fix(v1.6): harden backend and reliability boundaries - #152

… cleanup

- tests/test_config.py: split multi-import to satisfy ruff E401
- backends/retention.py: redact third-party RuntimeError in strict mode
  so provider credentials/cert paths cannot reach service logs; only the
  locally generated missing-key message passes through unchanged
- core/store.py: Store.close() now calls connector.close() (best-effort)
  so the SQLCipher key pragma is cleared on explicit shutdown, not just
  when the Store object is garbage-collected

Address two Codex review findings:

- scripts/start_dashboard.py: the broad Exception handler during eager
  config validation now emits a value-free diagnostic. UnsafeStateFile
  and other OSError subclasses can embed the configured ENGRAPHIS_ENV_FILE
  path; previously these bypassed the ValueError-specific sanitized branch
  and exposed tenant-identifying paths in service logs.

- engraphis/backends/extractor.py: guard client.close() calls in both
  llm_structured and llm strict-mode credential-rejection paths. If
  HTTP-client cleanup raises (proxy credentials, TLS failures, cert
  paths), the exception must not replace the sanitized missing-key
  diagnostic that dashboard startup surfaces.

Address three Codex review findings:

- config.py: normalize relay URL scheme to lowercase before validation.
  URL schemes are case-insensitive per RFC 3986; previously HTTPS://
  rejected while the downstream relay validator accepts it.

- mcp_server.py: eagerly construct the service when
  ENGRAPHIS_REQUIRE_EXACT_BACKENDS=true so missing models or credentials
  fail the process before the stdio/HTTP transport accepts traffic,
  matching the documented startup-failure contract.

- ledger.spec.js: wait for the debounced scene reload request after
  clearing the repository filter. The 250ms debounce plus network fetch
  made the assertion race the in-flight reload; waiting for the
  unfiltered /api/graph/scene request stabilizes the test.

Clearing the filter updates the renderer via setRepoFilter() but the
#graph-count text comes from the last server scene response. In
overview mode the debounced scheduleGraphRepositoryReload() does not
fire (only triggered for code overlay or full mode), so the count
stays at the filtered value. Click Reload data to fetch the
unfiltered scene, wait for the request, then assert.

… redacted selector errors

Address three Codex P2 findings:

- mcp_server.py: extract _eager_exact_backend_check() shared helper;
  mcp_http_cli.py and mcp_classic_cli.py now call it before accepting
  traffic, matching the startup-failure contract of engraphis-mcp.

- config.py: __post_init__ re-checks raw ENGRAPHIS_VECTOR_BACKEND
  against the known set when require_exact_backends=True. Previously
  _parse_vector_backend silently replaced typos with numpy, hiding the
  configuration error from the factory.

- extractor.py, graph_extractor.py: exact-mode rejection messages no
  longer interpolate the configured selector value (which could be an
  accidentally pasted credential). Report the setting name and valid
  kinds instead.

- test_mcp_server.py: update mock to include _eager_exact_backend_check
  so the DNS rebinding guard test stays green.

SimpleNamespace lacks __spec__/__loader__ attributes that the Python
import machinery requires for 'from X import Y' statements. This caused
ImportError on CI Python 3.10-3.14 when the test mock replaced
engraphis.mcp_server in sys.modules. Use types.ModuleType which
satisfies the import protocol across all supported versions.

Address three Codex P2 findings:

- core/store.py: Store.close() no longer calls connector.close(). An
  injected connector (connect parameter) may be shared across multiple
  Store instances — closing it blanks the SQLCipher key pragma and
  breaks surviving stores' _open_connection() calls (verified backups,
  secure-erasure helpers). The injector owns the connector lifecycle;
  Store never closes what it didn't create.

- mcp_http_cli.py, mcp_classic_cli.py: wrap the new
  _eager_exact_backend_check import in try/except ImportError so test
  mocks that replace engraphis.mcp_server in sys.modules with a minimal
  stand-in get a no-op fallback. Production callers always find the
  real symbol; tests that mock the module stay green across Python
  3.10-3.14 without requiring __spec__ on the mock.

…tor in exact mode

Address two Codex P2 findings:

- config.py: remove interpolated port and embed_dim values from
  out-of-range ValueError messages. Dashboard startup prints these
  verbatim; a parseable but out-of-range numeric secret (e.g. a pasted
  token that happens to look like a number) would otherwise reach
  operator logs. The setting name and valid range are sufficient for
  diagnosis.

- config.py: __post_init__ exact-mode vector backend check now rejects
  blank and whitespace-only values, not just unrecognized non-empty
  strings. Previously ENGRAPHIS_VECTOR_BACKEND='' with
  require_exact_backends=true passed the truthiness guard and silently
  degraded to numpy via _parse_vector_backend.

…wn\n\nMemoryService.create() builds the SQLCipher connector via connector_from_env()\nbut never closed it — Store.close() correctly refuses to close an injected\nconnector (it may be shared), and MemoryEngine doesn't retain a reference.\nThe SQLCipher key pragma therefore survived MemoryService.close() until GC.\n\nAdd owned_connector tracking: create() passes the connector it built to\n__init__, and close() clears the key material after the Store shuts down.\nInjected connectors (caller-owned) remain untouched.\n\nAddresses Codex review thread: encrypted_db.py:204 (P2)

… raises

Wrap engine.close()/store.close() in try/finally so the SQLCipher key
pragma is always cleared, even when a backend collaborator fails during
shutdown. Previously the connector cleanup ran only on the success path.

Addresses Codex review thread: service.py:1232 (P2)

…iagnostics

The _startup_error fallback branches for ValueError and RuntimeError were
echoing third-party exception messages verbatim. AutoTokenizer.from_pretrained
and provider client constructors can embed configured model paths, endpoint
URLs, or proxy credentials in their error text.

Replace both fallbacks with value-free diagnostics pointing to
engraphis-init --check. The known-safe branches (trusted config,
require_exact_backends) retain their actionable messages.

Addresses Codex review thread: start_dashboard.py:196 (P2)

The trusted config / environment variable substring check in _startup_error
still echoed arbitrary third-party ValueError messages verbatim when the
message happened to contain those phrases. A library raising
ValueError("environment variable failure loading C:/tenant/private/...")
would pass the check and leak the path.

Replace with unconditional value-free diagnostic. All ValueErrors from
dashboard construction now route to engraphis-init --check.

Addresses Codex review thread: start_dashboard.py:199 (P2)