fix: revalidate the vault price history cache on live universe loads

[miohtama]

Why

A live hyper-ai executor failed to start with DataTooOld:

tradeexecutor.strategy.universe_model.DataTooOld: Candle data 2023-03-01 - 2026-07-24 is too old to work with
we require threshold 2026-07-25, diff is 1 day, asked best before duration is 2 days

All 327 Hyperliquid vaults reported a newest source timestamp of 2026-07-24 14:1x,
three days before the 12:10 crash. The data itself was fine — checking the source
directly, cleaned-vault-prices-1h.parquet had last-modified: 2026-07-27 09:53:55 and
carried rows up to 09:32 that morning, including for the exact vaults in the log
(HLP 0xdfc2…f303 → 09:30:06, Citadel 0xda51…c167 → 09:30:07). The executor was
reading a stale local cache.

fetch_vault_price_history() trusts a cached parquet for 24 hours with no remote check:

• Inside the window no HEAD request is made, so a parquet published after the last
  check is not picked up, and nothing reports that the local copy has fallen behind.
  Reproduced in isolation: cache mtime 23h old, remote newer with a different ETag and
  size → no HEAD, no download, stale file served.
• The window is measured from the last validation, not the last download, because a
  HEAD check that finds the remote unchanged calls os.utime(path, None).

Scope note: a 24h window bounds staleness at roughly 24h plus pipeline lag, so this alone
does not account for a three-day gap. This is hardening of a real defect, not a confirmed
root cause. Closing out the remaining gap needs the executor's own
Vault price history cache hit/expired and Vault history freshness summary lines from
the incident, which were not in the captured output. A second candidate is still open:
_is_matching_http_cache_metadata() falls back to remote_last_modified <= local_mtime
even when sidecar metadata disagrees, which is unsound once mtime has been bumped —
currently masked by the Content-Length guard, and left for a follow-up.

Lessons learnt

• An mtime-keyed cache window silently changes meaning once anything touches mtime for
  bookkeeping. It stops reading as "age of the data" and starts meaning "time since we
  last looked", which is not what a bare timedelta(hours=24) conveys.
• A cache whose expiry is invisible in the happy path gives no signal when it goes wrong.
  The blind window emitted no log line; the staleness surfaced three layers downstream as
  a DataTooOld crash behind a 300-line table dump, and a Sentry Message too long
  logging error on the way.
• Freshness policy belongs to the caller, not the transport. A backtest wants the cheap
  cached path; a live executor must not guess. Same function, opposite requirements.
• Verifying "is the upstream data actually stale?" against the source took one HEAD
  request and one parquet read, and ruled out the entire data-pipeline branch of the
  investigation immediately. Worth doing before reading any cache code.

Summary

• load_partial_data() passes revalidate=execution_context.mode.is_live_trading() to
  client.transport.fetch_vault_price_history(), and to the client.fetch_vault_price_history()
  fallback branch, so live universe loads always validate the cache against the remote.
  Backtests keep the plain cached path.
• Bumps the deps/trading-strategy submodule to pick up the revalidate flag.
• One new test, test_load_partial_data_revalidates_vault_history_cache_only_when_live,
  captures the flag the transport receives and asserts True for real_trading and
  False for backtesting.
• Updates the vault-data pipeline notes in test_hyper_ai_stale_data_indicators.py,
  including dropping a stale cache.py:625 line reference.
• Existing stubs in test_vault_live_cutoff.py accept the new keyword.

Verified against the live server: with a 25-minute-old cache — deep inside the window
that produced the incident — the HEAD now runs and correctly short-circuits the 197 MB
download because the remote is unchanged.

Merge order: tradingstrategy-ai/trading-strategy#243 must land first; the submodule
pointer here should then be repointed at the merged commit rather than the branch commit.

🤖 Generated with Claude Code