diff --git a/src/components/insights/confidence-meter.tsx b/src/components/insights/confidence-meter.tsx
index 7f33106ea..198222662 100644
--- a/src/components/insights/confidence-meter.tsx
+++ b/src/components/insights/confidence-meter.tsx
@@ -10,7 +10,7 @@ export type ConfidenceBand = "high" | "medium" | "low" | "draft";
 export type ConfidenceMeterVariant = "bars" | "ring";
 
 export interface ConfidenceMeterProps {
-  /** Score from `computeConfidence()` — 0..100, integer expected. */
+  /** Score carried on the insight payload — 0..100, integer expected. */
   value: number;
   /** Visual variant. Defaults to `bars`. */
   variant?: ConfidenceMeterVariant;
diff --git a/src/lib/ai/__tests__/citation-coverage-logging.test.ts b/src/lib/ai/__tests__/citation-coverage-logging.test.ts
index 1dd249802..84e575d88 100644
--- a/src/lib/ai/__tests__/citation-coverage-logging.test.ts
+++ b/src/lib/ai/__tests__/citation-coverage-logging.test.ts
@@ -1,7 +1,4 @@
-import { describe, it, expect, beforeEach, vi } from "vitest";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn } from "../types";
+import { describe, it, expect, vi } from "vitest";
 import {
   detectsNormativeClaim,
   computeCitationCoverage,
@@ -236,97 +233,3 @@ describe("computeCitationCoverage()", () => {
     expect(result.uncitedNormativeRecommendationIds).toEqual(["uncited-1"]);
   });
 });
-
-describe("generateInsight() — citation-coverage annotation", () => {
-  beforeEach(() => {
-    annotateMock.mockReset();
-  });
-
-  /**
-   * v1.4.16 phase B5d — `applyConfidenceOverride()` also calls
-   * `annotate()` (action: "confidence.override.applied"), so the
-   * citation-coverage assertions filter by meta-key rather than by
-   * call count.
-   */
-  function findCoverageCall(calls: Array<unknown[]>): Record<string, unknown> {
-    for (const call of calls) {
-      const arg = call[0] as { meta?: Record<string, unknown> } | undefined;
-      const meta = arg?.meta;
-      if (meta && "ai_total_recommendations" in meta) {
-        return meta;
-      }
-    }
-    throw new Error("citation-coverage annotate() call not found");
-  }
-
-  it("emits an annotate() call with the coverage breakdown on success", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify({
-        summary: "x",
-        recommendations: [
-          {
-            id: "rec-cited",
-            text: "Aim for a target below 140/90",
-            severity: "important",
-            metricSource: baseMetricSource,
-            rationale: baseRationale,
-            referenceId: knownRefId,
-          },
-        ],
-        citations: [baseCitation],
-        warnings: [],
-      }),
-    });
-    await generateInsight(
-      provider,
-      singleUserTurn({ system: "system", user: "user" }),
-    );
-    const meta = findCoverageCall(annotateMock.mock.calls);
-    expect(meta).toMatchObject({
-      ai_total_recommendations: 1,
-      ai_normative_recommendations: 1,
-      ai_cited_normative_recommendations: 1,
-      ai_uncited_normative_recommendation_ids: [],
-    });
-  });
-
-  it("annotation flags the uncited normative rec", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify({
-        summary: "x",
-        recommendations: [
-          {
-            id: "rec-naked",
-            text: "Your BP should stay below 140/90",
-            severity: "important",
-            metricSource: baseMetricSource,
-            rationale: baseRationale,
-          },
-        ],
-        citations: [baseCitation],
-        warnings: [],
-      }),
-    });
-    await generateInsight(
-      provider,
-      singleUserTurn({ system: "system", user: "user" }),
-    );
-    const meta = findCoverageCall(annotateMock.mock.calls);
-    expect(meta).toMatchObject({
-      ai_normative_recommendations: 1,
-      ai_cited_normative_recommendations: 0,
-      ai_uncited_normative_recommendation_ids: ["rec-naked"],
-    });
-  });
-
-  it("does NOT annotate when generation throws (only on success)", async () => {
-    // Provider returns invalid JSON twice → InsightSchemaError surfaces.
-    const provider = new MockAIProvider({
-      responses: ["not json", "still not json"],
-    });
-    await expect(
-      generateInsight(provider, singleUserTurn({ system: "s", user: "u" })),
-    ).rejects.toThrow();
-    expect(annotateMock).not.toHaveBeenCalled();
-  });
-});
diff --git a/src/lib/ai/__tests__/citation-enforcement.test.ts b/src/lib/ai/__tests__/citation-enforcement.test.ts
index 684436283..4e4794500 100644
--- a/src/lib/ai/__tests__/citation-enforcement.test.ts
+++ b/src/lib/ai/__tests__/citation-enforcement.test.ts
@@ -1,14 +1,10 @@
 import { describe, it, expect } from "vitest";
 import {
   aiInsightResponseSchema,
-  findUncitedRecommendations,
   metricSourceSchema,
   aiRecommendationSchema,
   type AIInsightResponse,
 } from "../schema";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn } from "../types";
 
