrupted must not lose four hours of paid calls.
+
+    Measured the hard way: the 176-pair vfc baseline hit its timeout and wrote nothing, because scores were
+    accumulated in memory and persisted once at the end. An exception was already handled; a timeout, a kill and a
+    machine going away were not, and they cost the same money."""
+    from openultrasast.learning.rounds import run_baseline
+
+    taxonomy = load_families()
+    out = tmp_path / "out"
+    scored: list[str] = []
+
+    def build(_config):  # type: ignore[no-untyped-def]
+        def scan(root: Path) -> list[StaticFinding]:
+            scored.append(root.name)
+            # K=3 means three vulnerable-side scans per pair, so this gives up part way through the third pair
+            if len([item for item in scored if item == "vuln"]) > 6:
+                raise KeyboardInterrupt("the operator gave up")
+            return _steady(root)
+
+        return scan
+
+    cases = [_case(tmp_path, name) for name in ("a", "b", "c", "d")]
+    with pytest.raises(KeyboardInterrupt):
+        run_baseline(
+            cases,
+            taxonomy=taxonomy,
+            configs_dir=tmp_path / "configs",
+            scan_factory=build,
+            model="m",
+            out_dir=out,
+            k_runs=3,
+        )
+    checkpoint = out / "baseline" / "m" / "scores.jsonl"
+    assert checkpoint.is_file(), "the pairs already measured were not written down"
+    rows = [json.loads(line) for line in checkpoint.read_text().splitlines() if line.strip()]
+    assert [row["pair"] for row in rows] == ["a", "b"]
+    assert all(row["outcome"] == "pair_correct" for row in rows)
+
+
+def test_a_resumed_baseline_does_not_pay_twice(tmp_path: Path) -> None:
+    from openultrasast.learning.rounds import run_baseline
+
+    taxonomy = load_families()
+    out = tmp_path / "out"
+    cases = [_case(tmp_path, name) for name in ("a", "b")]
+    run_baseline(
+        cases,
+        taxonomy=taxonomy,
+        configs_dir=tmp_path / "configs",
+        scan_factory=_always(_steady),
+        model="m",
+        out_dir=out,
+        k_runs=3,
+    )
+    calls: list[str] = []
+
+    def counting(_config):  # type: ignore[no-untyped-def]
+        def scan(root: Path) -> list[StaticFinding]:
+            calls.append(root.name)
+            return _steady(root)
+
+        return scan
+
+    report = run_baseline(
+        cases,
+        taxonomy=taxonomy,
+        configs_dir=tmp_path / "configs",
+        scan_factory=counting,
+        model="m",
+        out_dir=out,
+        k_runs=3,
+        resume=True,
+    )
+    assert calls == [], "every pair was already measured and paid for"
+    assert report.metrics["injection"]["scorable"] == 2