Implement completeAuth chokepoint + caller migration

Implements all of unify-auth-completion (12/14 tasks done; manual
smoke + archive-time spec sync remain).

Design refinement during implementation: completeAuth supports two
response shapes via a `mode` parameter:
- mode: 'redirect' (loaders / direct browser navigation; auth.verify.tsx)
- mode: 'json' (action handlers called by imperative fetch from
  client forms; api.auth.login, api.auth.register)

Both modes share createSession + safeReturnTo + Set-Cookie. JSON mode
carries `{ ok: true, step: "done", redirectTo }` (the `step` field
preserves the existing client-form check).

Why two modes: passkey ceremonies are inherently imperative
(start → browser WebAuthn API → finish), so action handlers can't
move to <Form>/useFetcher. Picking option (B) from the design grill —
the chokepoint owns destination selection while clients navigate —
required this dual shape. The 3 hardcoded client-side targets
(returnTo ?? "/", "/", "/?add-passkey=1") collapse into 1 server-side
sanitization pass (safeReturnTo) inside completeAuth.

New module:
- apps/journal/app/lib/auth/session.ts: cookie session storage
  (sessionStorage, createSession, getSessionUser, destroySession)
  moved from auth.server.ts. Legacy import path kept via re-exports
  with @deprecated JSDoc.
- apps/journal/app/lib/auth/completion.ts: completeAuth + safeReturnTo.
- apps/journal/app/lib/auth/completion.test.ts: 10 contract tests
  covering both modes, returnTo sanitization (path-relative, protocol-
  relative, absolute-URL, malformed), Set-Cookie attachment, redirect
  status, JSON shape.

Caller migration:
- api.auth.register.ts passkey-finish → completeAuth(json)
- api.auth.login.ts finish-passkey → completeAuth(json)
- api.auth.login.ts verify-code → completeAuth(json)
- auth.verify.tsx magic-link consumer → completeAuth(redirect)

Client form updates:
- auth.login.tsx: pass returnTo in fetch body, read result.redirectTo
  on done.
- auth.register.tsx: pass returnTo: "/?add-passkey=1" for the magic-
  link verify-code path (preserves the post-register passkey prompt
  via the chokepoint's safeReturnTo, instead of hardcoding it
  client-side).

Verified:
- pnpm typecheck && pnpm lint: green across all 15 workspaces.
- pnpm --filter @trails-cool/journal test: 126 passed.
- pnpm test:e2e auth: 4/4 passed without modification — confirms the
  refactor is behaviour-preserving for the user-facing flows that
  matter most (passkey register + login).

Spec delta in openspec/changes/unify-auth-completion/specs/ applies at
/opsx:archive time.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

openspec/changes/unify-auth-completion/design.md

@@ -40,8 +40,6 @@ ADRs already recorded:
 ```ts
 async function completeAuth(input: {
   userId: string;
-  isNewRegistration: boolean;
-  termsVersion?: string;     // required if isNewRegistration; ignored otherwise
   request: Request;
   returnTo?: string | null;
 }): Promise<Response>
@@ -49,7 +47,9 @@ async function completeAuth(input: {
 
 Returns a `Response` (a `redirect` from React Router with the session `Set-Cookie` attached). Callers do `return await completeAuth(...)` from their action / loader. No headers-vs-redirect split: the chokepoint owns both halves.
 
-`termsVersion` is optional at the type level but required-when-`isNewRegistration` at runtime. Asserting it inside the function (throw on `isNewRegistration && !termsVersion`) keeps the call sites tidy. Alternative: a discriminated union forcing the compiler to require `termsVersion` when `isNewRegistration: true`. Lean toward the runtime assertion — three call sites, two of which set `isNewRegistration: false` — the type ergonomics aren't worth the union complexity.
+**Terms recording is NOT in `completeAuth`.** Both registration paths already write `terms_version` + `terms_accepted_at` at user creation: `finishRegistration` writes them when the user row is inserted (passkey path); `registerWithMagicLink` writes them when the user row is inserted (magic-link path). Magic-link registration *must* write terms at user creation, before verification, because the user row exists between register and verify — if terms weren't already written, the Terms gate would bounce the user on first authenticated visit.
+
+So `completeAuth` is purely **session + redirect**. No `isNewRegistration` flag, no `termsVersion` parameter. The earlier framing of an `isNewRegistration` discriminator was a mis-read — there is no code path that needs to record terms at completion time.
 
 ### Decision 3: File layout
 New directory `apps/journal/app/lib/auth/` containing: