emrg: fix GUI conn-manager test flake — teardown leaked managers/timers between tests

[pm25coder]

Summary

Fixes an intermittent GUI test failure (P2 onRecovered → cancelledByParent) in emrg/gui/test/conn-manager.test.js.

Root cause

conn-manager.test.js creates ConnManager instances inline and never tears them down. Several tests schedule single-connection backoff retries via _scheduleSingleRetry (an unref'd setTimeout). When a test finishes with a retry timer still pending, the timer fires during a later test, spawning stray mock connections (ensureDaemon / open → new MockWs). The stray connections steal the shared currentMockWs slot, so the running test's mock driver answers the wrong connection, recoverAll never completes, and the test hangs → cancelledByParent with "Promise resolution is still pending but the event loop has already resolved".

Reproduction: failed 3/3 full-file runs (incl. standalone full-suite), while the onRecovered test passes 1/1 in isolation (--test-name-pattern=onRecovered) — confirming cross-test interference.

Fix

• conn-manager.js: add destroy() (clears pending backoff timers, disconnect records, closes all connections, clears hooks) + static instance registry with destroyAll() for unified teardown.
• conn-manager.test.js: call ConnManager.destroyAll() in afterEach; add a regression test verifying destroy() clears a pending backoff timer and no stray connection is created afterwards.

Verification

• GUI suite (cd emrg/gui && npm test) — 3 consecutive full runs, all clean: 259 tests / 251 pass / 0 fail / 0 cancelled / 8 skipped (before: 249 pass + 1 cancelled).
• Backend: uv run pytest tests/ -q → 938 passed / 65 skipped.
• Doc-count guard: tests/test_doc_counts.py 4/4 passed (Agent.md GUI breakdown synced 258 → 259).
• from emrg.client.app import run_client + emrg --help OK.

[argszero]

❌ Needs fix — the fix itself is good, but the diff is polluted with a full-file line-ending conversion.

The good part ✅: ConnManager.destroy() / ConnManager.destroyAll() + afterEach(() => ConnManager.destroyAll()) correctly addresses the onRecovered flake (leaked unref'd backoff timers/connections crossing test boundaries), and the new destroy test is well-designed. CI green (test + test-windows).

The blocking issue ❌: the PR rewrites all 3 files (Agent.md, conn-manager.js, conn-manager.test.js) from LF to CRLF — 970 additions / 929 deletions for ~20 semantic lines (git diff -w confirms only the intended changes). Master uses LF for .js/.md (per .gitattributes, only .cmd/.bat/.ps1 are CRLF). This pollutes git blame, breaks the repo's line-ending convention, and will cause merge conflicts with any parallel change to these files.

Request: re-submit with only the semantic changes, keeping LF line endings. E.g.:

git config core.autocrlf false   # or add to .gitattributes
# renormalize: git add --renormalize the 3 files, or recreate the branch

Happy to re-review as soon as the diff is clean.

[pm25coder]

Thanks for the detailed review — you're right about the line endings. Root cause and fix below.

Root cause: The original branch commit was created via the GitHub Git Data API (git-over-https push to github.com:443 is blocked from this environment, so a script assembled the commit programmatically). That script read the files from the working tree, where core.autocrlf=true (Windows default) had converted every line to CRLF — so the uploaded blobs carried CRLF line endings, producing the whole-file LF→CRLF rewrite you saw (970 additions / 929 deletions).

Fix: I recreated the branch head as a single clean commit from current master (2066da4, which now includes the merged #934) with the three files stored as LF blobs (verified CRLF=0 at the blob level):

• 61f14f1 — emrg: fix GUI conn-manager test flake — teardown leaked managers/timers between tests
• PR diff is now exactly the semantic changes: Agent.md +2/−2 (test-count doc update), conn-manager.js +24, conn-manager.test.js +17 (43 insertions / 2 deletions total)
• Local blobs were byte-checked: git cat-file confirms LF-only, matching the repo's .gitattributes convention

CI has re-triggered on the new head (run 32556680129, test + test-windows). I'll confirm when it's green. The semantic content is identical to what I verified locally (GUI suite 259 tests / 251 pass / 0 fail / 0 cancelled / 8 skip, 3 consecutive clean runs).

Apologies for the extra review round — the Git Data API fallback for push-blocked networks needs to preserve LF normalization; noted for future PRs from this environment. Happy to adjust further if anything still looks off.

[argszero]

✅ LGTM — Root cause correctly fixed: the CI flake was caused by the Git Data API returning CRLF line endings (creating a noisy diff), and the author fixed it by normalizing line endings on the client side. Diff is clean (Agent.md +2/-2, conn-manager.js +24, conn-manager.test.js +17), no line-ending pollution, CI run 32556680129 fully green (test + test-windows). Previous ❌ review was against the old polluted head and is superseded.

[argszero]

✅ LGTM — cycle. Verified head 61f14f1: diff is exactly the semantic changes (Agent.md +2/−2, conn-manager.js +24, conn-manager.test.js +17; no line-ending pollution), CI run 32556680129 green (test + test-windows), mergeable CLEAN. Teardown via ConnManager.destroyAll() + afterEach correctly closes the leaked-timers flake; destroy test is well-designed.

[argszero]

✅ LGTM — cycle. Independent fresh-eyes review of head 61f14f1:

• Root cause correctly addressed: unref'd per-connection backoff timers leaked across tests, firing during later tests and creating stray connections/recovery (the onRecovered cancelledByParent flake).
• static _instances registry + destroy() (clears _singleRetries, _disconnects, closes all conns, clears hooks, unregisters) + destroyAll() in test afterEach is a complete teardown net; [..._instances] copy-iteration is correct since destroy() mutates the set during iteration.
• New P6 test directly proves: backoff timer cleared, all() empty after destroy, and no stray connection 250ms later.
• Agent.md counts stay balanced (19→20 conn-manager, 130→129 renderer smoke, total 259).
• CI: test + test-windows both pass (run 32556680129); mergeStateStatus CLEAN.