 /**
  * Phase C1 — citation-from-data hard requirement.
@@ -148,148 +144,3 @@ describe("citation-from-data — schema-level enforcement", () => {
     expect(result.success).toBe(true);
   });
 });
-
-describe("citation-from-data — cross-check enforcement", () => {
-  it("recommendation citing a (type, timeRange) absent from citations[] is flagged", () => {
-    const orphan: AIInsightResponse = {
-      ...baseValid,
-      recommendations: [
-        {
-          id: "rec-orphan",
-          text: "Cut sodium.",
-          severity: "suggestion",
-          metricSource: {
-            type: "diet",
-            timeRange: "last7days",
-            summary: "no diet logging available",
-          },
-          rationale: {
-            dataWindow: "last7days",
-            comparedTo: "general DGE intake guideline",
-            deviation: "n/a — no salt tracking present",
-          },
-        },
-      ],
-      citations: [baseCitation], // diet citation missing
-    };
-    const missing = findUncitedRecommendations(orphan);
-    expect(missing).toHaveLength(1);
-    expect(missing[0].recommendationId).toBe("rec-orphan");
-    expect(missing[0].missing).toEqual({
-      type: "diet",
-      timeRange: "last7days",
-    });
-  });
-
-  it("partial-key match on type alone is NOT enough — timeRange must also match", () => {
-    const orphan: AIInsightResponse = {
-      ...baseValid,
-      recommendations: [
-        {
-          id: "rec-mismatch",
-          text: "X",
-          severity: "info",
-          metricSource: {
-            type: "bloodPressure",
-            timeRange: "last30days", // citation only covers last7days
-            summary: "avg over 30 days",
-          },
-          rationale: {
-            dataWindow: "last30days",
-            comparedTo: "your 90-day median",
-            deviation: "+2 mmHg over 30-day window",
-          },
-        },
-      ],
-      citations: [baseCitation],
-    };
-    const missing = findUncitedRecommendations(orphan);
-    expect(missing).toHaveLength(1);
-  });
-
-  it("multiple recommendations sharing the same citation pass once that one citation is present", () => {
-    const ok: AIInsightResponse = {
-      ...baseValid,
-      recommendations: [
-        baseValid.recommendations[0],
-        {
-          id: "rec-2",
-          text: "Continue daily logging.",
-          severity: "info",
-          metricSource: { ...baseMetricSource },
-          rationale: baseRationale,
-        },
-      ],
-      citations: [baseCitation],
-    };
-    const missing = findUncitedRecommendations(ok);
-    expect(missing).toEqual([]);
-  });
-});
-
-describe("citation-from-data — wrapper end-to-end", () => {
-  it("uncited recommendation triggers retry; second try with proper citation passes", async () => {
-    const orphan = {
-      ...baseValid,
-      recommendations: [
-        {
-          id: "rec-orphan",
-          text: "Cut sodium.",
-          severity: "suggestion",
-          metricSource: {
-            type: "diet",
-            timeRange: "last7days",
-            summary: "no diet logging available",
-          },
-        },
-      ],
-      citations: [],
-    };
-    const provider = new MockAIProvider({
-      responses: [JSON.stringify(orphan), JSON.stringify(baseValid)],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "system", user: "user" }),
-    );
-    expect(outcome.retried).toBe(true);
-    expect(outcome.parsed.recommendations[0].id).toBe("rec-1");
-  });
-
-  it("retry-attempt that drops the recommendation rather than fixing the citation also passes", async () => {
-    const orphan = {
-      ...baseValid,
-      recommendations: [
-        {
-          id: "rec-orphan",
-          text: "Cut sodium.",
-          severity: "suggestion",
-          metricSource: {
-            type: "diet",
-            timeRange: "last7days",
-            summary: "no diet logging available",
-          },
-        },
-      ],
-      citations: [],
-    };
-    // Second attempt — model decides to drop the unsupported recommendation
-    // entirely. This is the explicitly-allowed "OMIT it" path from the
-    // retry-correction message.
-    const droppedRecs = {
-      summary: "Insufficient data for actionable recommendations.",
-      recommendations: [],
-      citations: [],
-      warnings: [],
-    };
-    const provider = new MockAIProvider({
-      responses: [JSON.stringify(orphan), JSON.stringify(droppedRecs)],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "system", user: "user" }),
-    );
-    expect(outcome.retried).toBe(true);
-    expect(outcome.parsed.recommendations).toHaveLength(0);
-  });
-});
diff --git a/src/lib/ai/__tests__/confidence-schema.test.ts b/src/lib/ai/__tests__/confidence-schema.test.ts
index 2795a6980..64a77271c 100644
--- a/src/lib/ai/__tests__/confidence-schema.test.ts
+++ b/src/lib/ai/__tests__/confidence-schema.test.ts
@@ -5,7 +5,7 @@ import { aiInsightResponseSchema, aiRecommendationSchema } from "../schema";
  * v1.4.16 phase B5d — `recommendation.confidence` schema field.
  *
  * The wrapper overrides the model's confidence with a deterministic
- * server-computed value (`computeConfidence()`), but the schema needs
+ * value the model emits, but the schema needs
  * to ACCEPT a model-supplied number so payloads round-trip cleanly.
  * The field is `.optional()` because:
  *   - The wrapper fills it post-validation regardless.
diff --git a/src/lib/ai/__tests__/confidence-wrapper.test.ts b/src/lib/ai/__tests__/confidence-wrapper.test.ts
deleted file mode 100644
index 81068d00a..000000000
--- a/src/lib/ai/__tests__/confidence-wrapper.test.ts
+++ /dev/null
@@ -1,203 +0,0 @@
-import { describe, it, expect } from "vitest";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn } from "../types";
-
-/**
- * v1.4.16 phase B5d — wrapper post-validation override.
- *
- * The model's self-reported `confidence` is DISCARDED. After zod
- * parse succeeds, `generateInsight()` derives `ConfidenceInputs` per
- * recommendation and overwrites `rec.confidence` with the value
- * `computeConfidence()` produces.
- *
- * Default-on per research §7 question 1 (maintainer-acked). The override
- * happens whether or not the caller supplies a bucket-series-aware
- * resolver — when the resolver is missing we fall back to inputs
- * derivable from the parsed payload alone (`metricSource.n`,
- * recencyDays=0, deviationStdRatio=null) so the contract is "the
- * confidence on the response NEVER comes from the LLM".
- */
-
-const validResponseWithModelConfidence = {
-  summary: "Your blood pressure is trending high.",
-  recommendations: [
-    {
-      id: "rec-1",
-      text: "Discuss lifestyle and possible medication review with your physician.",
-      severity: "important",
-      // Model claims 99 — wrapper must overwrite this.
-      confidence: 99,
-      metricSource: {
-        type: "bloodPressure",
-        timeRange: "last7days",
-        summary: "avg 142/88 across 12 readings",
-        n: 12,
-      },
-      rationale: {
-        dataWindow: "last7days",
-        comparedTo: "your 90-day median (128/82)",
-        deviation: "+14/+6 mmHg above baseline over 7 of 7 days",
-      },
-    },
-  ],
-  citations: [
-    {
-      type: "bloodPressure",
-      timeRange: "last7days",
-      summary: "avg 142/88 across 12 readings",
-    },
-  ],
-  warnings: [],
-};
-
-describe("generateInsight() — confidence override", () => {
-  it("overrides model-supplied confidence with deterministic computation (default fallback)", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(validResponseWithModelConfidence),
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "user" }),
-    );
-    const rec = outcome.parsed.recommendations[0];
-    // Model claimed 99; wrapper must have replaced it.
-    // n=12, recencyDays=0 (default), ratio=null (default) →
-    // nScore = 10 + 10*log10(12) ≈ 20.79
-    // recency = 30 * (1 - 0/30) = 30
-    // signal = null → 15
-    // total ≈ 65.79 → 66
-    expect(rec.confidence).toBe(66);
-    expect(rec.confidence).not.toBe(99);
-  });
-
-  it("fills in confidence even when the model omitted it", async () => {
-    const noConfidence = {
-      ...validResponseWithModelConfidence,
-      recommendations: [
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          confidence: undefined,
-        },
-      ],
-    };
-    // strip the undefined to emulate the model not emitting the key
-    const stringified = JSON.stringify(noConfidence);
-    const provider = new MockAIProvider({ responses: stringified });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "u" }),
-    );
-    expect(outcome.parsed.recommendations[0].confidence).toBe(66);
-  });
-
-  it("derives n from metricSource.n on the parsed payload (no caller injection)", async () => {
-    // v1.4.16 phase D reconcile (simplify F2) — the wrapper used to
-    // accept an optional `confidenceContext` resolver, but no production
-    // caller ever passed one. Inputs are derived from the payload alone
-    // today; v1.4.17's feedback ratchet will re-introduce a resolver
-    // injection when a real second caller exists.
-    const fourteenN = {
-      ...validResponseWithModelConfidence,
-      recommendations: [
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          metricSource: {
-            ...validResponseWithModelConfidence.recommendations[0].metricSource,
-            n: 14,
-          },
-        },
-      ],
-    };
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(fourteenN),
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "s", user: "u" }),
-    );
-    // n=14, recencyDays=0, ratio=null → 10 + 10*log10(14) + 30 + 15 ≈ 66.46
-    expect(outcome.parsed.recommendations[0].confidence).toBe(66);
-  });
-
-  it("hard-caps confidence to <=15 when n<3 (small-data shield)", async () => {
-    const tinyN = {
-      ...validResponseWithModelConfidence,
-      recommendations: [
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          metricSource: {
-            ...validResponseWithModelConfidence.recommendations[0].metricSource,
-            n: 1,
-          },
-        },
-      ],
-    };
-    const provider = new MockAIProvider({ responses: JSON.stringify(tinyN) });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "s", user: "u" }),
-    );
-    expect(outcome.parsed.recommendations[0].confidence).toBeLessThanOrEqual(
-      15,
-    );
-  });
-
-  it("treats missing metricSource.n as n=0 (hard-cap)", async () => {
-    const noN = {
-      ...validResponseWithModelConfidence,
-      recommendations: [
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          metricSource: {
-            type: "bloodPressure",
-            timeRange: "last7days",
-            summary: "avg 142/88 across 12 readings",
-            // n omitted
-          },
-        },
-      ],
-    };
-    const provider = new MockAIProvider({ responses: JSON.stringify(noN) });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "s", user: "u" }),
-    );
-    // n=0 → hard-cap floor 10
-    expect(outcome.parsed.recommendations[0].confidence).toBe(10);
-  });
-
-  it("overrides confidence on every recommendation in payload order", async () => {
-    const twoRecs = {
-      ...validResponseWithModelConfidence,
-      recommendations: [
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          id: "rec-a",
-          confidence: 99,
-        },
-        {
-          ...validResponseWithModelConfidence.recommendations[0],
-          id: "rec-b",
-          confidence: 1,
-          metricSource: {
-            ...validResponseWithModelConfidence.recommendations[0].metricSource,
-            n: 5,
-          },
-        },
-      ],
-    };
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(twoRecs),
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "s", user: "u" }),
-    );
-    // Model claimed (99, 1); both must be discarded and replaced with
-    // the deterministic value (n=12 → 66, n=5 → 62 from the saturating
-    // formula 10 + 10*log10(5) + 30 + 15 ≈ 61.99).
-    expect(outcome.parsed.recommendations[0].confidence).toBe(66);
-    expect(outcome.parsed.recommendations[1].confidence).toBe(62);
-  });
-});
diff --git a/src/lib/ai/__tests__/confidence.test.ts b/src/lib/ai/__tests__/confidence.test.ts
deleted file mode 100644
index e632982e6..000000000
--- a/src/lib/ai/__tests__/confidence.test.ts
+++ /dev/null
@@ -1,150 +0,0 @@
-import { describe, it, expect } from "vitest";
-import { computeConfidence, type ConfidenceInputs } from "../confidence";
-
-/**
- * v1.4.16 phase B5d — `computeConfidence()` deterministic scorer.
- *
- * The wrapper feeds three inputs:
- *   - n: sample count behind the cited data window
- *     (from `metricSource.n` when available; 0 otherwise).
- *   - recencyDays: days since the most recent sample in the window.
- *   - deviationStdRatio: |deviation| / stdev of the user's 90-day
- *     baseline. `null` when the baseline is too thin to compute.
- *
- * Score breakdown (max 100):
- *   - n contribution: log-saturating curve, max 40 at n=30 ish.
- *     n<3 hard-caps the whole score at ≤15 (5 * n).
- *   - recencyScore: 30 at <2d, decays linearly to 0 at 30d, floored at 0.
- *   - signalScore: 30 at |z|≥1.5, 0 at z=0, 15 when ratio is null.
- *
- * 12 fixtures cover the input matrix per research §2.C: small-n hard
- * cap, mid-n + recency interplay, saturated-n + signal extremes,
- * null-ratio neutral path.
- */
-
-interface Fixture {
-  name: string;
-  inputs: ConfidenceInputs;
-  expected: number;
-}
-
-const FIXTURES: Fixture[] = [
-  // n hard-cap: n<3 always returns max(10, 5*n)
-  {
-    name: "n=0 → hard-cap 10 floor",
-    inputs: { n: 0, recencyDays: 0, deviationStdRatio: 2 },
-    expected: 10,
-  },
-  {
-    name: "n=1 → hard-cap 10 floor",
-    inputs: { n: 1, recencyDays: 0, deviationStdRatio: 2 },
-    expected: 10,
-  },
-  {
-    name: "n=2 → hard-cap 10 (5*2)",
-    inputs: { n: 2, recencyDays: 0, deviationStdRatio: 2 },
-    expected: 10,
-  },
-  // n=3: nScore = 10 + 10*log10(3) ≈ 14.77;
-  // recency=1 → 30 * (1 - 1/30) = 29; signal=null → 15
-  // total ≈ 14.77 + 29 + 15 = 58.77 → 59
-  {
-    name: "n=3 fresh sample (1d old), null ratio → 59",
-    inputs: { n: 3, recencyDays: 1, deviationStdRatio: null },
-    expected: 59,
-  },
-  // n=14, recencyDays=1, ratio=2 → research-§2C calibration target ~94
-  // nScore = 10 + 10*log10(14) ≈ 21.46
-  // recencyScore = 30 * (1 - 1/30) = 29
-  // signalScore = min(30, 30 * 2/1.5) = 30
-  // total ≈ 80.46 → 80
-  {
-    name: "n=14, fresh, strong signal → ~80",
-    inputs: { n: 14, recencyDays: 1, deviationStdRatio: 2 },
-    expected: 80,
-  },
-  // n=30, recencyDays=20, null ratio → research-§2C calibration target ~52
-  // nScore = min(40, 10 + 10*log10(30)) = min(40, 24.77) = 24.77
-  // recencyScore = max(0, 30 * (1 - 20/30)) = 10
-  // signalScore = 15 (null)
-  // total ≈ 49.77 → 50
-  {
-    name: "n=30, stale (20d), null ratio → ~50",
-    inputs: { n: 30, recencyDays: 20, deviationStdRatio: null },
-    expected: 50,
-  },
-  // n=100 → nScore = min(40, 10 + 10*log10(100)) = 30 (curve, not cap)
-  // recency=0 → 30, signal=z=1.5 → 30, total 90
-  {
-    name: "n=100, fresh, strong signal → 90",
-    inputs: { n: 100, recencyDays: 0, deviationStdRatio: 1.5 },
-    expected: 90,
-  },
-  // n=10000 → log10 saturates at the 40 cap → max possible total 100
-  {
-    name: "n=10000 saturates the n curve at 40 → 100",
-    inputs: { n: 10000, recencyDays: 0, deviationStdRatio: 2 },
-    expected: 100,
-  },
-  // n=10, recency=30d → recency floors at 0; ratio=0 → signal=0
-  // nScore = 10 + 10*log10(10) = 20; recency=0; signal=0
-  // total = 20
-  {
-    name: "n=10, ancient (30d), zero deviation → 20",
-    inputs: { n: 10, recencyDays: 30, deviationStdRatio: 0 },
-    expected: 20,
-  },
-  // n=10, recency=45d → recencyScore floor 0 (negative clamped)
-  {
-    name: "n=10, very ancient (45d) clamps recency to 0",
-    inputs: { n: 10, recencyDays: 45, deviationStdRatio: null },
-    expected: 35, // 20 + 0 + 15
-  },
-  // n=20, fresh, mild signal (z=0.75 → half of cap)
-  // nScore = 10 + 10*log10(20) ≈ 23.01
-  // recency=0 → 30
-  // signal = 30 * 0.75/1.5 = 15
-  // total ≈ 68.01 → 68
-  {
-    name: "n=20, fresh, mild signal → ~68",
-    inputs: { n: 20, recencyDays: 0, deviationStdRatio: 0.75 },
-    expected: 68,
-  },
-  // Negative ratio: |z|=2 still hits cap of 30
-  {
-    name: "n=14, fresh, strong negative signal (z=-2) → 80",
-    inputs: { n: 14, recencyDays: 1, deviationStdRatio: -2 },
-    expected: 80,
-  },
-  // Edge: n=2, fresh, strong signal — hard-cap still applies
-  {
-    name: "n=2 hard-cap survives even with strong signal + fresh",
-    inputs: { n: 2, recencyDays: 0, deviationStdRatio: 2 },
-    expected: 10,
-  },
-];
-
-describe("computeConfidence()", () => {
-  it.each(FIXTURES)("$name", ({ inputs, expected }) => {
-    expect(computeConfidence(inputs)).toBe(expected);
-  });
-
-  it("returns an integer in [0, 100] for all fixtures", () => {
-    for (const f of FIXTURES) {
-      const v = computeConfidence(f.inputs);
-      expect(Number.isInteger(v)).toBe(true);
-      expect(v).toBeGreaterThanOrEqual(0);
-      expect(v).toBeLessThanOrEqual(100);
-    }
-  });
-
-  it("is monotonic in n when other inputs fixed (more data → never lower)", () => {
-    const baseInputs = { recencyDays: 1, deviationStdRatio: 1 };
-    let prev = -Infinity;
-    for (const n of [5, 10, 20, 50, 100]) {
-      const v = computeConfidence({ ...baseInputs, n });
-      expect(v).toBeGreaterThanOrEqual(prev);
-      prev = v;
-    }
-  });
-});
diff --git a/src/lib/ai/__tests__/generate-insight.test.ts b/src/lib/ai/__tests__/generate-insight.test.ts
deleted file mode 100644
index 392739fe5..000000000
--- a/src/lib/ai/__tests__/generate-insight.test.ts
+++ /dev/null
@@ -1,262 +0,0 @@
-import { describe, it, expect } from "vitest";
-import {
-  aiInsightResponseSchema,
-  findUncitedRecommendations,
-  InsightSchemaError,
-} from "../schema";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn, type CompletionParams } from "../types";
-
-/** Read the last user message's text out of a recorded CompletionParams. */
-function lastUserText(params: CompletionParams): string {
-  for (let i = params.messages.length - 1; i >= 0; i -= 1) {
-    const m = params.messages[i];
-    if (m.role !== "user") continue;
-    return typeof m.content === "string"
-      ? m.content
-      : m.content
-          .filter((p): p is { type: "text"; text: string } => p.type === "text")
-          .map((p) => p.text)
-          .join("");
-  }
-  return "";
-}
-
-const validResponse = {
-  summary: "Your blood pressure is trending high.",
-  recommendations: [
-    {
-      id: "rec-1",
-      text: "Discuss lifestyle and possible medication review with your physician.",
-      severity: "important",
-      metricSource: {
-        type: "bloodPressure",
-        timeRange: "last7days",
-        summary: "avg 142/88 across 12 readings",
-        n: 12,
-      },
-      rationale: {
-        dataWindow: "last7days",
-        comparedTo: "your 90-day median (128/82)",
-        deviation: "+14/+6 mmHg above baseline over 7 of 7 days",
-      },
-    },
-  ],
-  citations: [
-    {
-      type: "bloodPressure",
-      timeRange: "last7days",
-      summary: "avg 142/88 across 12 readings",
-    },
-  ],
-  warnings: [
-    {
-      topic: "blood_pressure",
-      message: "Stage 1 hypertension threshold crossed.",
-      severity: "important",
-    },
-  ],
-};
-
-describe("aiInsightResponseSchema", () => {
-  it("accepts a well-formed strict response", () => {
-    const result = aiInsightResponseSchema.safeParse(validResponse);
-    expect(result.success).toBe(true);
-  });
-
-  it("permits legacy passthrough fields (back-compat)", () => {
-    const withLegacy = {
-      ...validResponse,
-      classification: "grenzwertig",
-      findings: [{ label: "x", value: "y", assessment: "neutral" }],
-      disclaimer: "consult your doctor",
-    };
-    const result = aiInsightResponseSchema.safeParse(withLegacy);
-    expect(result.success).toBe(true);
-  });
-
-  it("rejects when summary is empty", () => {
-    const result = aiInsightResponseSchema.safeParse({
-      ...validResponse,
-      summary: "",
-    });
-    expect(result.success).toBe(false);
-  });
-
-  it("rejects a recommendation without metricSource.summary", () => {
-    const broken = {
-      ...validResponse,
-      recommendations: [
-        {
-          id: "rec-1",
-          text: "Walk more",
-          severity: "suggestion",
-          metricSource: {
-            type: "activity",
-            timeRange: "last7days",
-            summary: "", // empty — should fail
-          },
-        },
-      ],
-    };
-    const result = aiInsightResponseSchema.safeParse(broken);
-    expect(result.success).toBe(false);
-  });
-
-  it("rejects a recommendation with missing severity enum", () => {
-    const broken = {
-      ...validResponse,
-      recommendations: [
-        {
-          id: "rec-1",
-          text: "Walk more",
-          severity: "kinda-bad", // not in enum
-          metricSource: {
-            type: "activity",
-            timeRange: "last7days",
-            summary: "5,000 steps avg",
-          },
-        },
-      ],
-    };
-    const result = aiInsightResponseSchema.safeParse(broken);
-    expect(result.success).toBe(false);
-  });
-
-  it("findUncitedRecommendations flags recommendations whose metricSource is not in citations", () => {
-    const orphan = {
-      ...validResponse,
-      recommendations: [
-        validResponse.recommendations[0],
-        {
-          id: "rec-2",
-          text: "Reduce salt intake",
-          severity: "suggestion",
-          metricSource: {
-            type: "diet",
-            timeRange: "last30days",
-            summary: "no salt logging available",
-          },
-          rationale: {
-            dataWindow: "last30days",
-            comparedTo: "general DGE intake guideline",
-            deviation: "n/a — no salt tracking present",
-          },
-        },
-      ],
-    };
-    const parsed = aiInsightResponseSchema.parse(orphan);
-    const missing = findUncitedRecommendations(parsed);
-    expect(missing).toHaveLength(1);
-    expect(missing[0].recommendationId).toBe("rec-2");
-    expect(missing[0].missing.type).toBe("diet");
-  });
-});
-
-describe("generateInsight wrapper", () => {
-  it("returns the parsed response on first-attempt success", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(validResponse),
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "user" }),
-    );
-    expect(outcome.attempts).toBe(1);
-    expect(outcome.retried).toBe(false);
-    expect(outcome.parsed.summary).toMatch(/blood pressure/);
-    expect(provider.callCount).toBe(1);
-  });
-
-  it("retries once with correction message and succeeds on second attempt", async () => {
-    const provider = new MockAIProvider({
-      responses: ["definitely not json", JSON.stringify(validResponse)],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "Original user prompt." }),
-    );
-    expect(outcome.attempts).toBe(2);
-    expect(outcome.retried).toBe(true);
-    expect(provider.callCount).toBe(2);
-    // The retry call must include the original user prompt PLUS the
-    // corrective suffix referencing the schema.
-    const retryUserPrompt = lastUserText(provider.calls[1]);
-    expect(retryUserPrompt).toContain("Original user prompt.");
-    expect(retryUserPrompt).toContain("did not satisfy");
-    expect(retryUserPrompt).toContain("metricSource");
-  });
-
-  it("throws InsightSchemaError(422) when both attempts fail", async () => {
-    const provider = new MockAIProvider({
-      responses: ["nope", "still nope"],
-    });
-    let caught: unknown;
-    try {
-      await generateInsight(
-        provider,
-        singleUserTurn({ system: "sys", user: "user" }),
-      );
-    } catch (e) {
-      caught = e;
-    }
-    expect(caught).toBeInstanceOf(InsightSchemaError);
-    const err = caught as InsightSchemaError;
-    expect(err.httpStatus).toBe(422);
-    expect(err.attempts).toBe(2);
-    expect(provider.callCount).toBe(2);
-  });
-
-  it("retries when first attempt has uncited recommendations", async () => {
-    const orphanResponse = {
-      ...validResponse,
-      recommendations: [
-        {
-          id: "rec-orphan",
-          text: "Cut sugar",
-          severity: "suggestion",
-          metricSource: {
-            type: "diet",
-            timeRange: "last7days",
-            summary: "no sugar tracking present",
-          },
-          rationale: {
-            dataWindow: "last7days",
-            comparedTo: "general DGE intake guideline",
-            deviation: "n/a — no sugar tracking present",
-          },
-        },
-      ],
-      citations: [], // empty — fails cross-check
-    };
-    const provider = new MockAIProvider({
-      responses: [
-        JSON.stringify(orphanResponse),
-        JSON.stringify(validResponse),
-      ],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "u" }),
-    );
-    expect(outcome.attempts).toBe(2);
-    expect(outcome.retried).toBe(true);
-    expect(lastUserText(provider.calls[1])).toContain(
-      "metricSources not in citations",
-    );
-  });
-
-  it("propagates provider-level errors without retrying", async () => {
-    const provider = new MockAIProvider({
-      rejectWith: Object.assign(new Error("OpenAI request failed (500)"), {
-        httpStatus: 500,
-      }),
-    });
-    await expect(
-      generateInsight(provider, singleUserTurn({ system: "s", user: "u" })),
-    ).rejects.toThrow("OpenAI request failed (500)");
-    // Provider only called once — wrapper does not retry on provider errors.
-    expect(provider.callCount).toBe(1);
-  });
-});
diff --git a/src/lib/ai/__tests__/insight-generator-prompt.test.ts b/src/lib/ai/__tests__/insight-generator-prompt.test.ts
index 7174248f5..4988f48b9 100644
--- a/src/lib/ai/__tests__/insight-generator-prompt.test.ts
+++ b/src/lib/ai/__tests__/insight-generator-prompt.test.ts
@@ -5,10 +5,7 @@ import {
   OUT_OF_SCOPE_REFUSAL_EN,
   OUT_OF_SCOPE_REFUSAL_DE,
 } from "../prompts/insight-generator";
-import { aiInsightResponseSchema, findUncitedRecommendations } from "../schema";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn } from "../types";
+import { aiInsightResponseSchema } from "../schema";
 
