diff --git a/.claude/skills/data-fetching-perf-audit/SKILL.md b/.claude/skills/data-fetching-perf-audit/SKILL.md
new file mode 100644
index 000000000..65ee78a32
--- /dev/null
+++ b/.claude/skills/data-fetching-perf-audit/SKILL.md
@@ -0,0 +1,101 @@
+---
+name: data-fetching-perf-audit
+description: >-
+  Audit and optimize data fetching across TanStack Start route loaders, server
+  functions (createServerFn), and Drizzle/PlanetScale queries in wodsmith-start.
+  Use when a page or route group feels slow, when asked to do a "performance
+  analysis" / "perf audit" of data fetching, or to investigate loader
+  waterfalls, N+1 query patterns, sequential awaits of independent queries,
+  fake parallelism, over-fetching, or missing indexes. Triggers: slow compete
+  pages, slow organizer dashboard, "why is this loader slow", "optimize the
+  queries on X", "eliminate N+1", "reduce round trips", "batch these fetches".
+  Models the methodology proven in PRs #502 and #512.
+---
+
+# Data Fetching Performance Audit
+
+Find and fix data-fetching slowness in `wodsmith-start` — the layer between a
+route loader and the database. The wins are almost never "make the query
+faster"; they are **fewer round trips** and **real parallelism**. This skill
+encodes the audit methodology and the fix catalog proven in PR #502 (batch N+1
+across organizer routes) and PR #512 (eliminate waterfalls on public compete
+pages).
+
+## Why this stack is different
+
+Three facts drive every decision here. Internalize them before touching code:
+
+1. **A server fn called from a loader is an HTTP round trip.** On client-side
+   navigation, each `createServerFn` call the loader awaits is a separate
+   request. Three sequential `await fooFn()` calls = three serial network waves
+   before the page renders. Collapsing waves matters more than query speed.
+2. **`getDb()` wraps a single mysql2 connection, and mysql2 serializes commands
+   per connection.** `Promise.all([db.select()…, db.select()…])` on **one** `db`
+   runs sequentially on the wire — fake parallelism. Real parallelism requires
+   each branch to use its **own** `getDb()`.
+3. **The push-based DB workflow has no migrations dir.** Index changes are
+   schema edits applied with `pnpm db:push`, not generated migrations.
+
+## The audit workflow
+
+Work top-down: a page is slow because of how its loader composes server fns,
+which is slow because of how each server fn composes queries.
+
+1. **Scope.** Pick the route or route group (e.g. everything under
+   `/compete/organizer/$competitionId/`, or the public `/compete/$slug` tree).
+   Enumerate every route file in scope — don't sample.
+2. **Trace each loader → server fn → query chain.** For each route's `loader`,
+   list the server fns it awaits and whether they run sequentially or in
+   `Promise.all`. Open each server fn and list its internal DB queries and
+   whether *those* are sequential. Note anything called once per item in a list.
+3. **Classify** every finding against the [anti-pattern catalog](#anti-pattern-catalog).
+   Read `references/anti-patterns.md` for the detection signal, the fix, and
+   real before/after code for each.
+4. **Fix**, smallest blast radius first. Preserve loader data shapes so
+   components and their props are untouched (verify with `pnpm type-check`).
+   Keep single-item server fns alongside new batched ones — targeted refreshes
+   still need them.
+5. **Watch the two traps** in `references/stack-gotchas.md` — they will silently
+   undo your work (fake parallelism) or break the client build
+   (`cloudflare:workers` leaking into the bundle).
+6. **Verify** with the checklist below.
+
+For a thorough audit of a large route group, fan out: one agent per
+page/loader tracing its full chain, then an adversarial pass that re-checks
+each proposed fix for regressions (deferred-data drops, serialization,
+behavior changes). PR #512's analysis was a 10-agent trace reviewed by a
+20-agent adversarial pass. Scale the agent count to the scope.
+
+## Anti-pattern catalog
+
+Quick reference — full detection + fix + code in `references/anti-patterns.md`.
