test(cold-start): hermetic SessionStart banner — kill live-env dependency (#174 follow-up)

[cdeust]

Follow-up to #174 / #180 (merged). The first full-suite run after #180 surfaced two more failures in the same test-isolation family; this PR fixes the one that is a genuine test defect.

test_cold_start.py::test_normal_session_with_memories — live-environment dependency

Root cause: the test ran session_start.py as a subprocess under the inherited real environment. The SessionStart backend gate reads the developer's live backend marker (~/.claude/methodology/backend.json); the banner then reads the real ~/.claude/methodology/memory.db (or a live PostgreSQL); and the subprocess additionally spawned detached consolidate/reanalyze workers against the real HOME and repo. The in-process remember seed landed in the per-process test DB while the subprocess read a different live store — so the assertion outcome depended on whose machine ran it (failed on this developer's populated home, passed on an empty CI home). The seed's raw UPDATE ... NOW() is also PostgreSQL-only, silently mismatched under the SQLite default.

Fix (test-side — the SQLite banner path itself is already proven correct by tests_py/hooks/test_sqlite_hook_paths.py): rewritten to the canonical hermetic pattern — seed an ephemeral SqliteMemoryStore in tmp_path and drive _sqlite_context in-process via a monkeypatched get_shared_store, asserting the seeded memory is injected and no PostgreSQL install guidance leaks. Deterministic and environment-independent; same behavior verified (DB has memories -> banner injected, not cold start).

The second failure — test_I2_no_unauthorized_heat_writes — is NOT a code defect

test_I2 is a stateless static source scan (rglob mcp_server/**/*.py, compare SET heat_base sites to a line-pinned allow-list). It holds no runtime state, so it cannot carry hidden shared state between tests. Its one observed failure occurred in a working tree a second concurrent agent was actively mutating (running its own pytest tests_py/ and editing test_I2 + pg_store.py/sqlite_store.py on another branch in the same checkout); a git checkout/edit landing mid-run shifts the pinned offsets so the scan reads source that no longer matches the allow-list. It passes standalone and in every controlled/isolated run — no code change is warranted (hardening the scan against a concurrently-mutated tree would only weaken the invariant it guards).

Gates: ruff check + ruff format --check pass. Full suite verified green in an isolated worktree (own source via PYTHONPATH, own throwaway DB via conftest), immune to the concurrent agent's mutations.

🤖 Generated with Claude Code

[cdeust]

Full-suite proof (this machine, live ~/.claude present)

Run on current main + this cold_start fix, in an isolated git worktree (own source via PYTHONPATH, own per-process throwaway DB via conftest) so the result is immune to a second agent's concurrent pytest + git checkout activity in the shared checkout:

5352 passed, 12 skipped, 9 warnings in 2673.84s (0:44:33)
EXITCODE=0

Zero failures. The three originally-failing test_rebuild_profiles cases (fixed in merged #180), the rewritten test_cold_start::test_normal_session_with_memories, and test_I2_no_unauthorized_heat_writes are all green. I2 needed no code change — it is a stateless static source scan whose one prior failure was concurrent shared-checkout mutation, not hidden state.