 /**
  * Phase C1 — scope-hardened system prompt assertions.
@@ -137,27 +134,4 @@ describe("out-of-scope refusal payloads", () => {
     const result = aiInsightResponseSchema.safeParse(OUT_OF_SCOPE_REFUSAL_DE);
     expect(result.success).toBe(true);
   });
-
-  it("refusal payload has no orphan citations (trivially)", () => {
-    const parsed = aiInsightResponseSchema.parse(OUT_OF_SCOPE_REFUSAL_EN);
-    expect(findUncitedRecommendations(parsed)).toEqual([]);
-  });
-
-  it("wrapper accepts a model returning the refusal payload verbatim", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(OUT_OF_SCOPE_REFUSAL_EN),
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({
-        system: getStrictInsightsSystemPrompt("en"),
-        // Deliberately out-of-scope user prompt — model is instructed to
-        // return the refusal rather than fabricate health metrics.
-        user: "What's the weather in Berlin tomorrow and how do I write Python?",
-      }),
-    );
-    expect(outcome.attempts).toBe(1);
-    expect(outcome.parsed.recommendations).toEqual([]);
-    expect(outcome.parsed.summary).toContain("only summarise");
-  });
 });
diff --git a/src/lib/ai/__tests__/insight-response-schema.test.ts b/src/lib/ai/__tests__/insight-response-schema.test.ts
new file mode 100644
index 000000000..2d28a2c85
--- /dev/null
+++ b/src/lib/ai/__tests__/insight-response-schema.test.ts
@@ -0,0 +1,103 @@
+import { describe, it, expect } from "vitest";
+import { aiInsightResponseSchema } from "../schema";
+const validResponse = {
+  summary: "Your blood pressure is trending high.",
+  recommendations: [
+    {
+      id: "rec-1",
+      text: "Discuss lifestyle and possible medication review with your physician.",
+      severity: "important",
+      metricSource: {
+        type: "bloodPressure",
+        timeRange: "last7days",
+        summary: "avg 142/88 across 12 readings",
+        n: 12,
+      },
+      rationale: {
+        dataWindow: "last7days",
+        comparedTo: "your 90-day median (128/82)",
+        deviation: "+14/+6 mmHg above baseline over 7 of 7 days",
+      },
+    },
+  ],
+  citations: [
+    {
+      type: "bloodPressure",
+      timeRange: "last7days",
+      summary: "avg 142/88 across 12 readings",
+    },
+  ],
+  warnings: [
+    {
+      topic: "blood_pressure",
+      message: "Stage 1 hypertension threshold crossed.",
+      severity: "important",
+    },
+  ],
+};
+
+describe("aiInsightResponseSchema", () => {
+  it("accepts a well-formed strict response", () => {
+    const result = aiInsightResponseSchema.safeParse(validResponse);
+    expect(result.success).toBe(true);
+  });
+
+  it("permits legacy passthrough fields (back-compat)", () => {
+    const withLegacy = {
+      ...validResponse,
+      classification: "grenzwertig",
+      findings: [{ label: "x", value: "y", assessment: "neutral" }],
+      disclaimer: "consult your doctor",
+    };
+    const result = aiInsightResponseSchema.safeParse(withLegacy);
+    expect(result.success).toBe(true);
+  });
+
+  it("rejects when summary is empty", () => {
+    const result = aiInsightResponseSchema.safeParse({
+      ...validResponse,
+      summary: "",
+    });
+    expect(result.success).toBe(false);
+  });
+
+  it("rejects a recommendation without metricSource.summary", () => {
+    const broken = {
+      ...validResponse,
+      recommendations: [
+        {
+          id: "rec-1",
+          text: "Walk more",
+          severity: "suggestion",
+          metricSource: {
+            type: "activity",
+            timeRange: "last7days",
+            summary: "", // empty — should fail
+          },
+        },
+      ],
+    };
+    const result = aiInsightResponseSchema.safeParse(broken);
+    expect(result.success).toBe(false);
+  });
+
+  it("rejects a recommendation with missing severity enum", () => {
+    const broken = {
+      ...validResponse,
+      recommendations: [
+        {
+          id: "rec-1",
+          text: "Walk more",
+          severity: "kinda-bad", // not in enum
+          metricSource: {
+            type: "activity",
+            timeRange: "last7days",
+            summary: "5,000 steps avg",
+          },
+        },
+      ],
+    };
+    const result = aiInsightResponseSchema.safeParse(broken);
+    expect(result.success).toBe(false);
+  });
+});
diff --git a/src/lib/ai/__tests__/provider-runner.test.ts b/src/lib/ai/__tests__/provider-runner.test.ts
index 46a13fc93..b34a5fd1e 100644
--- a/src/lib/ai/__tests__/provider-runner.test.ts
+++ b/src/lib/ai/__tests__/provider-runner.test.ts
@@ -33,7 +33,6 @@ import {
   isHardProviderFailure,
   runRawCompletionWithFallback,
   runStreamingRawCompletionWithFallback,
-  runWithFallback,
 } from "../provider-runner";
 import {
   AUTH_FAILURE_COOLDOWN_MS,
@@ -129,15 +128,6 @@ describe("isHardProviderFailure", () => {
     expect(isHardProviderFailure(err(404))).toBe(true);
   });
 