+
+| Anti-pattern | Smell | Fix |
+|---|---|---|
+| **Loader waterfall** | Loader `await`s fn A, then B with A's result, then C — each a round trip on client nav | Consolidate into one server fn (or ≤2 parallel) that composes the work in-process |
+| **N+1 server-fn call** | `items.map(i => someFn({ data: i.id }))` / a per-item call inside a loop | One batched fn taking `ids[]`, one `inArray` query, returns a `Record<id, …>` map |
+| **Sequential independent awaits** | `const a = await x(); const b = await y()` where `y` doesn't use `a` | `const [a, b] = await Promise.all([x(), y()])` |
+| **Fake parallelism** | `Promise.all` over one shared `getDb()` instance | One `getDb()` per parallel branch (mysql2 serializes per connection) |
+| **Over-fetching** | `select()` whole rows / full `users` (ships `passwordHash`, profile JSON); filtering in JS | Project only consumed columns; push filters into SQL (`where`) |
+| **Scan-to-check** | Fetch all rows then `.find`/`.some` in JS to test existence | Single `EXISTS`-scoped / `inArray` query |
+| **Missing hot-path index** | `inArray` / `eq` on a junction table (`workout_movements`, `workout_tags`) doing full scans | Add `index(...)` in the schema, apply with `pnpm db:push` |
+| **Dead query** | A fetched value that nothing in the loader/component consumes | Delete it |
+| **Per-item on critical path** | Non-essential data blocking first render | Defer it off the loader (chain un-awaited, hydrate client-side) |
+
+## Security while optimizing
+
+Tracing every fetch surfaces authz holes — fix them in the same pass (PR #512
+found two). Watch for: client-supplied `userId`/ids trusted without a session
+check (IDOR — derive identity from `getSessionFromCookie()`), and unpublished/
+draft data (e.g. heats with athlete names) returned to anonymous viewers. See
+`references/stack-gotchas.md`.
+
+## Verification checklist
+
+- [ ] `pnpm type-check` — loader data shapes unchanged (component contracts intact)
+- [ ] `pnpm build` — confirms no `@/db` / `cloudflare:workers` leaked into the client bundle
+- [ ] `pnpm test` — at parity with `main` (run suspected pre-existing failures on a clean `main` to confirm)
+- [ ] `pnpm check` (Biome) clean on changed files
+- [ ] If indexes changed: `pnpm db:push` applied (no migration files)
+- [ ] `lat check` passes; update `lat.md/architecture.md` to document new consolidated/batched fns
+- [ ] Each parallel branch verified to use its own `getDb()` (no fake parallelism)
diff --git a/.claude/skills/data-fetching-perf-audit/references/anti-patterns.md b/.claude/skills/data-fetching-perf-audit/references/anti-patterns.md
new file mode 100644
index 000000000..3bcdb9266
--- /dev/null
+++ b/.claude/skills/data-fetching-perf-audit/references/anti-patterns.md
@@ -0,0 +1,273 @@
+# Anti-pattern catalog: detection + fix + code
+
+Each section: how to spot it, how to fix it, and real before/after from
+`wodsmith-start` (PRs #502, #512). Code is illustrative — match the existing
+file's imports and style when applying.
+
+## Contents
+
+1. [Loader waterfall → consolidated server fn](#1-loader-waterfall)
+2. [N+1 server-fn call → batched map fn](#2-n1-server-fn-call)
+3. [Sequential independent awaits → Promise.all](#3-sequential-independent-awaits)
+4. [Fake parallelism → one getDb() per branch](#4-fake-parallelism)
+5. [Over-fetching → project + filter in SQL](#5-over-fetching)
+6. [Scan-to-check → EXISTS-scoped query](#6-scan-to-check)
+7. [Missing hot-path index](#7-missing-hot-path-index)
+8. [Dead query → delete](#8-dead-query)
+9. [Per-item on critical path → defer](#9-per-item-on-critical-path)
+
+---
+
+## 1. Loader waterfall
+
+**Detect.** A route `loader` awaits one server fn, then awaits a second using
+the first's result, then a third. Each await is an HTTP round trip on
+client-side navigation. A page that issues 3–6 sequential waves pays 3–6 serial
+network latencies before render.
+
+**Fix.** Collapse the waves into **one consolidated server fn** that composes
+the existing fns/queries in-process (server-to-server, no extra round trips), or
+into **≤2 parallel** fns when one is public and one is session-specific. Both
+take the same input (e.g. `slug`) so neither waits on the other — one extra
+cheap indexed lookup beats a serial HTTP wave. Keep the loader's returned data
+shape identical so components don't change.
