plan: switching paths — two routes to the same square

Planning only; no source files touched. The BFS visited set in exploreMoves
keys on (coord, entry), so two legal routes that rejoin through the same port
collapse to whichever is shorter — silently discarding the option that couples
cars on the way.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQAJ4dND7enLj54eLyiF2C

docs/plans/switching-paths.md

@@ -0,0 +1,206 @@
+# Plan — Train Switching Paths: two routes to the same square
+
+**Status:** planned, not built. Written against `461c4d9` (v0.4.8).
+**Reported by Jesse:** *"There are times when a train can take two different paths to get to a
+destination. See game at undo 379. Train 11 can go from Eastern limits to the straight at 1,1 two
+different ways. It can go through the refinery and pick up the tanker on its nose, or it could go
+straight straight and then make the curve and skip the refinery and not pick up the car. The player
+should have the two different options."*
+
+---
+
+## 1. The defect
+
+A district with a passing loop — turnout down off the main, curve, industry, curve, turnout back up —
+offers two routes from one end to the rejoin square. Both are legal under §2.4 (a Move runs any
+distance without changing direction; neither route ever leaves a card by the port it entered). They
+differ in what they couple.
+
+**The engine offers only one of them, and which one is an artifact of search order.**
+
+`exploreMoves` (`src/engine/track.ts:386`) keys its visited set on destination *and entry port*:
+
+```ts
+// src/engine/track.ts:492
+const key = `${coordKey(node.coord)}|${node.entry}`;
+if (seen.has(key)) continue;
+```
+
+Both routes arrive at the rejoin square through the **same port**, so they collide on that key. The
+walk is breadth-first, so the shorter route is recorded and the longer one is discarded before it can
+become a destination.
+
+### Reproduction
+
+Fixture — Running Track on row 0, industry spur on row −1:
+
+| | col 0 | col 1 | col 2 | col 3 | col 4 |
+|---|---|---|---|---|---|
+| **row 0** | straight | turnout `{w,e,s}` | straight | turnout `{e,w,s}` | Limits (east) |
+| **row −1** | — | curve `ne` | **Refinery + tanker** | curve `nw` | — |
+
+Walking west from `(0,4)`, today's engine returns:
+
+```
+(0,2)  entry=e  couples=[]        path=(0,3)
+(-1,3) entry=n  couples=[]        path=(0,3)
+(-1,2) entry=e  couples=[tanker]  path=(0,3)->(-1,3)
+(0,0)  entry=e  couples=[]        path=(0,3)->(0,2)->(0,1)      ← only one route to (0,0)
+(-1,1) entry=e  couples=[tanker]  path=(0,3)->(-1,3)->(-1,2)
+```
+
+The refinery route to `(0,0)` — which couples the tanker — is absent. A simple-path prototype on the
+same fixture returns both. Verified identical on v0.4.7 and v0.4.8.
+
+### It is not cosmetic
+
+`dest.path` feeds `carsCoupled.from` (`src/engine/apply.ts`, the `switch.move` execute branch at
+`:1303`) — the list of cards swept clean when cars are lifted. The route reaches the reducer, so
+"which route" must be part of the intent, not just the button text.
+
+### How often
+
+Instrumented over 60 developer-bot games (17,884 destinations enumerated):
+
+| measure | count |
+|---|---|
+| destinations enumerated | 17,884 |
+| squares reachable by 2+ **distinct paths** | 208 (~1.2%) |
+| squares where the routes **differed in what they couple** | **0** |
+
+So the geometry occurs and today's engine silently discards a route about 1% of the time — but the
+bot never builds a district where the discarded route *matters*. Two consequences:
+
+- **Balance impact is expected to be nil.** Still confirm with `compare.ts` (see §6).
+- **The bot will never regression-test this.** It needs explicit unit tests.
+
+---
+
+## 2. The rules question, and Jesse's ruling
+
+§A.4 says: *"If you move your Crew Tray into cars on your track, you **MUST** pick them up. You may
+not go around them."* Whether taking the main instead of the siding counts as "going around" decides
+whether this feature is legal at all.
+
+**Ruling (Jesse):** the player may choose the path. Different paths may pick up different cars, or
+none.
+
+The permissive reading is well-supported by the text: §A.4 says *"cars on your track"*, and the
+industry spur is a different track. Declining to enter it is not going around the tanker — it is not
+going to the tanker's track at all, which is ordinary railroading. The mandatory half still bites in
+full: once a route is chosen, every car standing on it couples, and there is no route that passes a
+car and leaves it.
+
+**Doc changes:** add the clarifying sentence to §A.4 in `docs/rules/rules-v0.2.md`, and record the
+ruling in `docs/rules/open-questions.md`.
+
+---
+
+## 3. Design decisions
+
+**The discriminator is `via`: one intermediate `GridCoord` on the chosen route** (Jesse's call).
+Not an index into the destination list — intents are the canonical record and `undo` replays history
+minus one intent, so an order-dependent field would silently reinterpret saved games.
+
+```ts
+// src/engine/intents.ts:19
+| { type: 'switch.move'; trayId: TrayId; to: GridCoord; reverse: boolean; via?: GridCoord }
+```
+
+- `via` names an **intermediate** card — never the start, never the destination.
+- **Resolution:** among destinations at `to` matching `reverse`, take the first enumerated route whose
+  path contains `via`. Enumeration stays shortest-first.