-  it("does NOT flag an InsightSchemaError as hard — the strict JSON 422 surface still bubbles", async () => {
-    const { InsightSchemaError } = await import("../schema");
-    expect(
-      isHardProviderFailure(
-        new InsightSchemaError("bad json", { attempts: 2 }),
-      ),
-    ).toBe(false);
-  });
-
   it("flags network errors (no httpStatus) as hard", () => {
     expect(isHardProviderFailure(new Error("ECONNRESET"))).toBe(true);
   });
@@ -151,140 +141,7 @@ describe("isHardProviderFailure", () => {
   });
 });
 
-describe("runWithFallback — happy path", () => {
-  it("returns the parsed response from the first working provider", async () => {
-    const codex = new ScriptedProvider({
-      type: "codex",
-      script: [{ ok: true }],
-    });
-    const result = await runWithFallback({
-      userId: "u1",
-      providers: [{ providerType: "codex", instance: codex }],
-      params: singleUserTurn({ system: "s", user: "u" }),
-    });
-    expect(result.parsed.summary).toBe("ok");
-    expect(result.raw.providerType).toBe("codex");
-    expect(result.fallbackHops).toEqual([]);
-    expect(codex.callCount).toBe(1);
-  });
-
-  it("attaches the provider type to the outcome for B5e feedback attribution", async () => {
-    const openai = new ScriptedProvider({
-      type: "admin-key",
-      script: [{ ok: true }],
-    });
-    const result = await runWithFallback({
-      userId: "u1",
-      providers: [{ providerType: "openai", instance: openai }],
-      params: singleUserTurn({ system: "s", user: "u" }),
-    });
-    expect(result.workingProvider.providerType).toBe("openai");
-    expect(result.raw.providerType).toBe("admin-key");
-  });
-});
-
-describe("runWithFallback — primary fails 401, secondary succeeds", () => {
-  it("falls through to the secondary and surfaces the failure reason in the hop log", async () => {
-    const codex = new ScriptedProvider({
-      type: "codex",
-      script: [{ ok: false, error: err(401, "OAuth expired") }],
-    });
-    const openai = new ScriptedProvider({
-      type: "admin-key",
-      script: [{ ok: true }],
-    });
-    const result = await runWithFallback({
-      userId: "u1",
-      providers: [
-        { providerType: "codex", instance: codex },
-        { providerType: "openai", instance: openai },
-      ],
-      params: singleUserTurn({ system: "s", user: "u" }),
-    });
-    expect(result.parsed.summary).toBe("ok");
-    expect(result.workingProvider.providerType).toBe("openai");
-    expect(result.fallbackHops).toHaveLength(1);
-    expect(result.fallbackHops[0]).toMatchObject({
-      providerType: "codex",
-      attempt: 1,
-    });
-    expect(result.fallbackHops[0].failureReason).toMatch(/401/);
-  });
-});
-
-describe("runWithFallback — all-fail cascade", () => {
-  it("throws AllProvidersFailedError when every chain entry fails hard", async () => {
-    const codex = new ScriptedProvider({
-      type: "codex",
-      script: [{ ok: false, error: err(0, "ECONNRESET") }],
-    });
-    const openai = new ScriptedProvider({
-      type: "admin-key",
-      script: [{ ok: false, error: err(503, "upstream down") }],
-    });
-    let caught: unknown;
-    try {
-      await runWithFallback({
-        userId: "u1",
-        providers: [
-          { providerType: "codex", instance: codex },
-          { providerType: "openai", instance: openai },
-        ],
-        params: singleUserTurn({ system: "s", user: "u" }),
-      });
-    } catch (e) {
-      caught = e;
-    }
-    expect(caught).toBeInstanceOf(AllProvidersFailedError);
-    const e = caught as AllProvidersFailedError;
-    expect(e.httpStatus).toBe(503);
-    expect(e.attempts).toHaveLength(2);
-    expect(e.attempts.map((a) => a.providerType)).toEqual(["codex", "openai"]);
-  });
-});
-
-describe("runWithFallback — non-hard errors bubble unchanged", () => {
-  it("re-throws a 422 schema error from the first provider rather than walking", async () => {
-    // schema-mismatch is the wrapper's job, not the runner's. The runner
-    // only walks on transport / auth / 5xx — bad JSON from the first
-    // provider must still propagate as the wrapper's
-    // InsightSchemaError so the user gets the existing 422 surface.
-    const breaking = new ScriptedProvider({
-      type: "codex",
-      // Both wrapper attempts return invalid JSON → wrapper raises
-      // InsightSchemaError(422). The runner does not walk to provider
-      // 2 because that error is not a "hard provider failure".
-      script: [
-        { ok: true, content: "not-json" },
-        { ok: true, content: "still-not-json" },
-      ],
-    });
-    const openai = new ScriptedProvider({
-      type: "admin-key",
-      script: [{ ok: true }],
-    });
-    let caught: unknown;
-    try {
-      await runWithFallback({
-        userId: "u1",
-        providers: [
-          { providerType: "codex", instance: breaking },
-          { providerType: "openai", instance: openai },
-        ],
-        params: singleUserTurn({ system: "s", user: "u" }),
-      });
-    } catch (e) {
-      caught = e;
-    }
-    // Wrapper threw — fallback runner should NOT have escalated to
-    // openai because schema-mismatch is not a hard provider failure.
-    expect(caught).toBeDefined();
-    expect((caught as Error).name).toBe("InsightSchemaError");
-    expect(openai.callCount).toBe(0);
-  });
-});
-
-describe("runWithFallback — last-working cache", () => {
+describe("runRawCompletionWithFallback — last-working cache", () => {
   it("caches the working provider so the next call starts there", async () => {
     const codex = new ScriptedProvider({
       type: "codex",
@@ -296,7 +153,7 @@ describe("runWithFallback — last-working cache", () => {
     });
 
     // First call: codex fails, openai succeeds.
-    await runWithFallback({
+    await runRawCompletionWithFallback({
       userId: "u-cache",
       providers: [
         { providerType: "codex", instance: codex },
@@ -309,7 +166,7 @@ describe("runWithFallback — last-working cache", () => {
 
     // Second call: cache reorders so openai is tried first; codex is
     // never re-invoked even though it appears first in the input chain.
-    await runWithFallback({
+    await runRawCompletionWithFallback({
       userId: "u-cache",
       providers: [
         { providerType: "codex", instance: codex },
@@ -339,7 +196,7 @@ describe("runWithFallback — last-working cache", () => {
     });
 
     // First call: codex fails, openai succeeds, cache = openai.
-    await runWithFallback({
+    await runRawCompletionWithFallback({
       userId: "u-ttl",
       providers: [
         { providerType: "codex", instance: codex },
@@ -355,7 +212,7 @@ describe("runWithFallback — last-working cache", () => {
 
     // Next call walks the original priority order again (codex first,
     // and this time codex's second scripted entry succeeds).
-    await runWithFallback({
+    await runRawCompletionWithFallback({
       userId: "u-ttl",
       providers: [
         { providerType: "codex", instance: codex },
@@ -371,7 +228,7 @@ describe("runWithFallback — last-working cache", () => {
       type: "admin-key",
       script: [{ ok: true }],
     });
-    await runWithFallback({
+    await runRawCompletionWithFallback({
       userId: "u-clear",
       providers: [{ providerType: "openai", instance: ok }],
       params: singleUserTurn({ system: "s", user: "u" }),
@@ -502,11 +359,11 @@ describe("runRawCompletionWithFallback — legacy route shim", () => {
   });
 });
 