+
+Reference fns: `getPublicCompetitionPageDataFn` + `getViewerCompetitionContextFn`
+(`src/server-fns/competition-page-fns.ts`), `getPublicWorkoutsPageDataFn`
+(`competition-workouts-page-fns.ts`), `getPublicEventPageDataFn`
+(`competition-event-page-fns.ts`).
+
+```ts
+// BEFORE — loader: 3 sequential round trips
+const comp = await getCompetitionFn({ data: { slug } })
+const divisions = await getDivisionsFn({ data: { competitionId: comp.id } })
+const sponsors = await getSponsorsFn({ data: { competitionId: comp.id } })
+
+// AFTER — one consolidated fn; public branches fan out in-process, each on its
+// own connection. Anonymous viewers get empty stubs with zero DB work.
+export const getPublicCompetitionPageDataFn = createServerFn({ method: "GET" })
+  .inputValidator((d: unknown) => competitionPageInputSchema.parse(d))
+  .handler(async ({ data }) => {
+    const { competition } = await getCompetitionBySlugFn({ data: { slug: data.slug } })
+    if (!competition) return { competition: null, divisions: [], sponsors: emptySponsors }
+    const [divisionsResult, sponsors, judgesScheduleResult] = await Promise.all([
+      getPublicCompetitionDivisionsForCompetition({ competition }),
+      getCompetitionSponsorsFn({ data: { competitionId: competition.id } }),
+      hasJudgesScheduleFn({ data: { competitionId: competition.id } }),
+    ])
+    return { competition, divisions: divisionsResult.divisions, sponsors, /* … */ }
+  })
+```
+
+Split public vs. session data when one half needs no session: the public fn
+does zero session work and is cacheable; the viewer fn reads the session **once**
+and short-circuits to empty stubs for anonymous visitors.
+
+---
+
+## 2. N+1 server-fn call
+
+**Detect.** A server fn (or multi-query fn) invoked once per item in a list:
+`volunteers.map(v => canInputScoresFn({ data: v.userId }))`, or a parent fetching
+child data once per child. Cost scales with list length — the classic N+1.
+
+**Fix.** Add a **batched fn** that takes an array of ids, issues **one
+`inArray` query**, and returns a `Record<id, …>` map (pre-filled with defaults so
+every requested id is present). Callers do a map lookup instead of a call.
+Exploit shared scope: if children share a parent's competition-scoped data, fetch
+it once for all of them rather than per child.
+
+Reference fns: `getScoreAccessMapFn` (`src/server-fns/volunteer-fns.ts`),
+`getJudgeSchedulingDataForEventsFn` (`judge-scheduling-fns.ts`),
+`getEventScoreEntryDataWithHeatsBatchFn` (`competition-score-fns.ts`),
+`getHeatsByWorkoutIdsInternal` (`competition-heats-fns.ts`).
+
+```ts
+// BEFORE — N queries, one per volunteer
+const access = {}
+for (const v of volunteers) {
+  access[v.userId] = await canInputScoresFn({ data: { userId: v.userId, competitionTeamId } })
+}
+
+// AFTER — one inArray query, returns userId -> boolean map
+export const getScoreAccessMapFn = createServerFn({ method: "GET" })
+  .inputValidator((d: unknown) =>
+    z.object({ userIds: z.array(z.string()), competitionTeamId: competitionTeamIdSchema }).parse(d),
+  )
+  .handler(async ({ data }): Promise<Record<string, boolean>> => {
+    const accessMap: Record<string, boolean> = {}
+    for (const userId of data.userIds) accessMap[userId] = false  // defaults
+    if (data.userIds.length === 0) return accessMap
+    const db = getDb()
+    const entitlements = await db
+      .select({ userId: entitlementTable.userId })
+      .from(entitlementTable)
+      .where(and(
+        inArray(entitlementTable.userId, data.userIds),
+        eq(entitlementTable.teamId, data.competitionTeamId),
+        eq(entitlementTable.entitlementTypeId, SCORE_INPUT_TYPE_ID),
+        isNull(entitlementTable.deletedAt),
+      ))
+    for (const e of entitlements) if (e.userId) accessMap[e.userId] = true
+    return accessMap
+  })
+```
+
+When batching collapses N calls to 1, the per-item result may now be missing for
+a bad id. Preserve prior behavior explicitly — e.g. `throw` "Event not found" if a
+child's data is absent from the batch result (matching the old per-child error)
+rather than silently dropping it. Also avoid re-introducing O(items × rows): group
+rows by their key in a single pass, don't `.filter` all rows per item.
