docs: add ADR-0012 (confluence signal engine) + glossary terms (M22 slice 12)

ADR-0012 documents the 34-slot confluence signal engine architecture: slot
catalog, redundancy-aware rack evaluation, CandleProvider seam, closed-loop
signal history, and picture-change alert producer. CONTEXT.md gains a new
'Confluence Signal Engine (M22)' glossary section.

CONTEXT.md

@@ -152,6 +152,28 @@ Screener output policy (both modules): symbol + "why matched" + one-tap "open in
 
 **Portfolio Impact Commentary** — LLM-generated (Druckenmiller lens), two-horizon: short-term reaction risk (per-event, per-correlation-cluster, with historical averages + sample size + disclaimer, never a forecast) and long-term structural read (regime-shift framing). Cites M18 regime + the user's actual holdings. Always on (matches Analyst Voice "always on" decision).
 
+## Confluence Signal Engine (M22)
+
+**Confluence Signal Engine** — The multi-axis evidence aggregator that produces a per-symbol "picture quality" (strong/moderate/weak-bullish, mixed, weak/moderate/strong-bearish, sparse) from34 independently-evaluable slots across 6 families: technical (15), institutional (5), macro (5), seasonal (5), flows (3), sentiment (1). ADR-0007: describes the picture, never recommends action. ADR-0012.
+
+**Slot** — A named, independently-evaluable check whose firing state contributes bullish or bearish evidence to a confluence rack. Each slot has a `SlotBody` (bull / bear / exit), a `SlotFamily`, a `SlotGranularity` (1d / 1wk), and an ADR-safe `explain` note (evidence sentence, never a directive). The 34-slot catalog is defined in `confluenceSlots.ts`.
+
+**Rack** — A named subset of the 34 slots that evaluates a symbol's confluence. Can be a system preset (Full Confluence, Technical Momentum, Macro+Flows+Sentiment) or user-created. A rack evaluation produces the redundancy-discounted evidence totals and picture quality label.
+
+**Redundancy Group** — A set of slots that measure the same underlying condition (e.g. goldenCross + trendAlignment + pullbackToEMA21 all measure trend state). Evidence within a group decays geometrically (1 + 0.5 + 0.25 ...) so correlated signals count once, not triple. Defined in `confluenceLibrary.ts`.
+
+**Picture Quality** — The evidence-based label for a symbol's confluence: strong/moderate/weak-bullish, mixed, weak/moderate/strong-bearish, sparse. Labeled from redundancy-discounted evidence totals using direction ratio (0.6) and magnitude thresholds (strong ≥ 4.0, moderate ≥ 2.0, sparse < 1.0 total evidence).
+
+**CandleProvider** — The seam that resolves a symbol's candles per-granularity (1d/1wk) from the cache, with a realtime fold-in of the freshest live quote. Used by confluence evaluators instead of `cache.get` inline so a future realtime/replay source can slot in without touching slot logic. `candleProvider.ts`.
+
+**Signal History** — The closed-loop log: every slot fire is recorded with the as-of date, rack, and picture quality. A resolver later checks whether price moved the expected direction over 4 weeks (bull → up, bear/exit → down). Verdicts: `real` (confirmed), `false_alarm`, or `deferred` (not enough forward bars). Per-slot reliability weights (0.5–1.25) allow the rack to self-tune.
+
+**Reliability Weight** — A 0.5–1.25 multiplier applied to a slot's evidence based on its historical follow-through: ≥8 resolved fires at ≥90% hit rate → 1.25×; <2 resolved or ≤50% → 0.5×; thin sample → 0.75×. The rack can multiply per-slot evidence by this to self-tune.
+
+**Confluence Change** — A detected shift in picture quality tier or net evidence (≥ 0.35 shift). Driven by the `confluence_change` alert producer. Throttled to 5/hr. ADR-0007 framing: "the picture has changed," never "act now."
+
+**Confluence Universe** — The 15 research symbols + SPY benchmark that the confluence engine tracks: PLTR, NVDA, AMD, AAPL, MSFT, SMH, XOM, JPM, UNH, COST, AMZN, CAT, LMT, LIN, NEE. Pinned into the demand set on startup.
+
 **Regime History** — Timeline of regime classifications over time; cross-references Rotation Signal History. Pedagogy for "did we detect the shift correctly" — beginner learns which classifier calls were early vs whipsawed.
 
 ## LLM Data Provenance (ADR-0006)