-describe("runWithFallback — empty input", () => {
+describe("runRawCompletionWithFallback — empty input", () => {
   it("throws AllProvidersFailedError with httpStatus 422 when zero providers configured", async () => {
     let caught: unknown;
     try {
-      await runWithFallback({
+      await runRawCompletionWithFallback({
         userId: "u-empty",
         providers: [],
         params: singleUserTurn({ system: "s", user: "u" }),
diff --git a/src/lib/ai/__tests__/rationale-schema.test.ts b/src/lib/ai/__tests__/rationale-schema.test.ts
index 0708d6d1a..9bd1eeac0 100644
--- a/src/lib/ai/__tests__/rationale-schema.test.ts
+++ b/src/lib/ai/__tests__/rationale-schema.test.ts
@@ -3,7 +3,6 @@ import {
   aiInsightResponseSchema,
   aiRecommendationRationaleSchema,
   aiRecommendationSchema,
-  findRecommendationsMissingRationale,
   type AIInsightResponse,
 } from "../schema";
 
@@ -195,77 +194,3 @@ describe("aiInsightResponseSchema — rationale integration", () => {
     expect(parsed.recommendations[0].rationale).toEqual(baseRationale);
   });
 });
-
-describe("findRecommendationsMissingRationale()", () => {
-  it("returns an empty array for a payload where every rec has rationale", () => {
-    expect(findRecommendationsMissingRationale(baseValid)).toEqual([]);
-  });
-
-  it("flags a legacy payload whose recs predate B5c (rationale absent)", () => {
-    // Legacy payloads from before B5c had no rationale field. Once
-    // the strict parser flips to mandatory rationale, those payloads
-    // would fail parse — but we keep `.passthrough()` for one
-    // milestone and use `findRecommendationsMissingRationale()` to
-    // *detect* legacy shape so the UI can prompt regeneration.
-    //
-    // Cast through unknown because the strict input type already
-    // requires rationale; this helper exists to handle the legacy
-    // mismatch defensively at runtime.
-    const legacy = {
-      summary: "Older insight from v1.4.15.",
-      recommendations: [
-        {
-          id: "rec-old-1",
-          text: "Walk more",
-          severity: "suggestion",
-          metricSource: {
-            type: "activity",
-            timeRange: "last7days",
-            summary: "5,000 steps avg",
-          },
-          // no rationale
-        },
-        {
-          id: "rec-old-2",
-          text: "Hydrate",
-          severity: "info",
-          metricSource: {
-            type: "fluid",
-            timeRange: "last7days",
-            summary: "low intake logged",
-          },
-          // no rationale
-        },
-      ],
-      citations: [],
-      warnings: [],
-    } as unknown as AIInsightResponse;
-
-    const missing = findRecommendationsMissingRationale(legacy);
-    expect(missing).toEqual(["rec-old-1", "rec-old-2"]);
-  });
-
-  it("flags only the recs that are missing, not the well-formed ones", () => {
-    const partial = {
-      summary: "Mixed payload",
-      recommendations: [
-        baseRec,
-        {
-          id: "rec-bad",
-          text: "Hydrate",
-          severity: "info",
-          metricSource: {
-            type: "fluid",
-            timeRange: "last7days",
-            summary: "low intake logged",
-          },
-          // no rationale
-        },
-      ],
-      citations: [baseCitation],
-      warnings: [],
-    } as unknown as AIInsightResponse;
-
-    expect(findRecommendationsMissingRationale(partial)).toEqual(["rec-bad"]);
-  });
-});
diff --git a/src/lib/ai/__tests__/rationale-wrapper.test.ts b/src/lib/ai/__tests__/rationale-wrapper.test.ts
deleted file mode 100644
index ecf9d78fd..000000000
--- a/src/lib/ai/__tests__/rationale-wrapper.test.ts
+++ /dev/null
@@ -1,172 +0,0 @@
-import { describe, it, expect, beforeEach, vi } from "vitest";
-import { generateInsight } from "../generate-insight";
-import { MockAIProvider } from "../mock-client";
-import { singleUserTurn } from "../types";
-
-const annotateMock = vi.fn();
-
-vi.mock("@/lib/logging/context", () => ({
-  annotate: (fields: { meta?: Record<string, unknown> }) =>
-    annotateMock(fields),
-  getEvent: () => null,
-}));
-
-/**
- * v1.4.16 phase B5c — wrapper-level rationale handling.
- *
- *   1. The corrective retry message lists the rationale fields so a
- *      first-attempt response without rationale gets a targeted
- *      reprompt (instead of the model guessing what's wrong).
- *   2. The wrapper's post-validation `annotate()` carries a
- *      `ai_rationale_coverage_*` breakdown so the admin AI quality
- *      dashboard can chart per-payload coverage of the new field.
- *      Mirrors the citation-coverage annotations from B5a.
- */
-
-const baseRationale = {
-  dataWindow: "last7days" as const,
-  comparedTo: "your 90-day median (122/78)",
-  deviation: "+16/+8 mmHg above baseline over 9 of 9 readings",
-};
-
-const baseMetricSource = {
-  type: "bloodPressure",
-  timeRange: "last7days",
-  summary: "avg 138/86 across 9 readings",
-};
-
-const baseCitation = {
-  type: "bloodPressure",
-  timeRange: "last7days",
-  summary: "avg 138/86 across 9 readings",
-};
-
-const validResponse = {
-  summary: "BP runs slightly above your 90-day median.",
-  recommendations: [
-    {
-      id: "rec-1",
-      text: "Discuss home BP log with your physician.",
-      severity: "important",
-      metricSource: baseMetricSource,
-      rationale: baseRationale,
-    },
-  ],
-  citations: [baseCitation],
-  warnings: [],
-};
-
-describe("generateInsight() — corrective retry mentions rationale", () => {
-  beforeEach(() => {
-    annotateMock.mockReset();
-  });
-
-  it("retries with a corrective message naming dataWindow / comparedTo / deviation when rationale missing", async () => {
-    const recWithoutRationale = {
-      id: "rec-1",
-      text: "Walk more",
-      severity: "suggestion",
-      metricSource: baseMetricSource,
-      // rationale missing
-    };
-    const broken = {
-      ...validResponse,
-      recommendations: [recWithoutRationale],
-    };
-    const provider = new MockAIProvider({
-      responses: [JSON.stringify(broken), JSON.stringify(validResponse)],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "Original user prompt." }),
-    );
-    expect(outcome.attempts).toBe(2);
-    expect(outcome.retried).toBe(true);
-    const retryUserPrompt = provider.calls[1].messages[0].content as string;
-    expect(retryUserPrompt).toContain("Original user prompt.");
-    expect(retryUserPrompt).toContain("rationale");
-    expect(retryUserPrompt).toContain("dataWindow");
-    expect(retryUserPrompt).toContain("comparedTo");
-    expect(retryUserPrompt).toContain("deviation");
-  });
-
-  it("retries when first attempt has empty rationale.deviation", async () => {
-    const recWithEmptyDeviation = {
-      id: "rec-1",
-      text: "Walk more",
-      severity: "suggestion",
-      metricSource: baseMetricSource,
-      rationale: {
-        dataWindow: "last7days",
-        comparedTo: "your 90-day median",
-        deviation: "", // empty — should fail
-      },
-    };
-    const broken = {
-      ...validResponse,
-      recommendations: [recWithEmptyDeviation],
-    };
-    const provider = new MockAIProvider({
-      responses: [JSON.stringify(broken), JSON.stringify(validResponse)],
-    });
-    const outcome = await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "user" }),
-    );
-    expect(outcome.attempts).toBe(2);
-    expect(outcome.retried).toBe(true);
-  });
-});
-
-describe("generateInsight() — rationale-coverage annotation", () => {
-  beforeEach(() => {
-    annotateMock.mockReset();
-  });
-
-  it("annotates ai_rationale_coverage_* fields on a successful generation", async () => {
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(validResponse),
-    });
-    await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "user" }),
-    );
-    expect(annotateMock).toHaveBeenCalled();
-    const meta = annotateMock.mock.calls[0][0]?.meta as Record<string, unknown>;
-    expect(meta).toMatchObject({
-      ai_total_recommendations: 1,
-      ai_rationale_total_recommendations: 1,
-      ai_rationale_with_rationale: 1,
-      ai_rationale_missing_recommendation_ids: [],
-    });
-  });
-
-  it("rationale-coverage = total recommendations when every rec carries one", async () => {
-    const twoRecs = {
-      ...validResponse,
-      recommendations: [
-        validResponse.recommendations[0],
-        {
-          id: "rec-2",
-          text: "Continue daily logging.",
-          severity: "info",
-          metricSource: baseMetricSource,
-          rationale: baseRationale,
-        },
-      ],
-    };
-    const provider = new MockAIProvider({
-      responses: JSON.stringify(twoRecs),
-    });
-    await generateInsight(
-      provider,
-      singleUserTurn({ system: "sys", user: "user" }),
-    );
-    const meta = annotateMock.mock.calls[0][0]?.meta as Record<string, unknown>;
-    expect(meta).toMatchObject({
-      ai_rationale_total_recommendations: 2,
-      ai_rationale_with_rationale: 2,
-      ai_rationale_missing_recommendation_ids: [],
-    });
-  });
-});
diff --git a/src/lib/ai/citation-coverage.ts b/src/lib/ai/citation-coverage.ts
index c6c249bdc..06dea7174 100644
--- a/src/lib/ai/citation-coverage.ts
+++ b/src/lib/ai/citation-coverage.ts
@@ -1,12 +1,20 @@
 /**
  * Reachability note. This module was written for `generateInsight()`, whose
- * only caller is `runWithFallback()` — and `runWithFallback()` has no
- * production caller left (every live surface routes through
- * `runRawCompletionWithFallback` instead). The coverage annotation therefore
- * never executed in production: an audit surface that implied enforcement it
- * did not perform. Rather than delete working, tested logic, the input type is
- * widened to the structural shape both payloads share and the comprehensive
- * post-parse step now calls it, so the admin quality dashboard sees real data.
+ * only caller was `runWithFallback()` — and that had no production caller,
+ * because every live surface routes through `runRawCompletionWithFallback`.
+ * The coverage annotation therefore never executed: an audit surface that
+ * implied enforcement it did not perform.
+ *
+ * v1.30.24 kept the logic and widened the input type to the structural shape
+ * both payloads share, so the comprehensive post-parse step could call it and
+ * the admin quality dashboard could see real data. The dead wrapper and chain
+ * runner around it have since been removed outright, so this module now has
+ * exactly one caller and it is a live one.
+ *
+ * Observational by design: an uncited normative recommendation is an
+ * annotation, never a parse failure. It is a METRIC, not a gate — the citation
+ * ENFORCEMENT that once sat beside it is gone rather than dormant, so nothing
+ * here should be read as withholding output.
  */
 
