stage-0 warehouse: recorder folds its own tape to parquet; Mac mirrors every 15 min

The nightly Mac-coupled bulk ingest was the fragile link (today: Mac slept
through 08:00, the pipeline started 09:31, and the sftp bulk pull wedged in
timeout cascades holding the db lock — digest blocked behind it). Stage 0
moves the fold onto the box and reduces the Mac to an incremental mirror:

- recorder/fold.py (sidecar, capture stays PID 1): closed gz segments ->
  zstd parquet /data/parquet/<fam>/date=*/segment.parquet, row-parity
  verified, manifest-logged, THEN gz deleted. Deletion invariant STRONGER:
  raw needs a verified parquet; parquet needs a Mac ACK before the disk
  guard may prune it. duckdb capped 384MB; VM 256MB -> 1GB.
- recorder/sync_tape.py (com.jaxperro.tape-sync every 15 min + daily.sh):
  manifest-driven incremental sftp pull, per-file row verify, append into
  rtds.duckdb NATIVE tables (views-over-parquet rejected: sim's per-asset
  point queries would crawl), segment-keyed idempotence shared with the
  legacy ingest.py (kept as fallback), ack back to the box.
- recorder/bootstrap_parquet.py: pre-fold history exported to the mirror,
  parity OK (13.84M trades + 3.07M aux). live/parquet/ is now the complete
  durable layer Stage 1 (MotherDuck/ClickHouse) would consume.
- research/tape.py connect(): brief retry — the 15-min sync holds the
  write lock for seconds.

First run: backlog folded in <60s on the box, 36/36 files mirrored+
verified, rtds.duckdb 13.8M -> 18.2M trades, tape age 24h+ -> ~15-45 min.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fly.recorder.toml

@@ -15,4 +15,4 @@ primary_region = "arn"
 
 [[vm]]
   size = "shared-cpu-1x"
-  memory = "256mb"
+  memory = "1gb"          # fold sidecar: duckdb capped at 384MB; capture headroom