feat: sub-region-level territory — ownership, LLM and map at region granularity (v0.19.0)

Move territory from whole-country (ISO_A3) to individual sub-region id
granularity (PR 3 of #16):

- new server/regions.ts (turf-free): loadRegions (boot), expandTerritory
  (game creation), migrateTerritoryToRegions (legacy games on db.load),
  regionNamesFor / regionLabel (LLM prompt), graceful ISO_A3 fallback
- game creation expands each preset nation's ISO_A3 list to its sub-region
  ids; pre-v0.19 saved games migrate transparently on read (idempotent)
- LLM contract: territoryChanges gain/lose sub-region ids (one region at a
  time; gain moves out of the prior owner; lost-without-gainer = unowned);
  user prompt carries a top-level regions reference table
  ({id: 'name (parent)'} over the state's territory ids)
- WorldMap: choropleth fill per sub-region (countries can split between
  nations), ownership via properties.regionId through the shared
  mapFeatureOwner, flash lights exactly the changed features
  (featuresByRegionIds)
- deterministic stub: one region-level move per non-empty round (player
  annexes the lexicographically-first region of the first other nation)
- tests: new server/regions.test.ts (real geo + fallback + seams);
  map/engine tests re-pinned to region ids; scripts/regions.test.ts stays green

Measured full wwii-1939 user prompt (411 regions): ~24.7k chars / ~6.2k
tokens — far under the 80k frontier limit, so the full regions table is
always sent (no frontier-limited mode needed).

Part of #16

server/stub.ts

@@ -13,7 +13,8 @@ export interface StubPayload {
     turn: number;
     date: string;
     playerNationId: string;
-    nations: Record<string, { stats: Record<string, number> }>;
+    /** As of v0.19 territory is a list of sub-region ids (not ISO_A3 codes). */
+    nations: Record<string, { stats: Record<string, number>; territory?: string[] }>;
   };
   /** The raw pending round the engine emitted (one entry per queued item). */
   pending?: RoundItem[];
@@ -30,7 +31,12 @@ function truncate(s: string, n: number): string {
  * single political "The world moves on" event when the round is empty — so a
  * caller can predict the exact next state for assertions. It always sets the
  * player's military to (current + 5) and allies the player with the first other
- * nation, advancing the date by 2 days.
+ * nation (by key order), advancing the date by 2 days. As of v0.19, when the
+ * round is non-empty it also makes exactly **one** territory move: the player
+ * annexes the lexicographically-first sub-region id of that first other
+ * nation's territory (an empty round makes no move, and a player with nothing
+ * to annex from that nation also makes no move) — so unit tests and e2e can
+ * assert the exact region that changes hands.
  */
 export function deterministicTurn(p: StubPayload): TurnResult {
   const state = p.state;
@@ -41,6 +47,9 @@ export function deterministicTurn(p: StubPayload): TurnResult {
     ? (Object.keys(state.nations).find((id) => id !== pid) ?? '')
     : 'us';
   const items = p.pending ?? [];
+  // The one deterministic per-turn territory move (see the doc above).
+  const otherTerritory = [...(state?.nations?.[otherId]?.territory ?? [])].sort();
+  const gain = items.length > 0 && otherTerritory.length > 0 ? [otherTerritory[0]] : [];
 
   const events: TurnEventSpec[] =
     items.length === 0
@@ -76,7 +85,7 @@ export function deterministicTurn(p: StubPayload): TurnResult {
     events,
     stateChanges: [{ nation: pid, stat: 'military', op: 'set', value: mil }],
     relations: otherId ? [{ a: pid, b: otherId, set: 'ally' }] : [],
-    territoryChanges: [],
+    territoryChanges: gain.length > 0 ? [{ nation: pid, gain, lose: [] }] : [],
   };
 }