fix(tui-gateway): WS disconnect/reconnect TOCTOU fix + real regression test (RAH-05 + RAH-06)

[JoaoMarcos44]

Summary

Closes #77192 (RAH-06), closes #77191 (RAH-05).

Combines the WS disconnect/reconnect TOCTOU fix (RAH-06) with its regression test fix (RAH-05) into one PR — RAH-05's test exercises exactly the revalidation logic RAH-06 introduces, so splitting them into two PRs meant the second one couldn't pass CI on its own without the first merged first. Together they tell one coherent story: a real race, a real fix, and a test that actually proves the fix works.

RAH-06 — WS disconnect/reconnect TOCTOU race

While validating a prior investigation's claim that this race was already fixed, found that the fix commit existed only on branch fix/ws-disconnect-reconnect-transport-race — never merged into main:

$ git merge-base --is-ancestor 13241a6a3 main
# not an ancestor

main's _close_sessions_for_transport() had the original bug: a session snapshot taken under _sessions_lock, released, then each session mutated (closed or re-pointed to the detached sentinel) without re-validating transport ownership — racing against session.resume's warm-reuse rebind of session["transport"] under _session_resume_lock. A reconnect landing in that window could be silently undone: the disconnecting transport's teardown either force-closed the just-reattached session, or stomped its new transport back to the detached sentinel, making a live reconnect look orphaned and eligible for grace-reap.

Fix: cherry-picked the existing, already-authored fix (13241a6a3) — applies cleanly onto identical surrounding code. Revalidates transport ownership under the same _session_resume_lock → _sessions_lock ordering already used by the orphan-reap timer, immediately before claiming (close) or repointing (detach) each session. Slow teardown work still runs after releasing both locks.

RAH-05 — the existing regression test didn't exercise the race

The pre-existing regression test for this exact fix
(test_close_sessions_for_transport_skips_session_reattached_mid_teardown)
started both test sessions already pointing at new_transport. Since
_close_sessions_for_transport() filters owned_sids by the old
transport, those sessions never entered the snapshot in the first place
— the revalidation-under-lock logic RAH-06 adds was never reached. The
test passed identically before and after RAH-06's fix and proved
nothing about it.

Fix: rewrote the test to start the session on old_transport (so
the snapshot captures it) and inject the reattach — via a thin wrapper
around the real _session_resume_lock — strictly between the snapshot
and the per-sid claim, matching the actual race window.

%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#00f0ff', 'mainBkg': '#0a0a16', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#ff007f', 'lineColor': '#00f0ff'}}}%%
graph TD
    A[🔌 WS disconnect] -->|snapshot owned sids| B[⚡ _sessions_lock]
    C[🔁 session.resume reattach] -->|rebind transport| D[⚡ _session_resume_lock]
    B --> D
    D -->|revalidate ownership per sid| E{Still old transport?}
    E -->|yes| F[🚀 Claim: close or detach]
    E -->|no: reattached| G[✅ Skip — live reconnect preserved]

    H[🧪 Old test: sessions start on new_transport] -->|never in owned_sids| I[🚫 Revalidation branch never runs]
    J[🧪 New test: session starts on old_transport] -->|_RaceLock reattaches mid-claim| D

Loading

Infographic :

Test plan

• tests/test_tui_gateway_server.py -k close_sessions_for_transport — 2 passed
• New test confirmed to fail against the pre-RAH-06 implementation (reaped == 1, not 0)
• python -m py_compile tui_gateway/server.py tests/test_tui_gateway_server.py

Supersedes #77205 and #77206 (closed — split into separate PRs originally, recombined here since one depended on the other).

[alt-glitch]

This was generated by AI during triage.

Related to #77129: it carries the same core repair, while this PR adds the real regression interleaving identified in #77191. Please choose or consolidate the implementation and coverage.

[JoaoMarcos44]

Duplicate root cause with #77129 (opened independently for the same TOCTOU bug, tracked as #77127 there). Consolidated: pushed this PR's real regression-race test on top of #77129's branch (fix/ws-disconnect-reconnect-transport-race), replacing its shallow version. #77129 now carries the fix + the real interleaving test, and closes #77127, #77191, and #77192. Closing this one in favor of #77129.