feat: port the four asobi_lua commits that landed after the merge snapshot

[Taure]

asobi_lua was merged into asobi at 733cec4, snapshotting asobi_lua at b01c5dc. Four asobi_lua commits landed after that point and never made it across. asobi_lua#135 retires that repo and deletes its src/, so merging it without these first would regress the shipped runtime.

This ports all four, one commit each, in dependency order. Unblocks widgrensit/asobi_lua#135 and #346.

What landed

asobi_lua	Ported as	Applied
#131 af391ff atomize spawn-template base_state / zone.spawn overrides	c8504e9	clean
#134 1bc28be validate game.broadcast event names	97ebac6	reworked
#133 3354e5e real-Postgres CT for game.storage.*	ed7a63a	reworked
#132 4b4bf8e game-declared registration mode	1736cd5	reworked

The two reworks worth reviewing

#134 - one owner instead of two copies. In separate repos, asobi_lua had to restate the length, charset and reserved-name rules that asobi_ws_handler already enforces at the socket boundary (#303); its own comment said "both lists must move together". In one repo they do not have to. asobi_ws_handler now exports event_name_binary/1 next to reserved_event_names/0, and asobi_lua_api:do_broadcast/4 calls it. The core guard is unchanged and still the last line of defence; this is purely the script-facing half, so an author gets { error = "..." } at the game.broadcast call site instead of a dropped event and a server-side log line. game_broadcast_rejects_all_reserved iterates asobi_ws_handler:reserved_event_names(), so a name added to core is covered without touching the Lua tests.

#132 - registration is a second layer, not an overwrite. The original called application:set_env(asobi, registration, Mode) straight from the loader. In this repo that is both an unowned writer (ADR 0006: asobi_game_config is the only one) and the exact trust inversion ADR 0006 exists to stop - a game bundle overwriting operator config. It is also a security posture, so a bundle widening a closed deployment back to open is not a trade worth making for a smaller diff.

So registration gets the same shape as the mode registry:

• asobi_game_config:apply_config/1 writes the declared value to script_registration, never to the operator's registration key;
• asobi_registration:classify/0 composes the two, operator layer on top.

A bundle can pick a posture for a deployment that states none - which is the actual bug, since an engine-hosted game has no sys.config to edit and every hosted game silently ran open - and can never widen one that does. Omitting the global leaves both keys alone; an unrecognised value is logged and dropped by the loader rather than downgrading to open.

What proves each fix landed

Every ported fix was verified by reverting the source change and confirming the test fails.

• feat(world): build per-zone state in the zone process #131 - asobi_lua_zone_spawn_tests:lua_spawn_reaches_zone/0 reads maps:get(health, Goblin) (atom key) before any tick has returned entities; asobi_lua_world_tests:spawn_templates_decodes_test/0 asserts #{hp => 10} for base_state. Reverting asobi_lua_api.erl + asobi_lua_world.erl: 2 failures.
• Persist/restore entity timers across zone reap (serialise/deserialise) #134 - assert_broadcast_rejected/1 asserts a binary error came back and ?assertNot(meck:called(asobi_match_server, broadcast_event, '_')). Reverting do_broadcast/4: 5 failures.
• fix(world): snapshot and restore per-zone state across idle reap #133 - two independent defect shapes, both re-checked here against the live database:
  • dropping the player_id IS NULL predicate from storage_find/3 fails global_write_does_not_clobber_player_row;
  • restoring the unscoped update_all/2 write fails set_twice_updates_the_row and player_write_does_not_clobber_global_row.
• Lazy-reaped Lua zones lose gameplay state on respawn #132 - registration_global_sets_mode/0 asserts asobi_registration:mode() =:= closed and check(password) =:= {deny, ~"registration_closed"} (the effective posture, not the raw key). Reverting all three source files: 3 failures. operator_registration_beats_script/0 is the guard on the rework specifically - making apply_config/1 write registration instead of script_registration fails it and nothing else.

Notes

• asobi_lua_storage_SUITE uses asobi_test_helpers:start/1 rather than the standalone pgo/kura_migrator boot the asobi_lua original needed, and ci.yml already passes extra-services-compose, so no CI change was required. No name clash with the existing asobi_storage_SUITE, which covers the core REST storage path rather than the Lua write path.
• Not ported: the asobi_lua AGENTS.md and ci.yml hunks from fix(world): snapshot and restore per-zone state across idle reap #133 (repo-specific; the equivalent asobi AGENTS.md line is included).

Checks

fmt --check, xref, dialyzer, ex_doc clean. eunit 1137/1137. ct 297 passed, 1 failed: asobi_iap_SUITE:apple_valid_jws_correct_bundle, which fails identically on a clean origin/main worktree at a5e27d3 and is untouched by this branch.

[github-actions]

🟡 Code Coverage — 73%

5629 of 7707 lines covered.