Bump to v0.21.0 (region transfers + getNeighbours tool, #16 P2)

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 All notable changes to **History** are listed here, newest first.
 
+## v0.21.0 — 2026-10-14 10:37
+- **Part 2 of #16: territory now moves region-by-region along real land borders, and the LLM can verify borders with a live tool.** v0.19 made ownership region-keyed, but the model's `territoryChanges` were still applied wholesale — any key could jump to any nation, effectively moving whole countries. Now:
+  - **Region-adjacency manifest (Stage 1).** `scripts/build-region-adjacency.mjs` (`npm run build:region-adjacency`) computes which of the 1,796 sub-regions share a land border: per candidate pair, a cheap bbox-proximity prefilter (0.05° gap reject) then a `@turf/boolean-intersects` arbiter. The output `public/region-adjacency.json` is symmetric and deduped — **1,796 entries, 4,470 border pairs (8,940 directed), max degree 19; 1,727 regions have ≥1 neighbour, 69 don't** — and is deterministic/idempotent (sha256 `e9f1d475…`; re-running is byte-identical). Vite copies it into `dist/` for prod; `server/regions.ts` loads it (turf-free — plain `JSON.parse`, no `@turf` import in any `.ts`) and degrades gracefully when it is missing.
+  - **`getNeighbours` LLM tool (Stage 2).** `server/llm.ts` gains `LlmClient.chatWithTools` — a bounded OpenAI function-calling loop: on a `finish_reason: "tool_calls"` response the client runs the injected executor for every call, appends the assistant tool-call message plus one `tool` result message per call, and re-sends until a final `content` (max **6** tool rounds, then one last *no-tools* request forces the answer). It shares the retry/backoff/`reasoning_effort: "low"` machinery with `chat()`; a tool-call round is a success (never treated as truncation); malformed tool arguments become an `{"error":"invalid tool arguments…"}` tool message, so a bad model reply can never throw. `tools` is included in the request body only when non-empty, so plain `chat()` bodies (and their tests) are unchanged.
+  - **Engine wiring.** `runTurn` takes an optional `getNeighbours` resolver — the server's jump route passes one only when the manifest loaded (`hasAdjacency()`). With both the resolver and a `chatWithTools`-capable client, the turn runs the tool path: the system prompt gains the contiguity rules (a nation may only GAIN a key that borders its own territory; every key has exactly one owner) and advertises the tool; the user payload gains a `frontier` field — the player's `playerFrontier`, the sorted/deduped set of *other-nation* keys bordering the player's territory (the natural targets of a player-led conquest). Without the manifest, or with a minimal chat-only client, the turn falls back to the legacy plain `chat()` path — everything keeps working, the tool/frontier simply switch off.
+  - **Sole-ownership enforcement.** `applyTurnResult` now applies territory snapshot-first: before mutating anything it computes each key's current owner, and a key transfers only when (a) exactly one nation claims to gain it (a double claim is rejected — the key stays), (b) it has exactly one current owner (pre-existing double ownership is never touched), and (c) the owner is not the gainer (self-gain is a no-op). A transfer removes the key from all nations first, then grants it to the gainer; lone `lose` entries are no-ops. Every key therefore ends the turn with exactly one owner — it can never be orphaned or duplicated, and a bad model answer degrades to "no transfer", never to a corrupt map.
+  - **Deterministic stub.** The stub now makes one contiguous conquest per turn: the player gains `frontier[0]` (sorted → stable) and its owner loses it. It answers plain content (no tool calls), so it naturally bypasses the tool loop in a single round; with an empty frontier (no controllable border — e.g. the UK in `wwii-1939`, whose only external neighbour is Ireland) nothing transfers and every existing e2e assertion stays stable.
+  - **No data-format change — no save wipe.** The `TerritoryChange` wire shape is unchanged and saved games load as-is, so the deploy is a plain rebuild (no `down -v`).
+- **Verification.** Typecheck, lint, **105 unit tests** (79 baseline + 26 new: tool-loop protocol — echo, parsed executor args, bounded rounds, malformed args, retry; engine tool-path wiring + fallback; `playerFrontier`; executor degradation; sole-ownership invariants; `neighboursOf`/`hasAdjacency`), the production build, and all **14 e2e** tests pass.
+- **Deploy:** manual `docker compose up -d --build` on the host (no `down -v`).
+- Part of #16 — P3 (region rename/split/merge) and P4 follow; #16 stays open.
+
 ## v0.20.0 — 2026-10-14 09:30
 - **Fixed: v0.19.0's region-level ownership did not activate in the production image.** The server-side region loader (`server/regions.ts`) only looked for the manifest + geo in `public/`, but in the Docker image Vite copies `public/*` → `dist/` and no `public/` directory is kept — so `loadRegions` found nothing, `expandCountriesToKeys` degraded to a no-op, and freshly-created games kept raw ISO_A3 codes (`DEU`, `GBR,IRL,IND`, …) instead of expanding to region keys. The choropleth therefore still rendered whole-country ownership and the model's region-key contract was inactive in prod (the client's own `/region-keys.json` fetch still worked, because the client is served from `dist/`). `loadRegions` now probes `public/` (dev) then `dist/` (prod, == `clientDir`) and uses the first directory holding both files, so one build works in both environments. A new test covers the prod-only (`dist/`), dev (`public/`), and empty-root (no data) layouts.
 - **Deploy:** `docker compose down -v` (drops the `history-data` volume — pre-v0.19 saves were already dropped by v0.19.0 and are dropped again here) then `up -d --build`.