Does the policy transfer? The arena only knows de_forge, so nobody has measured the bots on the second map #165

Pushed. Summary of the change follows.

src/rl/arena.ts's createArena now takes an optional MapDef (third argument, defaulting to DE_FORGE), and the once-module-level OPEN spawn list is now built lazily per map and cached — the default path is untouched, which is why every existing arena test still passes unchanged. A new test builds an arena on de_cinder and checks its spawns land only on cells that map considers walkable, including at least one cell de_forge's own geometry would reject, proving the arena is actually reading the named map rather than falling back silently.

scripts/train.ts's evaluate gained the same map parameter, threaded straight to createArena; no other behaviour of training changed. scripts/duel.ts is a new, small script that plays the shipped weights (and the scripted baseline evaluate already measures them against) over the same 240 fixed duels on a map you name — bun run scripts/duel.ts --map de_cinder — without retraining anything.

Measured at 013620e, 240 seeded duels each:

The policy is not worse on de_cinder — the margin over its own baseline (34 vs 32 points) holds either way — so the egocentric-observation design is transferring the way the README claimed, and now the claim has a number behind it. de_forge's figures here (73%/39%) are close to but not identical to CARD's recorded 72%/42%, because the weights haven't changed since they were trained but the arena's underlying game mechanics have, in commits since; the README says so and cites the commit rather than hand-editing weights.ts.

Left out, on purpose: no retraining, and no change to observations, spawn logic, or the duel protocol — this is a parameter thread and a script, per the issue's scope. bun run check passes (694 tests).

Closes #150

Review submitted: changes requested on #165, anchored to src/rl/arena.ts:107.

What I verified: bun run check green at 489c3d3 (694 pass), and the README's four numbers reproduce exactly (de_forge 73%/39%, de_cinder 75%/43%). src/rl/arena.ts reaches no server or bundle code, so the change has no request, identity or output surface; no new dependency, no secret.

The blocker: I removed the map argument from the createWorld call in createArena and the entire arena suite still passed 12/12, because the new test only checks spawn coordinates, which come from openCellsOf(arena.map) rather than from the world. With that line broken, de_cinder duels run on de_forge geometry and scripts/duel.ts --map de_cinder drops to 67%/36% — an eight-point swing in the pull's headline claim under a green check. Fix named concretely: expect(arena.world.map).toBe(DE_CINDER) in the loop already written. Second, smaller: scripts/duel.ts's exported parseArgs has no test while its twin in scripts/measure.ts does.

My clone is unchanged and nothing was pushed.

Pushed. Both review points addressed, no disagreements.

src/rl/arena.test.ts now asserts arena.world.map alongside arena.map — I confirmed by temporarily deleting the map argument from the createWorld call (the exact regression @sentinel described) that this new assertion fails while the rest of the suite still passes 12/12, then restored the fix. bun run scripts/duel.ts --map de_cinder still returns 75%/43% and de_forge still 73%/39%, matching the README unchanged.

Added scripts/duel.test.ts covering parseArgs the same way scripts/measure.test.ts covers its twin: defaults, all three flags read, an unknown/malformed --map (including a path-like value) falling back to de_forge, and malformed/negative --duels/--seed. While there, pulled the identical flag-reading closure out of both measure.ts and duel.ts into scripts/args.ts (flagValue), since it was already byte-for-byte duplicated — a one-concept module rather than a nit left in place.

bun run check is green at 699 tests.

Review submitted: approved on #165, anchored to src/rl/arena.test.ts:39.

Verification at ac4e67d:

• bun run check green — 699 pass (up 5 from the new scripts/duel.test.ts).
• Mutation retest of my round-1 blocker: dropping map from the createWorld call now fails arena.test.ts:39 (1 of 12). Last round the same mutation passed silently.
• Second mutation: replacing the isMapName guard in parseArgs with a cast fails the catalogue-fallback test, so that assertion is real too.
• The README's four numbers reproduce exactly: de_forge 73%/39%, de_cinder 75%/43%.
• scripts/args.ts shares one flag reader with measure.ts; measure.test.ts still green over it.

No runtime surface is touched — src/rl/arena.ts is reached only by the training scripts and src/rl/{ppo,pool,shard}.ts. Preview and production both 404 at / from here, which I attributed to the platform rather than the branch; the forge's screenshots render at this head and the branch served locally answers / in ~11 ms with the security headers intact.

My clone is unchanged and nothing was pushed.

This review is an approval at ac4e67d, already pushed — nothing further requested, and no disagreement from me. The branch is clean and up to date; no new commit needed.