fix(receipts): preserve sidecar evidence

Co-Authored-By: Codex <codex@openai.com>

Author: solomonneas

.claude/memory-handoffs/20260728-030215-issue-567-write-once-sidecars.md

@@ -0,0 +1,41 @@
+# Memory Handoff
+
+## Type
+
+bugfix
+
+## Title
+
+Write-once run sidecar revisions
+
+## Summary
+
+Resume and patch-reference updates previously replaced worker and synthesis sidecars in place. They now append private, immutable revisions while retaining the legacy files as current compatibility projections.
+
+## Durable facts
+
+- Initial emissions and every update append six-digit JSON revisions below `revisions/worker-results/` and `revisions/synthesis/`.
+- A first update of a legacy run imports the compatibility projection as revision 000001 before appending the new revision.
+- A failed compatibility projection write must not replace the old projection after immutable evidence has been written.
+- Revision files use mode 0600. POSIX writes fsync each new directory entry, the revision file, and the leaf directory before updating the projection.
+- Existing readers continue loading `worker-results.json` and `synthesis.json`; no receipt schema or dependency changed.
+
+## Evidence
+
+- files changed: `src/brigade/aboyeur.py`, `src/brigade/run_resume.py`, `tests/test_aboyeur.py`, `tests/test_run_resume.py`, `tests/test_runs_cmd.py`
+- RED receipts: `20260728-030303-work-verify-9ae890`, `20260728-031100-work-verify-052751`, `20260728-031433-work-verify-ca2331`, `20260728-031552-work-verify-cf2e04`
+- GREEN receipts: `20260728-031646-work-verify-d12faa`, `20260728-031656-work-verify-8bfb8d`, `20260728-031822-work-verify-4cb89a`, `20260728-031829-work-verify-d68b11`
+
+## Recommended memory action
+
+no-card
+
+## Target document
+
+.learnings/LEARNINGS.md
+
+## Suggested document content
+
+### Write-once run sidecar revisions
+
+Worker and synthesis evidence is append-only under `revisions/<sidecar>/`, while the original sidecar paths remain current compatibility projections. On the first update of a legacy run, archive the exact old bytes before appending the updated document. Persist revision evidence before replacing the projection, and retain the new revision if projection replacement fails. On POSIX, directory-entry fsync ordering is part of the durability contract.

src/brigade/aboyeur.py

@@ -492,6 +492,88 @@ def _write_json(path: Path, payload: object) -> None:
     localio.write_text_atomic(path, json.dumps(payload, indent=2, sort_keys=True) + "\n")
 