 /**
diff --git a/src/lib/ai/confidence.ts b/src/lib/ai/confidence.ts
deleted file mode 100644
index d69915fa2..000000000
--- a/src/lib/ai/confidence.ts
+++ /dev/null
@@ -1,102 +0,0 @@
-/**
- * v1.4.16 phase B5d — deterministic confidence-score computation per
- * recommendation.
- *
- * The maintainer's mandate: the model's self-reported confidence is unreliable
- * (calibrated probabilities are not a small-LLM strength), so the
- * wrapper OVERRIDES with a server-computed value derived from three
- * inputs that the cited data window actually exposes:
- *
- *   - n               sample count behind the cited data window
- *                     (`metricSource.n`).
- *   - recencyDays     days since the most recent sample in the window;
- *                     a 7-day rec built off data 25 days old should
- *                     not feel as confident as the same rec built off
- *                     data from this morning.
- *   - deviationStdRatio  |deviation| / stdev of the user's 90-day
- *                     baseline. `null` when the baseline is too thin
- *                     to compute a stdev — we fall back to a neutral
- *                     mid-band score rather than zero.
- *
- * Score components (sum, max 100):
- *
- *   - n contribution: log-saturating curve, capped at 40.
- *     n<3 hard-caps the whole score at ≤15 (5 * n) — three samples is
- *     the hard floor for any quantitative claim per research §2.A
- *     (Apple's "be quiet when unsure" stance).
- *
- *   - recency contribution: 30 at <2d old, decays linearly to 0 at
- *     30d, floored at 0. The shape matches the Oura "Pay Attention"
- *     band — old data is still data, just stale.
- *
- *   - signal contribution: 30 at |z|≥1.5 (clearly out-of-band),
- *     scaled linearly down to 0 at z=0. When ratio is null (no
- *     baseline yet), we award 15 — neither punishing the user for
- *     thin history nor over-claiming a strong deviation we can't
- *     measure.
- *
- * The formula is pre-feedback. v1.4.17 will fit weights to thumbs-up
- * rates (B5e feedback ratchet); for v1.4.16 the formula is fixed and
- * deterministic so admin observability + the "draft pill below 25"
- * UI affordance are reproducible across runs.
- */
-
-export interface ConfidenceInputs {
-  /** Sample count behind the cited data window (from metricSource.n). */
-  n: number;
-  /** Days since the most recent sample in the window. */
-  recencyDays: number;
-  /** |deviation| / stdev of the user's 90-day baseline, when computable. */
-  deviationStdRatio: number | null;
-}
-
-const HARD_CAP_FLOOR = 10;
-const N_CONTRIBUTION_MAX = 40;
-const RECENCY_CONTRIBUTION_MAX = 30;
-const RECENCY_DECAY_DAYS = 30;
-const SIGNAL_CONTRIBUTION_MAX = 30;
-const SIGNAL_SATURATION_RATIO = 1.5;
-const SIGNAL_NULL_NEUTRAL = 15;
-
-/**
- * Compute the deterministic confidence score (0..100) for a single
- * recommendation. Pure function — no I/O, no clock; safe to call from
- * test fixtures and from `generateInsight()` post-validation alike.
- */
-export function computeConfidence(inputs: ConfidenceInputs): number {
-  const { n, recencyDays, deviationStdRatio } = inputs;
-
-  // n hard cap: <3 samples means we cannot make a quantitative claim,
-  // so the whole score is bounded at 5*n (max 15 at n=2). Floor at
-  // 10 so n=0/n=1 still produce a non-zero "draft" signal — the UI
-  // shows the rec but tags it as low-confidence rather than hiding it.
-  if (n < 3) {
-    return Math.max(HARD_CAP_FLOOR, 5 * n);
-  }
-
-  // n contribution: log10 saturates so a 100-sample window doesn't
-  // outweigh a 30-sample window by 3×; both are "enough data" for the
-  // user-level signal we care about.
-  const nScore = Math.min(N_CONTRIBUTION_MAX, 10 + 10 * Math.log10(n));
-
-  // Recency contribution: <2d ≈ full score (30 * (1 - 0/30) = 30,
-  // 30 * (1 - 1/30) = 29; both round into the meter's top band).
-  // 30d+ floors to 0 so an ancient window cannot inflate the score.
-  const recencyScore = Math.max(
-    0,
-    RECENCY_CONTRIBUTION_MAX * (1 - recencyDays / RECENCY_DECAY_DAYS),
-  );
-
-  // Signal contribution: null ratio (no baseline) → neutral 15.
-  // Otherwise scale |z| linearly to the SIGNAL_SATURATION_RATIO cap.
-  const signalScore =
-    deviationStdRatio === null
-      ? SIGNAL_NULL_NEUTRAL
-      : Math.min(
-          SIGNAL_CONTRIBUTION_MAX,
-          SIGNAL_CONTRIBUTION_MAX *
-            (Math.abs(deviationStdRatio) / SIGNAL_SATURATION_RATIO),
-        );
-
-  return Math.round(nScore + recencyScore + signalScore);
-}
diff --git a/src/lib/ai/generate-insight.ts b/src/lib/ai/generate-insight.ts
deleted file mode 100644
index c861b790d..000000000
--- a/src/lib/ai/generate-insight.ts
+++ /dev/null
@@ -1,279 +0,0 @@
-import type { z } from "zod/v4";
-import {
-  appendToLastUserMessage,
-  type AIProvider,
-  type CompletionParams,
-  type CompletionResult,
-} from "./types";
-import {
-  aiInsightResponseSchema,
-  findRecommendationsMissingRationale,
-  findUncitedRecommendations,
-  InsightSchemaError,
-  type AIInsightResponse,
-} from "./schema";
-import { computeCitationCoverage } from "./citation-coverage";
-import { computeConfidence } from "./confidence";
-import { annotate } from "@/lib/logging/context";
-
-/**
- * Schema-enforced wrapper around `AIProvider.generateCompletion()`.
- *
- * Phase C1 (v1.4.15): the maintainer's mandate ("zero hallucinations, must
- * ground on user data") means a free-form LLM response is no longer
- * acceptable. This wrapper:
- *
- *   1. Calls the provider once.
- *   2. Parses the response as JSON.
- *   3. Validates it against `aiInsightResponseSchema` AND cross-checks
- *      that every recommendation's `metricSource` appears in
- *      `citations[]`.
- *   4. On failure: retries ONCE, prepending a corrective system
- *      message that includes the violated zod issues.
- *   5. On second failure: throws `InsightSchemaError` (httpStatus 422).
- *
- * The wrapper does not concern itself with provider-level retries
- * (HTTP 5xx, 401 token-refresh, etc.) — each provider client handles
- * those internally. This wrapper retries only on schema-violation.
- */
-
-export interface GenerateInsightOutcome {
-  parsed: AIInsightResponse;
-  raw: CompletionResult;
-  /** Number of provider calls made (1 on success, 2 on retry). */
-  attempts: number;
-  /** True iff the retry attempt produced the parsed response. */
-  retried: boolean;
-}
-
-/**
- * Build the corrective user-prompt suffix that explains to the model
- * what went wrong on the first attempt. Includes the violated zod
- * issues verbatim (truncated to 1KB) so the model can self-correct.
- */
-export function buildRetryCorrectionMessage(
-  reason: string,
-  details: string,
-): string {
-  return `
-Your previous response did not satisfy the required JSON schema.
-Reason: ${reason}
-Details: ${details.slice(0, 1024)}
-
-You MUST return a JSON object matching this schema exactly. Do NOT
-include prose, markdown fences, or commentary outside the JSON.
-
-Required top-level fields:
-  - summary: string (2-3 sentences)
-  - recommendations: array of objects with
-      { id: string,
-        text: string,
-        severity: "info" | "suggestion" | "important" | "urgent",
-        metricSource: { type: string, timeRange: string, summary: string, n?: number },
-        rationale: { dataWindow: "last7days" | "last30days" | "last90days" | "allTime",
-                     comparedTo: non-empty string,
-                     deviation: non-empty string } }
-  - citations: array of objects with { type: string, timeRange: string, summary: string }
-  - warnings: array of objects with { topic: string, message: string, severity?: same enum }
-
-Every recommendation's metricSource MUST also appear in citations[]
-(same type + timeRange). Every recommendation MUST carry a
-rationale object — dataWindow, comparedTo, and deviation are all
-required and rationale.dataWindow MUST equal metricSource.timeRange.
-If you cannot ground a recommendation in user data from the
-snapshot you were given, OMIT it. An empty recommendations[] is
-acceptable.
-`;
-}
-
-interface TryOutcome {
-  ok: boolean;
-  parsed?: AIInsightResponse;
-  raw: CompletionResult;
-  reason?: string;
-  details?: string;
-  zodIssues?: z.ZodIssue[];
-}
-
-async function tryOnce(
-  provider: AIProvider,
-  params: CompletionParams,
-): Promise<TryOutcome> {
-  const raw = await provider.generateCompletion(params);
-
-  // Parse JSON.
-  let json: unknown;
-  try {
-    json = JSON.parse(raw.content);
-  } catch (e) {
-    return {
-      ok: false,
-      raw,
-      reason: "Response was not valid JSON",
-      details: (e as Error).message,
-    };
-  }
-
-  // Schema validate.
-  const result = aiInsightResponseSchema.safeParse(json);
-  if (!result.success) {
-    return {
-      ok: false,
-      raw,
-      reason: "Response did not match the required JSON schema",
-      details: result.error.issues
-        .slice(0, 8)
-        .map((i) => `${i.path.join(".")}: ${i.message}`)
-        .join("; "),
-      zodIssues: result.error.issues,
-    };
-  }
-
-  // Cross-check citations.
-  const uncited = findUncitedRecommendations(result.data);
-  if (uncited.length > 0) {
-    const list = uncited
-      .slice(0, 4)
-      .map(
-        (u) =>
-          `${u.recommendationId} → ${u.missing.type}/${u.missing.timeRange}`,
-      )
-      .join(", ");
-    return {
-      ok: false,
-      raw,
-      reason: "Recommendations referenced metricSources not in citations[]",
-      details: `Missing citations: ${list}`,
-    };
-  }
-
-  return { ok: true, parsed: result.data, raw };
-}
-
-/**
- * Forward the citation-coverage breakdown into the active Wide-Event
- * via `annotate()` (no-op outside a request context). Keys are
- * snake_case to match the rest of the AI Wide-Event vocabulary.
- *
- * v1.4.16 phase B5c — additionally emit `ai_rationale_*` keys so the
- * admin AI quality dashboard can chart rationale-coverage per
- * payload alongside citation-coverage. Rationale is required by the
- * strict schema; the missing-ids list will only ever populate when
- * a legacy payload sneaks through `.passthrough()` — useful as an
- * observability tripwire for the migration window.
- */
-function annotateCitationCoverage(parsed: AIInsightResponse): void {
-  const coverage = computeCitationCoverage(parsed);
-  const rationaleMissing = findRecommendationsMissingRationale(parsed);
-  annotate({
-    meta: {
-      ai_total_recommendations: coverage.totalRecommendations,
-      ai_normative_recommendations: coverage.normativeRecommendations,
-      ai_cited_normative_recommendations:
-        coverage.citedNormativeRecommendations,
-      ai_uncited_normative_recommendation_ids:
-        coverage.uncitedNormativeRecommendationIds,
-      ai_rationale_total_recommendations: coverage.totalRecommendations,
-      ai_rationale_with_rationale:
-        coverage.totalRecommendations - rationaleMissing.length,
-      ai_rationale_missing_recommendation_ids: rationaleMissing,
-    },
-  });
-}
-
-/**
- * Apply the deterministic confidence override to every recommendation.
- * The model's value (if any) is REPLACED. Inputs are derived from the
- * parsed payload alone (n from `metricSource.n`, recencyDays=0,
- * deviationStdRatio=null) — the route layer doesn't pass bucket-series
- * data through here today. v1.4.17's feedback ratchet will need a
- * resolver injection; re-introduce one when the second caller exists.
- *
- * Emits a Wide-Event with per-rec model-vs-computed pairs so admin
- * observability can chart drift between what the LLM thinks and what
- * the formula produces.
- */
-function applyConfidenceOverride(parsed: AIInsightResponse): void {
-  const overrides: Array<{
-    id: string;
-    model: number | null;
-    computed: number;
-  }> = [];
-  for (const rec of parsed.recommendations) {
-    const computed = computeConfidence({
-      n: rec.metricSource.n ?? 0,
-      recencyDays: 0,
-      deviationStdRatio: null,
-    });
-    const modelValue =
-      typeof rec.confidence === "number" ? rec.confidence : null;
-    rec.confidence = computed;
-    overrides.push({ id: rec.id, model: modelValue, computed });
-  }
-  if (overrides.length > 0) {
-    annotate({
-      meta: {
-        ai_confidence_override_event: "confidence.override.applied",
-        ai_confidence_override_count: overrides.length,
-        ai_confidence_override_pairs: overrides,
-      },
-    });
-  }
-}
-
-/**
- * Run a schema-enforced insight generation. Up to 2 provider calls.
- * Throws `InsightSchemaError` (httpStatus 422) if both attempts fail
- * the schema; provider-level errors bubble untouched.
- *
- * v1.4.16 phase B5d — after a successful parse, every recommendation's
- * `confidence` field is overwritten with the deterministic
- * `computeConfidence()` value. The model's number (if any) is
- * discarded; admin observability sees both via the
- * `confidence.override.applied` Wide-Event annotation.
- */
-export async function generateInsight(
-  provider: AIProvider,
-  params: CompletionParams,
-): Promise<GenerateInsightOutcome> {
-  const first = await tryOnce(provider, params);
-  if (first.ok && first.parsed) {
-    annotateCitationCoverage(first.parsed);
-    applyConfidenceOverride(first.parsed);
-    return {
-      parsed: first.parsed,
-      raw: first.raw,
-      attempts: 1,
-      retried: false,
-    };
-  }
-
-  // Retry-once with corrective context appended to the user prompt.
-  const correction = buildRetryCorrectionMessage(
-    first.reason ?? "schema mismatch",
-    first.details ?? "",
-  );
-  const retryParams: CompletionParams = appendToLastUserMessage(
-    params,
-    `\n\n${correction}`,
-  );
-  const second = await tryOnce(provider, retryParams);
-  if (second.ok && second.parsed) {
-    annotateCitationCoverage(second.parsed);
-    applyConfidenceOverride(second.parsed);
-    return {
-      parsed: second.parsed,
-      raw: second.raw,
-      attempts: 2,
-      retried: true,
-    };
-  }
-
-  throw new InsightSchemaError(
-    second.reason ?? first.reason ?? "Schema validation failed twice",
-    {
-      issues: second.zodIssues ?? first.zodIssues,
-      attempts: 2,
-    },
-  );
-}
diff --git a/src/lib/ai/prompts/insight-generator.ts b/src/lib/ai/prompts/insight-generator.ts
index 184367af3..bc96be96b 100644
--- a/src/lib/ai/prompts/insight-generator.ts
+++ b/src/lib/ai/prompts/insight-generator.ts
@@ -986,7 +986,7 @@ const INSIGHTS_LOCALE_REPLY_FOOTER_FALLBACK: Record<
 
