feat(coach): scaffold conversation-summary + facts storage

Extract the AES-256-GCM Bytes codec from persistence into a shared
bytes-codec module so the conversation summary and the new personal-fact
store reuse one ArrayBuffer-backed implementation. Add the CoachFact
table (migration 0116) for durable, user-scoped facts the Coach extracts
from conversations — encrypted fact text, plain category/confidence/
source for rank-without-decrypt, soft-delete for a reversible forget.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

prisma/migrations/0116_v1111_coach_facts/migration.sql

@@ -0,0 +1,27 @@
+-- v1.11.1 — durable Coach personal facts (Epic B, B-W7).
+--
+-- Additive, multi-tenant-safe: one new table + index + FK, no column on an
+-- existing hot table, no backfill. The fact text is encrypted at rest
+-- (AES-256-GCM, same codec as coach_messages.encrypted_content); category /
+-- confidence / source stay plain so the injection picker can rank without a
+-- per-row decrypt. Soft-delete via deleted_at keeps "forget this" auditable.
+CREATE TABLE "coach_facts" (
+    "id" TEXT NOT NULL,
+    "user_id" TEXT NOT NULL,
+    "fact_encrypted" BYTEA NOT NULL,
+    "category" TEXT NOT NULL,
+    "confidence" INTEGER NOT NULL DEFAULT 50,
+    "source_conversation_id" TEXT,
+    "created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    "updated_at" TIMESTAMP(3) NOT NULL,
+    "deleted_at" TIMESTAMP(3),
+
+    CONSTRAINT "coach_facts_pkey" PRIMARY KEY ("id")
+);
+
+CREATE INDEX "coach_facts_user_id_deleted_at_confidence_idx"
+    ON "coach_facts" ("user_id", "deleted_at", "confidence");
+
+ALTER TABLE "coach_facts"
+    ADD CONSTRAINT "coach_facts_user_id_fkey"
+    FOREIGN KEY ("user_id") REFERENCES "users" ("id") ON DELETE CASCADE ON UPDATE CASCADE;

prisma/schema.prisma

@@ -351,6 +351,7 @@ model User {
   devices                Device[]
   integrationStatuses    IntegrationStatus[]
   coachConversations     CoachConversation[]
+  coachFacts             CoachFact[]
   coachUsage             CoachUsage[]
   // v1.4.25 W8d — workout passthrough (Apple HKWorkout / Withings
   // getworkouts / future Strava). Schema lands now so the v1.5 iOS
@@ -2687,6 +2688,37 @@ model CoachMessage {
   @@map("coach_messages")
 }
 
+// v1.11.1 (Epic B, B-W7) — durable, user-scoped personal facts the Coach
+// extracts from conversations: stable preferences, conditions the user states
+// about themselves, goals, constraints, durable life context. The fact TEXT is
+// encrypted at rest (AES-256-GCM, the same codec as CoachMessage). `category`,
+// `confidence`, and `sourceConversationId` stay plain so the injection picker
+// can rank without paying a per-row decrypt. Descriptive, never diagnostic.
+// Soft-delete via `deletedAt` so a user "forget this" action is reversible /
+// auditable and re-extraction can avoid resurrecting it.
+model CoachFact {
+  id                   String    @id @default(cuid())
+  userId               String    @map("user_id")
+  factEncrypted        Bytes     @map("fact_encrypted")
+  // App-side closed enum (not a DB enum, to keep the migration purely additive
+  // and avoid an enum-migration per new category): preference | condition |
+  // goal | constraint | context.
+  category             String
+  // 0..100 server-assigned extraction confidence.
+  confidence           Int       @default(50)
+  // Provenance — nullable so a future manual-entry surface can leave it null.
+  sourceConversationId String?   @map("source_conversation_id")
+  createdAt            DateTime  @default(now()) @map("created_at")
+  updatedAt            DateTime  @updatedAt @map("updated_at")
+  // Soft-delete: hides the fact from injection but keeps the row for audit.
+  deletedAt            DateTime? @map("deleted_at")
+
+  user User @relation(fields: [userId], references: [id], onDelete: Cascade)
+
+  @@index([userId, deletedAt, confidence])
+  @@map("coach_facts")
+}
+
 // Per-user, per-day token-spend ledger for the AI Coach. The dispatcher
 // reads `totalTokens` against `MAX_TOKENS_PER_USER_PER_DAY` before
 // hitting any provider, so a runaway client cannot drain the operator's