+
+---
+
+## 3. Sequential independent awaits
+
+**Detect.** Inside a fn or loader, `const a = await x(); const b = await y()`
+where `y` does **not** use `a`. They are independent but run serially.
+
+**Fix.** `const [a, b] = await Promise.all([x(), y()])`. But across **server fns**
+this is real concurrency; across **raw queries on one `db`** it is not — see #4.
+
+Reference: pricing loader (5 independent fetches → one `Promise.all`), sponsors
+loader, `getCompetitionRevenueStats` (`src/server/commerce/fee-calculator.ts`),
+all in PR #502.
+
+---
+
+## 4. Fake parallelism
+
+**Detect.** `Promise.all([...])` whose branches all use the **same** `db` /
+`getDb()` instance. mysql2 serializes commands on a single connection, so this
+runs sequentially despite the `Promise.all` — the most insidious trap because it
+*looks* parallel.
+
+**Fix.** Give each parallel branch its **own** `getDb()`. In-process server-fn
+composition gets this for free (each fn calls `getDb()` internally), but raw
+queries you parallelize by hand must each open a connection.
+
+Reference: `getCompetitionLeaderboard` (`src/server/competition-leaderboard.ts`)
+runs three dependency-ordered parallel groups, each branch on its own connection.
+
+```ts
+// BEFORE — fake parallelism: one connection serializes both
+const db = getDb()
+const [competition, registrations] = await Promise.all([
+  db.select().from(competitionsTable).where(...),
+  db.select().from(registrationsTable).where(...),
+])
+
+// AFTER — one connection per branch actually runs them concurrently
+const competitionDb = getDb()
+const registrationsDb = getDb()
+const [competition, registrations] = await Promise.all([
+  competitionDb.select().from(competitionsTable).where(...),
+  registrationsDb.select().from(registrationsTable).where(...),
+])
+```
+
+Dependency-order the groups: group 1 fetches what group 2 needs; within a group,
+everything is independent and each branch has its own connection.
+
+---
+
+## 5. Over-fetching
+
+**Detect.** `db.select()` (whole row) when only a few columns are used —
+especially joining `users`, which ships `passwordHash` and `athleteProfile` JSON
+for every athlete onto the wire. Or fetching all rows and filtering by a field in
+JS.
+
+**Fix.** Project only the consumed columns with `select({ … })`, and push the
+filter into SQL with `where(...)` instead of a JS `.filter`.
+
+Reference: leaderboard registrations now project only the
+registration/user/division columns the ranking consumes and filter division in
+SQL (`src/server/competition-leaderboard.ts`, PR #512).
+
+```ts
+// BEFORE — full user rows + JS division filter
+const regs = (await db.select().from(registrationsTable)
+  .innerJoin(users, eq(users.id, registrationsTable.userId)))
+  .filter(r => r.division_id === divisionId)
+
+// AFTER — projected columns + SQL filter
+const regs = await db
+  .select({ id: registrationsTable.id, name: users.firstName, divisionId: registrationsTable.divisionId })
+  .from(registrationsTable)
+  .innerJoin(users, eq(users.id, registrationsTable.userId))
+  .where(eq(registrationsTable.divisionId, divisionId))
+```
+
+---
+
+## 6. Scan-to-check
+
+**Detect.** Fetching every row in a scope, then `.find`/`.some` in JS to test
+whether one exists — e.g. loading every registration in a competition just to see
+if a given email has a pending team invite.
+
+**Fix.** One `EXISTS`-correlated or `inArray`-scoped query that answers the
+question in SQL.
+
+Reference: pending team invites now use a single EXISTS-scoped query on the
+invitee email instead of scanning every registration
+(`getPendingTeamInvitesForEmail`, `src/server/competition-detail.ts`, PR #512).
+
+---
+
+## 7. Missing hot-path index
+
+**Detect.** Frequent `inArray`/`eq` lookups on a junction/global table that has
+no index on the looked-up column — a full table scan on every hit. PR #512 found
+`workout_movements` and `workout_tags` (global junction tables) were full-scanned
+on hot `inArray` lookups, plus `judge_heat_assignments(rotationId)`.
