fix(infinifi): re-arm liquid reserves alert when cache is stale

[spalen0]

Problem

The Infinifi "Low Liquid Reserves" alert only fires on a downward crossing of the $8M threshold (last >= 8M and current < 8M). If the script doesn't run for an extended period, the cached last_reserves becomes stale and the alert can stay suppressed indefinitely, even when reserves remain well below threshold for the entire outage.

Fix

In protocols/infinifi/main.py, the Liquid Reserves alert block now also writes a *_ts cache key on every run. On read, if the cached value is older than LIQUID_RESERVES_STALE_AFTER_SECONDS (12h), the previous reading is treated as "at threshold", which re-arms the crossing detection so the alert fires again on the next successful run.

A logger.info records the re-arm event so it is visible in logs.

Audit of other monitors

A repo-wide audit (full report in the PR conversation) found ~35 other cache keys with the same shape (delta/change alerts gated on a last > 0 check or an if previous is not None initialisation). Highlights:

• CRITICAL/HIGH wrong-delta after outage: protocols/3jane/main.py (PPS, TVL, debt cap, insurance fund), protocols/strata/main.py (sUSDe/srUSDe rate, jr drain, strategy ratio, total deposits), protocols/ustb/main.py (Chainlink NAV, supply), protocols/stables/oracles.py (rate-oracle monotonicity, roundId regression), protocols/maple/main.py (PPS, TVL, delegate cover), protocols/lrt-pegs/origin_protocol.py (redeem value), protocols/apyusd/main.py (rate), protocols/rtoken/monitor_rtoken.py (StRSR rate), protocols/morpho/governance_v2.py (owner/curator/sentinels/allocators/adapters).
• False-positive risk on cache loss: protocols/3jane/main.py insurance fund outflow — previous_shares is treated as 0, so the entire fund balance looks like a one-step outflow on the first post-outage run.
• The only true STUCK_AFTER_OUTAGE bug is the Infinifi reserves alert fixed here; every other finding is a wrong-delta / cadence issue where the first run after cache loss silently rebaselines.

Suggested follow-up (separate PR)

Add a tiny pair of helpers to utils/cache.py:

def read_cache_fresh(key: str, max_age_seconds: int, filename: str = cache_filename):
    """Returns the cached value, or None if missing/stale (>max_age_seconds old)."""

def write_cache_fresh(key: str, value, filename: str = cache_filename):
    """Writes value + a current-timestamp marker for read_cache_fresh()."""

Then convert each of the 35 wrong-delta sites in small protocol-scoped PRs, picking a per-monitor max_age_seconds (4h for CRITICAL signals, 6–8h for MEDIUM/LOW). Happy to do that as a follow-up if this approach is approved.

Checklist

• ruff format
• ruff check
• mypy (ran locally clean)