test(acp): the dogfood server gets its own session dir, not the real one

The live-Kiro dogfood test keeps the developer's real HOME on purpose:
kiro-cli's auth and the installed `studyloop` agent live there, and running
against them is the point. The session DIRECTORY is a different axis. The
server it spawns resolves session_state.SESSION_DIR from
STUDYLOOP_SESSION_DIR, defaulting to the real ~/.config/studyloop, and the
conftest sets that variable per TEST in a function-scoped autouse fixture.
This server starts from a module-scoped fixture, before any test's fixtures
run, so `{**os.environ}` carried no override and every run wrote an
`acp-sql-<id>` directory (CLAUDE.md, the `studyloop` wrapper) into the
developer's real sessions folder. Nine had accumulated across the runs that
reached a verdict; the conftest's R-49d real-home guard named the tenth on
the first run after the trust-write and adapter fixes.

Every other spawned server gets its session dir from
_playwright_helpers._isolated_child_env; this launcher bypasses that helper
only to keep HOME, so it now sets the one variable itself (a fresh mkdtemp).

Verified live, twice, against real kiro-cli 2.29.0 after c1accbc:
run 1 (before this change) 1 passed in 47.6 s with the R-49d banner naming
the new real-home directory; run 2 (with it) 1 passed in 66.7 s, no banner,
no new entry under ~/.config/studyloop/sessions, one dir under TMPDIR.
Both runs also confirm the two fixes beneath them: ~/.claude/settings.json
was byte-identical (same SHA-256, size and mtime) across a real Kiro ACP
start and end, and the adapter teardown classified the 17 Sep
study-mentor.json.studyloop-backup as obsolete and left the installer's
symlink untouched (the discard itself hit this sandbox's EPERM on
~/.kiro/agents and was caught by cleanup._teardown_agent, so the session
end still completed; outside the sandbox it is removed).

Author: NetDevAutomate

CHANGELOG.md

@@ -181,7 +181,16 @@ experience may change before `1.0.0`.
   the real agent runs its session-start protocol (several tool calls, each a
   permission round-trip) before answering, and a timeout now reports the
   console's state instead of a bare "120000ms exceeded". First green run:
-  48 s.
+  48 s. The server it spawns also wrote each session's directory (`CLAUDE.md`,
+  the `studyloop` wrapper, IPC files) into the developer's real
+  `~/.config/studyloop/sessions/`: the test keeps the real `HOME` on purpose
+  (Kiro's auth and the installed `studyloop` agent live there), but the
+  module-scoped server fixture starts before the conftest's per-test
+  `STUDYLOOP_SESSION_DIR` isolation applies, so the server inherited no
+  override. The launcher now gives the server its own throwaway session dir,
+  as every other spawned server already gets from `_isolated_child_env`.
+  Found by the conftest's R-49d real-home guard on the first run after the
+  fixes above (nine `acp-sql-*` directories had accumulated).
 - The `voice-tts` OpenSpec specification described the architecture that was
   deleted, not the one that shipped: it required browser-side Kokoro via
   `transformers.js` and forbade any server-side synthesis for the web path,

packages/studyloop/tests/test_web_acp_dogfood_kiro.py

@@ -26,6 +26,7 @@
 import shutil
 import subprocess
 import sys
+import tempfile
 import time
 import urllib.error
 import urllib.request
@@ -160,6 +161,18 @@ def _start_web_server_real_kiro() -> subprocess.Popen:
     # developer's real, authenticated kiro-cli and StudyLoop's installed
     # `studyloop` agent -- that is what makes it dogfood.
     env.setdefault("SESSION_CONTEXT_SCOPE", "unclassified")
+    # The session DIRECTORY is not part of that: it is where the server
+    # writes CLAUDE.md, the `studyloop` wrapper and the IPC files for each
+    # session, and session_state.SESSION_DIR defaults to the REAL
+    # ~/.config/studyloop unless STUDYLOOP_SESSION_DIR says otherwise. The
+    # conftest autouse fixture sets that variable per TEST, but this server
+    # starts from a module-scoped fixture, before any test's fixtures run,
+    # so `{**os.environ}` carried no override and every dogfood run left an
+    # `acp-sql-<id>` directory in the developer's real sessions folder
+    # (R-49d banner, 2026-10-12). Every other spawned server gets this from
+    # _playwright_helpers._isolated_child_env; this launcher bypasses that
+    # helper only to keep HOME, so it sets the one variable itself.
+    env["STUDYLOOP_SESSION_DIR"] = tempfile.mkdtemp(prefix="studyloop-dogfood-session-dir-")
 
     cmd = [sys.executable, "-m", "studyloop.cli", "web", "--port", str(WEB_PORT)]
     proc = subprocess.Popen(