+
+**Fix.** Add the index in the Drizzle schema's table-extras callback, then apply
+with `pnpm db:push` — **no migration files** (push-based workflow; no migrations
+dir exists). Flag the `db:push` step explicitly in the PR.
+
+```ts
+// src/db/schemas/workouts.ts
+export const workoutTagsTable = mysqlTable(
+  "workout_tags",
+  { ...commonColumns, id: varchar({ length: 255 }).primaryKey(),
+    workoutId: varchar({ length: 255 }).notNull(), tagId: varchar({ length: 255 }).notNull() },
+  (table) => [
+    index("workout_tags_workout_idx").on(table.workoutId),
+    index("workout_tags_tag_idx").on(table.tagId),
+  ],
+)
+```
+
+---
+
+## 8. Dead query
+
+**Detect.** A fetched value that nothing downstream reads — a `totalEvents` count
+returned but never rendered, a `getTeamContactEmailFn` loader fetch whose result
+is unused, a duplicate mapping query. Grep the symbol; if no consumer, it's dead.
+
+**Fix.** Delete it. PR #512 removed `totalEvents`, a dead `getTeamContactEmailFn`
+fetch, a duplicate leaderboard mappings query, and redundant inline registration
+fns. Each deletion is one fewer query on the critical path.
+
+---
+
+## 9. Per-item on critical path
+
+**Detect.** Non-essential, expensive data (heats, schedules) awaited in the
+loader, blocking first paint, when the component could hydrate it after render.
+
+**Fix.** Keep it **off** the loader's critical path: chain the fn off the
+page-data promise **un-awaited** and let the component resolve it client-side
+(deferred data). Distinct from #1 — here the goal isn't fewer round trips but
+keeping a slow fetch out of the blocking path.
+
+Reference: event-detail published heats — the route chains `getPublicEventHeatsFn`
+off the page-data promise un-awaited and the heat-schedule component resolves it
+client-side (`src/routes/compete/$slug/workouts/$eventId.tsx`, PR #512).
+
+> Regression watch: when deferring, ensure the deferred promise is actually
+> forwarded to the component and not accidentally awaited or dropped — PR #512's
+> adversarial review caught a heats-deferral regression.
diff --git a/.claude/skills/data-fetching-perf-audit/references/stack-gotchas.md b/.claude/skills/data-fetching-perf-audit/references/stack-gotchas.md
new file mode 100644
index 000000000..5f0d93c27
--- /dev/null
+++ b/.claude/skills/data-fetching-perf-audit/references/stack-gotchas.md
@@ -0,0 +1,125 @@
+# Stack gotchas
+
+Codebase-specific traps that silently undo a perf fix or break the build. The
+first two are the ones that bite hardest; read them before optimizing.
+
+## Contents
+
+1. [mysql2 single-connection serialization](#1-mysql2-single-connection-serialization)
+2. [cloudflare:workers leaking into the client bundle](#2-cloudflareworkers-leaking-into-the-client-bundle)
+3. [Push-based DB workflow (indexes)](#3-push-based-db-workflow)
+4. [Preserve loader data shapes](#4-preserve-loader-data-shapes)
+5. [Server fn = HTTP round trip; in-process composition is free](#5-server-fn--http-round-trip)
+6. [Security holes surface during the audit](#6-security-holes-surface-during-the-audit)
+7. [Caching is deferred, not free](#7-caching-is-deferred-not-free)
+
+---
+
+## 1. mysql2 single-connection serialization
+
+`getDb()` returns a Drizzle instance wrapping a **single mysql2 connection**, and
+mysql2 serializes commands per connection. Consequences:
+
+- `Promise.all([db.q1(), db.q2()])` on **one** `db` is **fake parallelism** — the
+  queries run back-to-back on the wire. The `Promise.all` makes it *look*
+  concurrent; it isn't.
+- Real concurrency requires **one `getDb()` per parallel branch**. In-process
+  server-fn composition gets this automatically (each fn opens its own
+  connection). Hand-written parallel raw queries must each call `getDb()`.
+
+```ts
+// Each parallel branch gets its OWN getDb() connection: a Drizzle instance
+// wraps a single mysql2 connection which serializes commands, so sharing one
+// instance across Promise.all branches would still execute sequentially.
