Bump to v0.22.0 (no man's land, #16 P3) + AGENTS.md/CHANGELOG

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to **History** are listed here, newest first.
 
+## v0.22.0 — 2026-10-14 12:02
+- **Part 3 of #16: no man's land — a region can be left unowned mid-game and re-claimed later.** After a war or cataclysm the LLM can leave a region without an owner: it lists the key in the new `TurnResult.unowned` field (and narrates the cause in `events`); the engine removes the key from its sole owner, and the region ends the turn ownerless. A later ordinary `gain` on that key is a **re-claim** — the key is granted to the (sole) gainer and removed from the no-man's-land list. A fresh game always starts fully owned (no man's land can only arise mid-game).
+  - **Engine (`server/engine.ts`).** `applyTurnResult` applies **releases first**: a key is released only when it has exactly one current owner (0 owners — a preset never assigned it — is a no-op, so no man's land can never be fabricated from an ordinary unowned region; 2+ owners — corruption — is a no-op; already-unowned is a no-op). Claims are then resolved against the post-release ownership snapshot: sole gainer + sole owner ≠ gainer = a transfer (the v0.21 rules, unchanged); sole gainer + no owner = a re-claim; a double claim is still rejected. Invariants after every turn: no key sits in two or more territories, a released key sits in none, and `GameState.unowned` holds exactly the keys released mid-game. A lone release counts as `isMeaningful` (it survives the retry/nudge path instead of being replaced by the generic "unclear" result). The system prompt gains the JSON field, the no-man's-land rule (in both tool-mode and plain-mode), and the hard constraint that `unowned` may contain only region keys a nation currently owns. The user payload always carries the current `GameState.unowned` as `unowned` (even when empty), and the jump route's `regions` reference table now includes the released keys so the model can name them when re-claiming.
+  - **New required `GameState.unowned` field** (`shared/types.ts`): the cumulative list of keys released mid-game. A no-man's-land region is **distinct from a preset-unassigned region** — the presets assign only ~25–30% of the 1,796 sub-regions (e.g. `wwii-1939` starts with 411 owned / 1,385 unowned), so ~1,300 ownerless regions exist at init; those keep the existing faint look, and only `GameState.unowned` keys get the distinct one. Legacy SQLite rows saved before v0.22 load with the field defaulted to `[]` (`server/db.ts`), and game creation sets it to `[]`.
+  - **Map presentation (a new contract in `shared/protocol.ts`).** `regionFillStyle(owner, isNoMansLand)` + `unownedTooltipLabel` decide the look: an owned region is unchanged (nation color + `OWNED_STROKE`); a released region gets the distinct visible grey `NO_MANS_LAND_FILL = '#3f3f46'` (hue-less by construction, so it can never collide with a nation's `hsl(h 58% 46%)` color) with a brighter border (`NO_MANS_LAND_STROKE`) and the tooltip sub-line "No man's land"; every other ownerless region keeps the faint `rgba(148,163,184,0.05)` fill and "No owner". The client (`WorldMap`, via a new `unowned` prop from `App.tsx`) is the sole consumer of the distinction — the `RegionMap` simply has no entry for an unowned key (`buildRegionMap` is unchanged). Clicking a no-man's-land region does not change the dossier (there is no owner to focus), and a release/re-claim flashes like any other ownership change.
+  - **Stub + tests.** The deterministic stub is deliberately **unchanged** — it never releases a region (it always returns `unowned: []` and ignores the payload's `unowned` field), because its one-conquest behavior is what the e2e assertions rely on; the release/re-claim path is therefore covered by **unit tests only** (a new `applyTurnResult (no man's land)` suite: release, re-claim, release+claim of the same key, a double claim on a released key, the four no-op release cases, a legacy state without the field, and a `runTurn` end-to-end test with a narrated release; new `buildUserPrompt`/`buildSystemPrompt` contract tests; a presentation-contract suite in `server/map.test.ts`; a legacy-row normalization test in `server/db.test.ts`). The unit suite grows **105 → 127**.
+  - **No data-format wipe.** The `GameState` shape only gains a field that legacy rows load with, so existing saves remain compatible: the deploy is a plain `up -d --build` (no `down -v`).
+- **Verification.** Typecheck, lint, **127 unit tests**, and the production build all pass. The e2e suite (CI, `LLM_STUB=1`) carries no map-color/no-man's-land assertions, so it stays green with the unchanged stub.
+- **Deploy:** manual `docker compose up -d --build` on the host (no `down -v`).
+- Part of #16 — no man's land (P3); #16 stays open.
+
 ## 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.