src/lib/ai/coach/bytes-codec.ts

@@ -0,0 +1,32 @@
+/**
+ * Shared AES-256-GCM ↔ `Bytes` codec for the Coach's encrypted columns.
+ *
+ * `coach_messages.encrypted_content`, `coach_conversations.summary_encrypted`,
+ * and `coach_facts.fact_encrypted` all store the same payload shape: the
+ * `encrypt()` string format from `@/lib/crypto` (`"<keyId>.<base64>"`) encoded
+ * as UTF-8 bytes. Centralised here so there is one ArrayBuffer-backed
+ * implementation every Coach persistence path shares.
+ *
+ * Prisma's `Bytes` type maps to `Uint8Array<ArrayBuffer>`, not Node's
+ * `Buffer<ArrayBufferLike>`, so we allocate a fresh ArrayBuffer-backed
+ * `Uint8Array` for writes to keep the structural type stable across Node
+ * versions.
+ */
+import { Buffer } from "node:buffer";
+
+import { decrypt, encrypt } from "@/lib/crypto";
+
+/** Encrypt a UTF-8 string into the `Bytes` payload the schema stores. */
+export function encryptToBytes(plaintext: string): Uint8Array<ArrayBuffer> {
+  const ciphertext = encrypt(plaintext);
+  const encoded = Buffer.from(ciphertext, "utf8");
+  const out = new Uint8Array(new ArrayBuffer(encoded.byteLength));
+  out.set(encoded);
+  return out;
+}
+
+/** Decrypt a `Bytes` payload back to its plaintext. Throws on a bad key id. */
+export function decryptFromBytes(buf: Uint8Array): string {
+  const text = Buffer.from(buf).toString("utf8");
+  return decrypt(text);
+}

src/lib/ai/coach/persistence.ts

@@ -10,10 +10,8 @@
  * provenance (window names, metric tags, sample counts) and never raw
  * values, so it can be queried without decryption for analytics.
  */
-import { Buffer } from "node:buffer";
-
 import { prisma } from "@/lib/db";
-import { encrypt, decrypt } from "@/lib/crypto";
+import { decryptFromBytes, encryptToBytes } from "./bytes-codec";
 
 import type {
   CoachConversationDTO,
@@ -49,27 +47,6 @@ export function summariseTitle(input: string): string {
   return `${cut.trimEnd()}…`;
 }
 
-/**
- * Encode a UTF-8 string as the AES-256-GCM payload format the schema
- * stores in `coach_messages.encrypted_content`.
- *
- * Prisma's `Bytes` type maps to `Uint8Array<ArrayBuffer>`, not Node's
- * `Buffer<ArrayBufferLike>`. We allocate a fresh ArrayBuffer-backed
- * Uint8Array so the structural type matches across Node versions.
- */
-function encryptToBytes(plaintext: string): Uint8Array<ArrayBuffer> {
-  const ciphertext = encrypt(plaintext);
-  const encoded = Buffer.from(ciphertext, "utf8");
-  const out = new Uint8Array(new ArrayBuffer(encoded.byteLength));
-  out.set(encoded);
-  return out;
-}
-
-function decryptFromBytes(buf: Uint8Array): string {
-  const text = Buffer.from(buf).toString("utf8");
-  return decrypt(text);
-}
-
 function provenanceToJson(provenance: CoachProvenance | null): string | null {
   if (!provenance) return null;
   return JSON.stringify(provenance);