+const competitionDb = getDb()
+const registrationsDb = getDb()
+const [competition, registrations] = await Promise.all([
+  competitionDb.select()...,
+  registrationsDb.select()...,
+])
+```
+
+> Do **not** "fix" this by memoizing `getDb()` into a shared pooled instance —
+> naive memoization would serialize all parallel queries again. Per-request
+> connection pooling is a deliberate infra decision, not a perf-audit change.
+
+## 2. cloudflare:workers leaking into the client bundle
+
+`@/db` imports `cloudflare:workers`, which Vite **cannot resolve in the browser**.
+TanStack Start's client compile strips `createServerFn` **handler bodies** (and
+their imports), but it **cannot strip exported plain functions**. So an exported
+plain helper that calls `getDb()`, living in a `src/server-fns/*` module, drags
+`@/db` → `cloudflare:workers` into the client graph and breaks the build:
+
+```
+[vite:import-analysis] Failed to resolve import "cloudflare:workers" from "src/db/index.ts"
+```
+
+This bites specifically when consolidating fns: extracting shared DB helpers as
+exported plain functions is the natural refactor, and it's the one that breaks.
+
+**Fix — follow the `src/server/` convention:**
+
+- DB-backed plain helpers go in `src/server/*.ts` with `import "server-only"` at
+  the top (e.g. `src/server/competition-detail.ts`,
+  `src/server/competition-divisions.ts`).
+- Client-safe pieces (Zod parsers, constants, plain types like
+  `parseCompetitionSettings`, `PublicCompetitionDivision`) go in `src/utils/*`
+  (e.g. `src/utils/competition-settings.ts`) so server modules can use them with
+  no cycle. Re-export from the original `*-fns.ts` so existing importers (incl.
+  client routes) stay unchanged.
+- `src/server-fns/*` files then reference the helpers **only inside handler
+  bodies**, which the client compiler strips.
+
+Always run `pnpm build` (not just `type-check`) after a consolidation — the
+client-bundle resolution error only appears at build time.
+
+## 3. Push-based DB workflow
+
+No migrations directory exists. Schema/index changes are applied with
+`pnpm db:push` during development. When a perf fix adds an index, the index lives
+in the Drizzle schema's table-extras callback and the PR must call out the
+`pnpm db:push` step explicitly. Do **not** hand-write SQL migrations.
+
+## 4. Preserve loader data shapes
+
+The goal is faster loading with **zero** change to what components receive. Keep
+each loader's returned object shape identical (same keys, same types) so child
+components and their props are untouched. `pnpm type-check` is the guardrail — if
+it passes, the contract held. The only acceptable shape change is deleting a
+genuinely dead field (e.g. `organizerContactEmail`), and only after confirming no
+consumer reads it.
+
+## 5. Server fn = HTTP round trip
+
+A `createServerFn` call awaited in a loader is a network request on client-side
+navigation. Composing fns **in-process** (one server fn calling another's logic
+server-side) costs no extra round trip. That's why consolidation works: the
+loader makes 1–2 calls, and the fan-out happens server-to-server. When you need
+two loader-level calls, make them parallel and have each resolve its own
+prerequisites (e.g. both take the `slug`) so neither blocks the other.
+
+## 6. Security holes surface during the audit
+
+Tracing every fetch is also an authz audit. Two real holes found in PR #512:
+
+- **IDOR:** `getUserCompetitionRegistrationsFn` accepted a client-supplied
+  `userId` with no session check — any visitor could read another user's
+  registrations. Fix: derive the user from `getSessionFromCookie()`, never trust
+  client input for identity.
+- **Draft exposure:** `getHeatsForCompetitionFn` returned unpublished heats —
+  including athlete names — to anonymous visitors. Fix: filter
+  `schedulePublishedAt IS NOT NULL` unless the viewer is a site admin or holds
+  owner/admin/`manage_programming` on the organizing team. Add regression tests
+  and verify every existing caller is unaffected.
+
+When over-fetching also leaks sensitive columns (`passwordHash`, profile JSON via
+full `users` rows), the projection fix in anti-pattern #5 is a security fix too.
+
+## 7. Caching is deferred, not free
+
+These optimizations are about call structure, not caching. Caching the
+now-consolidated public payloads (KV), enabling Hyperdrive query caching, and
+per-request connection pooling are separate infra decisions — Hyperdrive query
+caching is explicitly disabled in `alchemy.run.ts`; verify *why* before flipping
+it. Don't bundle these into a structural perf audit.