 /**
  * Returns the active scope-hardened system prompt for a given locale.
- * The route consumes this via `generateInsight()`.
+ * Consumed by the comprehensive generation path.
  */
 export function getStrictInsightsSystemPrompt(locale: Locale): string {
   if (locale === "de") return SYSTEM_PROMPT_DE;
diff --git a/src/lib/ai/provider-runner.ts b/src/lib/ai/provider-runner.ts
index d10faed99..cc62c83da 100644
--- a/src/lib/ai/provider-runner.ts
+++ b/src/lib/ai/provider-runner.ts
@@ -1,9 +1,4 @@
 import type { AIProvider, CompletionParams, CompletionResult } from "./types";
-import {
-  generateInsight,
-  type GenerateInsightOutcome,
-} from "./generate-insight";
-import { InsightSchemaError } from "./schema";
 import type { ProviderChainType } from "./provider-chain";
 import {
   postgresProviderHealthLedger,
@@ -25,19 +20,17 @@ const NEVER_SKIP: ReadonlySet<ProviderChainType> = new Set(["local"]);
 /**
  * v1.4.16 phase B5b — multi-provider redundancy runner.
  *
- * The chain runner wraps `generateInsight()` with a hard-failure
+ * The chain runner wraps a provider completion with a hard-failure
  * walkthrough across an ordered list of providers. The motivation is
  * the maintainer's "ein Provider-Ausfall darf die Insights nicht zerschiessen"
  * mandate: a 401 from Codex (OAuth expired), a 503 from OpenAI
  * (upstream brown-out), or a network reset MUST cascade to the next
  * configured provider rather than 422-ing the whole request.
  *
- * Hard failures (`isHardProviderFailure`) cascade. Schema failures
- * (the existing 422 surface) do NOT cascade — they bubble through
- * `InsightSchemaError` so the user still sees the helpful "the model
- * returned malformed JSON" diagnostic. Walking on schema errors would
- * mask provider-specific prompt-following issues that the v1.4.15
- * citation-enforcement guard exists to surface.
+ * Hard failures (`isHardProviderFailure`) cascade. A malformed-JSON
+ * reply is NOT a hard failure — it is a prompt-following issue, so it
+ * bubbles to the caller's own validator rather than masking itself as
+ * provider unavailability and walking the chain.
  *
  * Cross-feature coupling:
  *   - B5a's `MEDICAL_REFERENCES` is shared via the system prompt; each
@@ -120,9 +113,9 @@ function rememberWorkingProvider(
  *   - Network errors (no `httpStatus`) — DNS, ECONNRESET, timeout.
  *
  * No:
- *   - `InsightSchemaError` — surfaces the wrapper's existing 422 so the
- *     user sees "model JSON malformed" rather than getting silent
- *     provider drift. This is the ONLY non-hard class.
+ *   - Anything with a non-error status. Callers that parse a structured
+ *     reply raise their own validation error, which carries no
+ *     `httpStatus` from a provider wire and is theirs to surface.
  *
  * v1.21.3 — a 4xx that the PROVIDER WIRE returned (tagged `upstream`,
  * e.g. a Codex `400` rejecting the tool-call / structured request shape)
@@ -132,12 +125,9 @@ function rememberWorkingProvider(
  * provider error rethrew as an HTTP 500. Treating every upstream 4xx as
  * hard means it cascades to the next provider and, on exhaustion, is
  * wrapped in `AllProvidersFailedError` so the route surfaces a graceful
- * `coach.provider.*` frame instead of a 500. `InsightSchemaError`
- * (no `upstream` tag) is still excluded, so the strict JSON surface keeps
- * its 422 "model JSON malformed" diagnostic.
+ * `coach.provider.*` frame instead of a 500.
  */
 export function isHardProviderFailure(error: unknown): boolean {
-  if (error instanceof InsightSchemaError) return false;
   const err = error as { httpStatus?: number; name?: string } | null;
   const status = err?.httpStatus;
   if (typeof status !== "number" || status <= 0) {
@@ -169,28 +159,6 @@ export interface FallbackHop {
   httpStatus: number | null;
 }
 
-export interface RunWithFallbackParams {
-  userId: string;
-  providers: ProviderChainResolved[];
-  params: CompletionParams;
-  /**
-   * v1.11.0 W1 — durable provider-health ledger. Defaults to the
-   * Postgres-backed implementation in production; unit tests of the pure
-   * chain pass an in-memory / no-op ledger to avoid standing up a DB.
-   * The ledger is read-through (skip-hint reorder) + write-through
-   * (record outcome); it never gates generation — a ledger error always
-   * fails open to today's behaviour.
-   */
-  ledger?: ProviderHealthLedger;
-}
-
-export interface RunWithFallbackResult extends GenerateInsightOutcome {
-  /** Provider that actually produced the parsed response. */
-  workingProvider: ProviderChainResolved;
-  /** Hops that failed before the working provider succeeded. */
-  fallbackHops: FallbackHop[];
-}
-
 /**
  * Hard-fail wrapper around `AllProvidersFailedError.attempts` so the
  * route layer can pick a single representative status (the worst of
@@ -392,78 +360,6 @@ export function classifyErrorBody(body: unknown): string | null {
   return `unclassified:${body.length}chars`;
 }
 
-/**
- * Run an insight generation across an ordered chain of providers,
- * walking past hard provider failures. Schema-class failures (422)
- * bubble unchanged from the underlying wrapper.
- */
-export async function runWithFallback(
-  args: RunWithFallbackParams,
-): Promise<RunWithFallbackResult> {
-  const { userId, providers, params } = args;
-  const ledger = args.ledger ?? postgresProviderHealthLedger;
-
-  if (providers.length === 0) {
-    throw new AllProvidersFailedError([]);
-  }
-
-  const ordered = await resolveChainOrder(userId, providers, ledger);
-  const hops: FallbackHop[] = [];
-
-  for (let i = 0; i < ordered.length; i += 1) {
-    const candidate = ordered[i];
-    try {
-      const outcome = await generateInsight(candidate.instance, params);
-      rememberWorkingProvider(userId, candidate.providerType);
-      void ledger.recordSuccess(userId, candidate.providerType);
-      annotate({
-        meta: {
-          ai_chain_working_provider: candidate.providerType,
-          ai_chain_fallback_count: hops.length,
-        },
-      });
-      return {
-        ...outcome,
-        workingProvider: candidate,
-        fallbackHops: hops,
-      };
-    } catch (error) {
-      if (!isHardProviderFailure(error)) {
-        // Schema/validation error — bubble immediately. NOT recorded in
-        // the health ledger: a malformed-JSON reply is a prompt-following
-        // issue, not a provider-availability failure.
-        throw error;
-      }
-      const summary = summariseError(error);
-      void ledger.recordFailure(userId, candidate.providerType, summary.status);
-      const hop: FallbackHop = {
-        providerType: candidate.providerType,
-        attempt: i + 1,
-        failureReason: summary.reason,
-        httpStatus: summary.status,
-      };
-      hops.push(hop);
-      annotate({
-        meta: {
-          [`ai_chain_hop_${i + 1}_provider`]: candidate.providerType,
-          [`ai_chain_hop_${i + 1}_status`]: summary.status,
-          [`ai_chain_hop_${i + 1}_reason`]: summary.reason.slice(0, 240),
-          [`ai_chain_hop_${i + 1}_body`]: summary.bodyExcerpt,
-        },
-      });
-    }
-  }
-
-  // Every chain entry hard-failed.
-  annotate({
-    meta: {
-      ai_chain_outcome: "all-failed",
-      ai_chain_fallback_count: hops.length,
-    },
-  });
-  throw new AllProvidersFailedError(hops);
-}
-
 /**
  * Shared chain walker for the two RAW runners (streaming + buffered). Owns the
  * one set of chain / fallback / health-ledger semantics — empty-chain guard,
@@ -593,14 +489,15 @@ export interface RunRawWithFallbackResult {
 }
 
