Merge pull request #104 from abinaze/fix/r5-wire-cache-ttl

Wire get_cache_ttl() into _STATS_TTL and _VERIFY_TTL (R-5, partial)

api/dependencies.py

@@ -17,12 +17,20 @@ from dotenv import load_dotenv
 from loguru import logger
 from neo4j import GraphDatabase
 from neo4j.exceptions import ServiceUnavailable, AuthError
+from config.model_selector import get_cache_ttl
 
 load_dotenv()
 
 _driver          = None
 _last_verified_at = 0.0
-_VERIFY_TTL      = 30.0   # seconds between connectivity re-checks
+# R-5 fix: was a hardcoded 30.0. This TTL exists specifically to prevent
+# the connection-storm scenario the module docstring above describes
+# (many concurrent requests all re-verifying at once) -- a concern that
+# scales with concurrent traffic, which is exactly what PROFILE's tiers
+# already track (a "low" desktop install is single-user; "high" is a
+# multi-user server). Same reasoning as _STATS_TTL in api/main.py: pure
+# freshness/overhead trade-off, no analytical-completeness question.
+_VERIFY_TTL      = float(get_cache_ttl())   # seconds between connectivity re-checks
 _driver_lock     = threading.Lock()
 
 

api/main.py

@@ -16,6 +16,7 @@ from api.models import HealthResponse, StatsResponse
 from processing.alias_graph import AliasGraph          # Phase 32
 from processing.entity_resolver_v2 import EntityResolverV2  # Phase 32
 from config.runtime_profile import PROFILE as _RUNTIME_PROFILE
+from config.model_selector import get_cache_ttl
 import config.settings  # noqa: F401 -- Phase D1: importing this loads any
                          # local desktop config file into os.environ as a
                          # side effect (mirrors load_dotenv() above). Must
@@ -149,7 +150,16 @@ def health_check():
 # H-07 FIX: module-level cache vars (must be at module scope, not inside function)
 _stats_cache      = None
 _stats_cached_at  = 0.0
-_STATS_TTL        = 60.0
+# R-5 fix: was a hardcoded 60.0 (which happened to equal PROFILE["high"]'s
+# own cache_ttl_seconds, coincidentally -- this constant predates the
+# tiered PROFILE system entirely). Pure cache-freshness knob: changes how
+# OFTEN /stats recomputes, never WHAT it computes -- unlike graph_depth/
+# investigation_layers, there is no analytical-completeness trade-off
+# here, so this one genuinely is the mechanical swap R-2/R-3/R-4 already
+# were, not a design question. A low-resource desktop install now caches
+# longer (300s) to avoid frequent full-graph scans; a high-tier server
+# keeps the original 60s for fresher numbers.
+_STATS_TTL        = float(get_cache_ttl())
 import threading as _threading          # B-01 FIX: was missing entirely
 _stats_lock       = _threading.Lock()  # B-01 FIX: NameError on every GET /stats