v0.117.0

Cal #38.B — god-node alarm-fatigue suppression + edge-count staleness stamp (receipt #10 findings #4 + #5)

Receipt #10 finding #4: "God-node warnings have alarm fatigue — PScope/AppError/UUID fire on nearly every PR. Needs a project ubiquitous-types suppression list." Finding #5: "Edge-count drift — cached preflight (AppError 893) vs live MCP (1074) disagree; pick one source of truth."

B1 — ubiquitous-type discrimination (alarm fatigue). blastRadius no longer returns an opaque god_node_match boolean alone — it now itemizes god_node_matches: [{symbol, ubiquitous}] and a discriminating_god_node_match flag. The preflight god-node warning fires the ⚠️ only for a discriminating match (a non-ubiquitous god-node); a match on ubiquitous types alone downgrades to an ℹ️ info note, killing the every-PR alarm fatigue.

The ubiquitous set is auto-derived via degree-DOMINANCE, not a flat top-K. A flat top-K can't work: the god-node match window is the top-10, so any flat top-K ≥ 10 would mark every match ubiquitous and suppress all warnings. Instead, of the top-10 god-nodes, those whose degree is ≥ _UBIQUITOUS_DOMINANCE_FACTOR (2) × the match-window floor degree are flagged ubiquitous — separating the fires-on-every-PR hubs from merely-high domain symbols. Fails safe: a flat degree distribution yields an empty set (no suppression → warnings still fire).

Greenfield's Q1a chose hybrid (auto + override). New config.graphify.ubiquitous_types knob: a plain name is FORCE-ADDED to suppression; a !-prefixed name is FORCE-KEPT (exempt — e.g. when a normally-ubiquitous type is itself being refactored, you want it back in the warning set). effect_size routing is unchanged (touching a hub is still structurally large) — only the warning surface is quieted, which is where the fatigue actually was.

This generalizes greenfield's "K=15 flat list" suggestion per the devt-stays-general guardrail: degree-dominance is the general mechanism that works against devt's actual top-10 match semantics, and it self-tunes as the graph grows (a hardcoded count rots — receipt #10 watched AppError move 893→1074 in a day).

B2 — edge-count staleness stamp. The cached-vs-live divergence (893 vs 1074) is temporal, not a computation bug: preflight already overlays live godNodes() (inc+out degree), but the cached sidecar value is from an earlier graph snapshot than a later live MCP read. preflight-brief.json now carries god_nodes_built_at (the commit the cached counts were computed against), so the divergence is diagnosable as staleness rather than a silent contradiction. Per greenfield's Q5: the absolute count doesn't change any decision ("high-degree, weight it") — what erodes trust is the unexplained contradiction.

Drift-guard stack now 99-deep K94-K192. Smoke gate K192 asserts the discrimination (dominant hub → suppressed info note; moderate god-node → ⚠️ fires; !-force-keep → re-fires) with a synthetic dominance-distribution graph.

Cal #38 carryover: #38.C (incremental lane writes + query_graph AND-then-OR fallback + disk preflight), #38.D (comment-hygiene per [[devt-stays-general]]).