+- **`via` absent resolves exactly as today** (first route at `to`), so existing histories replay
+  unchanged. Note v0.4.8 already established that saves are *not* preserved across rule changes —
+  this is belt-and-braces, not a hard constraint.
+- **`legal.ts` emits a distinguishing `via`** per route, so a generated intent is never ambiguous.
+
+**Collapse routes whose outcome is identical.** Two routes that lift the same cars from the same
+cards are not a choice; offering both is noise in the action list and doubles the bot's branching
+factor for nothing.
+
+**The dedupe key must include the origin cards, not just car types.** Two routes can couple an
+identical car *multiset* lifted from *different* cards. Keying on type alone would collapse them and
+sweep the wrong card clean. Key on `(coord, entry, [(origin card, car)...])`.
+
+**Simple-path walk with an enumeration cap.** Relaxing `seen` to a per-path visited set keeps
+termination (no card twice within one path), but the number of simple paths is worst-case exponential
+in a dense district. Cap enumeration and prefer shortest routes when the cap bites.
+
+> Note: the 45°-diagonal matching rule appears to make closed loops geometrically impossible — a
+> rectangle of four curves cannot close, because the slopes do not match at the corners. **Do not
+> rely on this.** Use a path-local visited set and a cap regardless.
+
+---
+
+## 4. Change list
+
+| File | Site | Change |
+|---|---|---|
+| `src/engine/track.ts` | `exploreMoves` :386, `seen` :492 | Per-path visited set; dedupe on outcome; cap. Preserve shortest-first ordering. |
+| `src/engine/intents.ts` | :19 | Add `via?: GridCoord`. |
+| `src/engine/apply.ts` | `check` :623 | Select destination by `via`, not first-by-coord. |
+| `src/engine/apply.ts` | `execute` :1303 | Same selection; `carsCoupled.from` follows the chosen path. |
+| `src/engine/apply.ts` | `destinationsFor` :1210, `movesFor` :1260 | Unchanged signatures; `movesFor` still collapses to one square per coord for the board. |
+| `src/engine/legal.ts` | :90 | Emit one intent per distinct route, each with a distinguishing `via`. |
+| `src/sim/view.ts` | `describeIntent` :724 | Find the destination by `via` so the label names the cars *that route* couples. |
+| `src/web/game.ts` | `coordOf` :113, label dedupe ~:229 | Labels must differ per route or the label-dedupe drops one button. |
+| `src/web/main.ts` | `cellRef` :660 | Route-on-hover — see §5. |
+| `src/sim/narrate.ts` | move narration | Name the route when it is not the only one. |
+
+**`ownCutFor` (`apply.ts:1241`) needs no change** — it depends only on the exit port, not the route.
+
+---
+
+## 5. UI — route on hover (Jesse's choice)
+
+v0.4.8 already built the mechanism: `coordOf(intent)` → `cellRef` → `data-square="r,c"` on each
+action button, and hovering or focusing lights that square. This feature extends it from one square
+to a route.
+
+- Carry the path on the **menu**, resolved server-side, for the same reason `coordOf` is: a remote
+  client holds no `GameState`.
+- Emit the route as a second attribute (`data-route="r,c r,c …"`); hovering or focusing lights every
+  card the route runs over, with the destination keeping its existing stronger mark.
+- **On focus as well as hover**, matching v0.4.8 — the keyboard route is not a lesser one.
+
+Target rendering:
+
+```
+Switching Train 11, standing at (1, 4)
+  [ move to (1,1) ]
+  [ move to (1,1) — couples 1 tanker on the way (onto the nose) ]   ← hover
+```
+
+---
+
+## 6. Verification
+
+- **Unit test the fixture in §1** — both routes offered, one coupling the tanker and one not.
+- **`via` round-trip** — each intent `legal.ts` emits resolves back to the route it was generated for.
+- **Back-compat** — replay `public/replays/*.json` and assert identical outcomes with `via` absent.
+- **Origin-card dedupe** — two routes coupling identical car *types* from *different* cards must both
+  be offered, and each must sweep its own card.
+- **Cap** — a dense district must terminate and stay bounded.
+- **Balance** — `node src/sim/compare.ts 1600`. Expect no movement (the bot never builds this
+  geometry); keep only at t ≥ 3 per `README.md`, and read the better/worse/identical split.
+
+**Run the suite serially: `node --test --test-concurrency=1`.** The `npm test` script runs files in
+parallel and several of them invoke the static build against a shared `dist/`, which produces ~20
+spurious failures. Measured on `461c4d9`: parallel 20 fail / serial **575 pass, 0 fail**.
+
+> v0.4.8's CHANGELOG attributes those 20 failures to the tracked `node_modules` symlink. That fix was
+> real and worth making, but it is **not** the whole cause: on this checkout `tsc` resolves and
+> typecheck exits 0, yet the same 20 fail under parallel execution and pass under
+> `--test-concurrency=1`. The `dist/` race is still there. Worth fixing separately — give each
+> build-invoking test its own output directory.
+
+---
+
+## 7. Open question for Jesse
+
+A route's cards may be reachable only as a *subset* of a longer route's cards. If that ever happens,
+no single `via` uniquely selects the shorter route, and it is addressable only by the
+shortest-first tiebreak. It did not arise in any fixture or in 60 bot games, and it may be
+geometrically impossible — but if it does arise, the alternatives are to reject the ambiguous intent
+or to let shortest-first stand. **Shortest-first is assumed here.**