 /**
- * Sibling of `runWithFallback()` that calls `provider.generateCompletion`
- * directly (no strict-schema wrapper). Used by the legacy route at
- * `/api/insights/generate` which still consumes the rich legacy shape.
+ * Buffered runner: calls `provider.generateCompletion` directly and hands the
+ * raw reply back for the caller to validate. This is the ONLY buffered chain
+ * runner — a strict-schema sibling (`runWithFallback`) once sat beside it but
+ * had no production caller and was removed.
  *
- * Fallback policy is identical: hard provider failures cascade,
- * everything else (e.g. a 4xx-but-non-auth from a custom provider)
- * bubbles to the caller. The cache + structured logging are shared
- * with the strict variant.
+ * Fallback policy: hard provider failures cascade, everything else (e.g. a
+ * 4xx-but-non-auth from a custom provider) bubbles to the caller. The
+ * last-working cache + structured logging are shared with the streaming
+ * sibling via `runRawChain`.
  */
 export async function runRawCompletionWithFallback(args: {
   userId: string;
diff --git a/src/lib/ai/retry-correction.ts b/src/lib/ai/retry-correction.ts
new file mode 100644
index 000000000..26c912b49
--- /dev/null
+++ b/src/lib/ai/retry-correction.ts
@@ -0,0 +1,60 @@
+/**
+ * The corrective retry prompt for a structured-output generation that came
+ * back malformed.
+ *
+ * History worth keeping, because the shape of this module is a consequence of
+ * it: this file used to hold `generateInsight()`, a schema-enforced wrapper
+ * that parsed a reply against the strict `aiInsightResponseSchema`,
+ * cross-checked that every recommendation's `metricSource` appeared in
+ * `citations[]`, retried once with the message below, and threw
+ * `InsightSchemaError` on a second failure. That wrapper's only caller was
+ * `runWithFallback()`, which itself had no production caller — every live
+ * surface runs `runRawCompletionWithFallback` and validates with its own
+ * schema. So the enforcement never executed. It was removed rather than wired
+ * in: the live `insightResultSchema` has no `citations[]` field at all and
+ * treats `metricSource` as optional, so applying the strict cross-check to a
+ * live payload would have rejected essentially every generation.
+ *
+ * What survived is the part that was always reachable — the corrective
+ * message, which `comprehensive-generate.ts` sends on its own JSON-retry leg.
+ */
+
+/**
+ * Build the corrective user-prompt suffix that explains to the model
+ * what went wrong on the first attempt. Includes the violated zod
+ * issues verbatim (truncated to 1KB) so the model can self-correct.
+ */
+export function buildRetryCorrectionMessage(
+  reason: string,
+  details: string,
+): string {
+  return `
+Your previous response did not satisfy the required JSON schema.
+Reason: ${reason}
+Details: ${details.slice(0, 1024)}
+
+You MUST return a JSON object matching this schema exactly. Do NOT
+include prose, markdown fences, or commentary outside the JSON.
+
+Required top-level fields:
+  - summary: string (2-3 sentences)
+  - recommendations: array of objects with
+      { id: string,
+        text: string,
+        severity: "info" | "suggestion" | "important" | "urgent",
+        metricSource: { type: string, timeRange: string, summary: string, n?: number },
+        rationale: { dataWindow: "last7days" | "last30days" | "last90days" | "allTime",
+                     comparedTo: non-empty string,
+                     deviation: non-empty string } }
+  - citations: array of objects with { type: string, timeRange: string, summary: string }
+  - warnings: array of objects with { topic: string, message: string, severity?: same enum }
+
+Every recommendation's metricSource MUST also appear in citations[]
+(same type + timeRange). Every recommendation MUST carry a
+rationale object — dataWindow, comparedTo, and deviation are all
+required and rationale.dataWindow MUST equal metricSource.timeRange.
+If you cannot ground a recommendation in user data from the
+snapshot you were given, OMIT it. An empty recommendations[] is
+acceptable.
+`;
+}
diff --git a/src/lib/ai/schema.ts b/src/lib/ai/schema.ts
index 7a107c407..994b53fb4 100644
--- a/src/lib/ai/schema.ts
+++ b/src/lib/ai/schema.ts
@@ -18,7 +18,7 @@ import { MEDICAL_REFERENCE_IDS } from "./medical-references";
  *   - A `citations[]` array referencing the metrics the response drew
  *     on. May be empty when `recommendations[]` is empty (no data → no
  *     citation), but every recommendation's `metricSource` MUST also
- *     appear in `citations[]` (cross-check enforced by `validateInsightResponse`).
+ *     appear in `citations[]`.
  *   - A `warnings[]` array for guideline-flagged values (high BP,
  *     critical pulse, etc.). Free-form for now; v1.4.16 adds severity
  *     enum + medical-reference grounding.
@@ -137,9 +137,8 @@ export const aiRecommendationSchema = z
     /**
      * v1.4.16 phase B5c — per-recommendation explainability rationale.
      * Required on every rec so the UI's expand-card always has WHY +
-     * WINDOW + COMPARED-TO content to render. Legacy payloads (pre-
-     * B5c) ride through `findRecommendationsMissingRationale()` for
-     * the regenerate-CTA migration; the strict parser rejects them.
+     * WINDOW + COMPARED-TO content to render. The strict parser
+     * rejects a legacy (pre-B5c) payload that omits it.
      */
     rationale: aiRecommendationRationaleSchema,
     /**
@@ -159,11 +158,9 @@ export const aiRecommendationSchema = z
      *
      * Optional at parse-time:
      *   - The LLM may emit a value (we accept it so the payload round-
-     *     trips cleanly), but `generateInsight()` OVERRIDES with the
-     *     server-computed `computeConfidence()` post-validation. The
-     *     model's number is discarded — calibrated probabilities are
-     *     not a small-LLM strength and the deterministic path keeps
-     *     the v1.4.17 feedback ratchet reproducible.
+     *     trips cleanly). A deterministic server-side override once
+     *     replaced it post-validation, but that path never ran in
+     *     production and was removed with the rest of the wrapper.
      *   - Legacy cached payloads from before B5d landed have no
      *     confidence field; the meter falls back to a "draft" pill.
      *
@@ -507,87 +504,19 @@ export const aiInsightResponseSchema = z
 export type AIInsightResponse = z.infer<typeof aiInsightResponseSchema>;
 
 /**
- * Tagged error thrown by the schema-enforcement wrapper when even the
- * retry attempt fails to produce a valid payload. The route catches
- * this and surfaces a 422 to the client.
- */
-export class InsightSchemaError extends Error {
-  readonly httpStatus = 422;
-  readonly issues: z.ZodIssue[] | null;
-  readonly attempts: number;
-
-  constructor(
-    message: string,
-    options: { issues?: z.ZodIssue[]; attempts: number },
-  ) {
-    super(message);
-    this.name = "InsightSchemaError";
-    this.issues = options.issues ?? null;
-    this.attempts = options.attempts;
-  }
-}
-
-/**
- * v1.4.16 phase B5c — legacy-payload detector.
+ * The strict-contract ENFORCEMENT that used to live here is gone: an
+ * `InsightSchemaError` class no caller threw, a `findUncitedRecommendations`
+ * cross-check (every recommendation's `metricSource` must appear in
+ * `citations[]`) and a `findRecommendationsMissingRationale` legacy-payload
+ * detector. All three existed only for `generateInsight()`, whose sole caller
+ * `runWithFallback()` had no production caller — so none of them ever ran.
  *
- * The strict schema now requires `rationale` on every recommendation.
- * Cached payloads from v1.4.14/v1.4.15 predate the field — they would
- * fail `aiInsightResponseSchema.safeParse()` outright. Rather than
- * auto-regenerating on every read (expensive + surprising), we let
- * the route call this helper against the legacy-shape JSON so the
- * UI can show a "Insights updated — regenerate for new explainability
- * features" CTA. User-initiated regeneration stays the trigger.
- *
- * Input is intentionally typed loosely (the runtime shape is the
- * canonical AIInsightResponse, but legacy payloads omit `rationale`
- * which the static type now requires). Returns the ids of
- * recommendations missing rationale; an empty array means the
- * payload is well-shaped under B5c.
+ * The schema above stays because it is not a guard: it is the shape the
+ * corrective retry prompt (`retry-correction.ts`) describes to the model, and
+ * it is the reference definition for the strict payload. NOTE the live
+ * comprehensive path validates against the looser `insightResultSchema` in
+ * `types.ts`, which has no `citations[]` and treats `metricSource` as
+ * optional — so the retry prompt asks for a stricter shape than the live
+ * validator requires. That mismatch predates this cleanup and is left alone
+ * deliberately; changing the retry prompt changes model output.
  */
-export function findRecommendationsMissingRationale(
-  parsed: AIInsightResponse,
-): string[] {
-  const missing: string[] = [];
-  for (const rec of parsed.recommendations) {
-    // Defensive runtime check — the static type asserts rationale is
-    // present, but the helper exists precisely for legacy payloads
-    // where it isn't.
-    const r = (rec as { rationale?: unknown }).rationale;
-    if (r === undefined || r === null) {
-      missing.push(rec.id);
-    }
-  }
-  return missing;
-}
-
-/**
- * Cross-validation beyond zod: every recommendation's `metricSource`
- * must also appear in `citations[]` (same `type` + `timeRange`).
- * Returns the list of missing-citation issues; an empty list means
- * the response is internally consistent.
- */
-export function findUncitedRecommendations(parsed: AIInsightResponse): Array<{
-  recommendationId: string;
-  missing: { type: string; timeRange: string };
-}> {
-  const cited = new Set(
-    parsed.citations.map((c) => `${c.type}::${c.timeRange}`),
-  );
-  const missing: Array<{
-    recommendationId: string;
-    missing: { type: string; timeRange: string };
-  }> = [];
-  for (const rec of parsed.recommendations) {
-    const key = `${rec.metricSource.type}::${rec.metricSource.timeRange}`;
-    if (!cited.has(key)) {
-      missing.push({
-        recommendationId: rec.id,
-        missing: {
-          type: rec.metricSource.type,
-          timeRange: rec.metricSource.timeRange,
-        },
-      });
-    }
-  }
-  return missing;
-}
diff --git a/src/lib/insights/comprehensive-generate.ts b/src/lib/insights/comprehensive-generate.ts
index ecbb94c6c..943b4e17a 100644
--- a/src/lib/insights/comprehensive-generate.ts
+++ b/src/lib/insights/comprehensive-generate.ts
@@ -53,7 +53,7 @@ import {
   buildSystemPromptWithReferences,
   buildBriefingPersonalisationBlock,
 } from "@/lib/ai/prompts/insight-generator";
-import { buildRetryCorrectionMessage } from "@/lib/ai/generate-insight";
+import { buildRetryCorrectionMessage } from "@/lib/ai/retry-correction";
 import { singleUserTurn } from "@/lib/ai/types";
 import {
   buildAboutMeInsightBlock,
@@ -1042,8 +1042,8 @@ export async function generateComprehensiveInsight(
 
   // Citation coverage. This annotation is what the admin AI-quality dashboard
   // slices to track how many normative recommendations carry a curated
-  // medical-reference id. It used to hang off `generateInsight()`, which has no
-  // production caller, so the dashboard was reading a metric nothing emitted.
+  // medical-reference id. It used to hang off a wrapper that had no production
+  // caller, so the dashboard was reading a metric nothing emitted.
   // Observational only — an uncited normative rec is a logged warning, never a
   // reason to fail the generation.
   if (insights && typeof insights === "object") {