fix(bib): revert sleep to 3.7s — 1000/hr is the binding cap

Empirical: at 1.3s pacing the API hits 429 constantly and our 60s
backoff makes throughput WORSE than the conservative 3.7s pacing
(observed ~5/min with 16 workers + heavy 429s vs ~16/min steady at
3.7s). The 50/min burst cap is looser than the 1000/hr sustained cap
— 3.7s ≈ 970/hr stays under both.

Concurrent CDN attachment downloads + 16 workers stay (those don't
count against the API limit).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/bib/regulations_gov.py

@@ -89,7 +89,7 @@ class Client:
         self,
         api_key: str | None = None,
         *,
-        sleep: float = 1.3,  # ~46 req/min — under the 50/min burst cap
+        sleep: float = 3.7,  # ~970 req/hr — under the 1000/hr cap (binding)
         dl_workers: int = 16,  # concurrent attachment download threads
         client: httpx.Client | None = None,
     ) -> None:

src/cli/bib.py

@@ -345,10 +345,11 @@ def backfill_comments(
         help="Cap items processed this run. 0 = no cap (multi-day crawl).",
     ),
     sleep: float = typer.Option(
-        1.3,
+        3.7,
         "--sleep",
-        help="Seconds between API calls. 1.3s ≈ 46/min — under the "
+        help="Seconds between API calls. 3.7s ≈ 970/hr — just under the "
-        "50/min reg.gov burst cap. CDN downloads are concurrent and free.",
+        "1000/hr reg.gov hourly cap (the binding limit; the 50/min burst "
+        "cap is looser). CDN downloads are concurrent and don't count.",
     ),
     log_path: Path = typer.Option(
         Path("/tmp/bib-backfill-comments.log"),