[Taure]

world_input_crossing_touches_only_ring_delta/0 passed in a full eunit run and failed whenever the module was run on its own. That is the worse way round of the two: it was green for a reason that had nothing to do with the ring diff it exists to prove.

Cause

subscribe_interest_zones/4 resolves a player through find_player_pid/1, which reads pg:get_members(nova_scope, {player, PlayerId}). The test joined p1 without registering anything under that key, so the pid came back undefined, asobi_world_server skipped the subscribe entirely (asobi_world_server.erl:1056-1059), and every subscriber count in the test was 0.

In a full run it found a live process another module had left registered under {player, <<"p1">>} in the shared nova_scope, and passed on that.

Fix

The module already has fake_session/2 for exactly this, and its own documentation says so - "a live, pg-registered session for a player, so find_player_pid/1 ... resolve to a real process instead of silently falling through". Six other tests in the module use it. This one now does too, with a distinctive player id, since p1 is what made it borrowable in the first place.

Verification

By substitution rather than by the test simply going green: swapping the pg-registered session for a plain unregistered process reproduces the original failure exactly (expected: 1, got: 0). The assertion is now carried by the subscribe path rather than by whatever else happens to be running.

Also narrows two sys:get_state/1 reads in the same module - they are term(), and touching the module brings it into CI's EQW_SCOPE=diff gate, which requires zero errors in every module a PR touches. eqwalize-all: 195, from 197.

Checks

fmt --check, xref, dialyzer, elp lint (no new findings), eunit 2,398/0, CT 390/0, and the module now passes standalone.

Found during the review gate on #566.

---

Second commit: fix the mechanism, not just this test

nova_scope is shared by every module in a run and {player, Id} is not namespaced, so a registration that outlives its test is lent to any later test using the same id. That has now cost two silent wrong-reason passes, so the second commit fixes the mechanism.

fake_session/1,2 was hand-copied into four modules in two variants that differed only in whether they forwarded messages. It now lives in asobi_test_helpers, alongside:

• unique_session/1, deriving the id from unique_id/1. A distinctive id works only until someone else picks it - the same guarantee p1 had.
• assert_no_session/1, for tests that want the session-less path. Without it a leftover registration silently turns such a test into a different one, which is how this class stayed hidden.

The donor. asobi_world_chat_tests registers self() - the eunit runner, alive for the whole run - under {player, <<"p1">>} in seven tests, with the pg:leave as a trailing call rather than an after. Those tests assert in the middle, so a failing assertion skipped the leave and stranded the runner in the group for the rest of the run. with_registered_player/1 makes the pairing unconditional.

asobi_match_server_tests had two sessions spawned and never killed at all; both now released in an after.

Also narrows two sys:get_state/1 reads in asobi_world_server_tests, since CI eqwalizes every module a PR touches. eqwalize-all: 193, from 212 on main.

eunit 2,398/0, CT 390/0.

Gate

Architecture guardian: ACCEPT, and it caught a real flaw in the first commit - the session and world were released only from a block the failing assertion jumped over, recreating the leak class on the failure path. Fixed in 80b1229.

Security reviewer: the diff is clean; it found the production defect underneath, now fixed separately in #569.

A correction to an earlier version of this description. It said the donor was asobi_world_chat_tests and that the guardian had misattributed it. That is backwards. Every unregister_player() in the chat module was paired, so a green run left nothing behind - what that module had was the failure path. The donor was asobi_match_server_tests, which had two _-prefixed sessions on p1 and killed neither, permanently registered, in a module that sorts first. The guardian had it right; the security review did not, and this description repeated the security review.

Code reviewer: found that the helper five modules now depend on had no tests - and that who/1 had already regressed inside this PR, caught by review rather than by a test. Also that release_session/2 had no callers outside with_session/3, and that prop_reconnect_lifecycle was accumulating up to ~125 registered sessions per run. All worked in the final commit.

[github-actions]

🟡 Code Coverage — 76.9%

9396 of 12223 lines covered.

---

🟡 ELP Lint — 190 warnings

210 diagnostics found. See job logs for details.

---

✅ ELP eqWAlize

No diagnostics.