 
+def _revision_contains(revisions_dir: Path, projection: bytes) -> bool:
+    """Return whether a preserved sidecar revision matches ``projection``."""
+    if not revisions_dir.is_dir():
+        return False
+    for revision_path in revisions_dir.glob("*.json"):
+        try:
+            if revision_path.read_bytes() == projection:
+                return True
+        except OSError:
+            continue
+    return False
+
+
+def _fsync_directory(path: Path) -> None:
+    """Persist directory entries where the platform supports directory fsync."""
+    if os.name != "posix":
+        return
+    directory_fd = os.open(path, os.O_RDONLY)
+    try:
+        os.fsync(directory_fd)
+    finally:
+        os.close(directory_fd)
+
+
+def _ensure_directory_durable(path: Path) -> None:
+    """Create missing directory levels and persist each new parent entry."""
+    missing: list[Path] = []
+    current = path
+    while not current.exists():
+        missing.append(current)
+        current = current.parent
+    for directory in reversed(missing):
+        try:
+            directory.mkdir()
+        except FileExistsError:
+            if not directory.is_dir():
+                raise
+        else:
+            _fsync_directory(directory.parent)
+
+
+def _write_new_revision(revisions_dir: Path, encoded: bytes) -> None:
+    """Exclusively create the next immutable sidecar revision."""
+    _ensure_directory_durable(revisions_dir)
+    sequence = max(
+        (int(path.stem) for path in revisions_dir.glob("*.json") if path.stem.isascii() and path.stem.isdecimal()),
+        default=0,
+    )
+    while True:
+        sequence += 1
+        revision_path = revisions_dir / f"{sequence:06d}.json"
+        writer_acquired = False
+        try:
+            fd = os.open(revision_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
+            writer_acquired = True
+            with os.fdopen(fd, "wb") as handle:
+                handle.write(encoded)
+                handle.flush()
+                os.fsync(handle.fileno())
+            _fsync_directory(revisions_dir)
+        except FileExistsError:
+            continue
+        except BaseException:
+            if writer_acquired:
+                revision_path.unlink(missing_ok=True)
+            raise
+        return
+
+
+def write_sidecar_revision(run_dir: Path, filename: str, payload: object) -> None:
+    """Append an immutable sidecar revision, then update its compatibility file."""
+    projection_path = run_dir / filename
+    revisions_dir = run_dir / "revisions" / Path(filename).stem
+    if projection_path.exists():
+        legacy_projection = projection_path.read_bytes()
+        json.loads(legacy_projection)
+        if not _revision_contains(revisions_dir, legacy_projection):
+            _write_new_revision(revisions_dir, legacy_projection)
+    _write_new_revision(revisions_dir, (json.dumps(payload, indent=2, sort_keys=True) + "\n").encode("utf-8"))
+    _write_json(projection_path, payload)
+
+
 def _utc_iso(value: datetime) -> str:
     return value.astimezone(timezone.utc).isoformat().replace("+00:00", "Z")
 
@@ -1759,7 +1841,7 @@ def set_artifact_patch_ref(output_dir: Path, patch_ref: str = "changes.patch") -
         else:
             receipt_schema.stamp_synthesis_document(payload)
         try:
-            _write_json(path, payload)
+            write_sidecar_revision(output_dir, filename, payload)
         except OSError as exc:
             raise runguard.RunGuardError(f"failed to record artifact patch reference in {filename}: {exc}") from exc
 
@@ -2983,8 +3065,9 @@ def dispatch_interrupted() -> None:
     worker_results = _mark_noop_worker_results(worker_results, suspected_noop)
     if output_dir is not None:
         worker_results = _write_worker_logs(output_dir, worker_results)
-        _write_json(
-            output_dir / "worker-results.json",
+        write_sidecar_revision(
+            output_dir,
+            "worker-results.json",
             receipt_schema.worker_results_document(
                 _worker_payload(worker_results),
                 ground_truth=ground_truth,
@@ -3079,7 +3162,7 @@ def dispatch_interrupted() -> None:
                 ground_truth=ground_truth,
             )
         )
-        _write_json(output_dir / "synthesis.json", synthesis_payload)
+        write_sidecar_revision(output_dir, "synthesis.json", synthesis_payload)
     if not final.ok:
         if output_dir is not None:
             finished_at = datetime.now(timezone.utc)

src/brigade/run_resume.py

@@ -199,8 +199,9 @@ def _resume_locked(run_dir: Path) -> int:
         for r in results
     ]
     ground_truth = worker_data.get("ground_truth") or {}
-    aboyeur._write_json(
-        run_dir / "worker-results.json",
+    aboyeur.write_sidecar_revision(
+        run_dir,
+        "worker-results.json",
         receipt_schema.worker_results_document(
             aboyeur._worker_payload(worker_results),
             ground_truth=ground_truth,
@@ -226,8 +227,9 @@ def _resume_locked(run_dir: Path) -> int:
         reasoning=orchestrator.reasoning,
         env=dict(orchestrator.env) if orchestrator.env is not None else None,
     )
-    aboyeur._write_json(
-        run_dir / "synthesis.json",
+    aboyeur.write_sidecar_revision(
+        run_dir,
+        "synthesis.json",
         receipt_schema.synthesis_document(
             orchestrator=roster.orchestrator,
             result={"ok": final.ok, "detail": final.detail, "text": final.text},