MCPg v0.7.1

Changed

• New .mcpb bundle icon (packaging/mcpb/icon.png). Replaces the
  previous database-cylinder-plus-wordmark icon with an abstract
  "convergence" mark — several streams meeting at one bright point —
  designed to hold up at the small sizes directory listings actually
  render icons at, without relying on embedded text.

Security

• Code-scanning triage (CodeQL + OpenSSF Scorecard). Removed the
  unused, broken apisec-scan.yml workflow (pointed at a demo project
  with no real credentials). Added least-privilege permissions: blocks
  to the three ci.yml jobs CodeQL flagged
  (actions/missing-workflow-permissions) — all three only check out the
  repo and run local commands, so a single top-level contents: read
  covers them. Pinned Docker base images (Dockerfile,
  local-postgres.Dockerfile, .github/ci-postgres-warehousepg.Dockerfile,
  .github/ci-postgres-pg19.Dockerfile,
  scratch/pg_turboquant/docker/Dockerfile.dev) to @sha256: content
  digests; .github/ci-postgres.Dockerfile is intentionally left
  unpinned (parameterized by PG_MAJOR to drive the CI matrix) with a
  comment explaining the exception. Pinned the unversioned
  pip install huggingface_hub step in publish.yml's (non-blocking)
  HF-Space-refresh job to ==1.26.0.
• Hardened SqlDriver's query-error log line against clear-text secret
  leakage (CodeQL py/clear-text-logging-sensitive-data, high severity).
  _execute_with_connection's exception handler logged the raw query
  text verbatim. Root-caused: this is a real, reachable path —
  mcpg.redis_fdw.create_redis_user_mapping resolves a real secret from
  the configured secrets backend and interpolates it into a
  CREATE USER MAPPING ... OPTIONS (password '...') statement (SQL
  literal syntax, no = sign) before calling execute_query; if that
  call raises, the literal secret would have reached the log verbatim.
  obfuscate_password's existing DSN-style patterns (password=...)
  didn't match this =-less SQL-literal form. Added a new
  obfuscate_password pattern for bare-keyword SQL literals
  (password '...' / IDENTIFIED BY '...') and routed both the logged
  query and exception text through obfuscate_password in
  src/mcpg/sql/driver.py. Regression tests added in
  tests/unit/test_sql_kernel_obfuscate.py and
  tests/unit/test_sql_kernel_driver.py — confirmed against
  mcpg.redis_fdw.create_redis_user_mapping's actual generated SQL, so
  the pattern match is exact, not assumed. The test_sql_kernel_driver.py
  regression test as first committed (WIP commit 038b5dd) never actually
  exercised the fix — it needed two follow-up corrections (a broken mock
  of the async-context-manager protocol that caused a genuine infinite
  loop, then a test-isolation fix for cross-test logger-handler leakage
  from test_obs_logging.py) before its redaction assertion could
  genuinely fire; see the fix commit's message for the full root-cause
  trail. Verified redacted end-to-end (password '****' present, secret
  absent) as part of a full green pytest tests/unit tests/contract run.
• Fixed an incomplete-redaction bug in the SQL-literal obfuscate_password
  pattern added above. Its literal-body group was a plain [^']+, which
  stops at the first single quote — but create_redis_user_mapping
  escapes an embedded ' in the secret by doubling it (standard SQL
  string-literal syntax: password.replace("'", "''")), so a password
  containing a literal quote (e.g. it's-a-secret → it''s-a-secret) was
  only partially redacted (password '****''s-a-secret' — the real
  password's tail leaked in clear text). Changed the group to
  (?:[^']|'')*, which consumes doubled-quote pairs as part of the
  literal so the whole secret is matched; confirmed not vulnerable to
  catastrophic backtracking (the two alternatives never overlap at a
  given position — stress-tested at 50k adversarial chars, ~2ms).
  Regression test added in tests/unit/test_sql_kernel_obfuscate.py.
• Reviewed the CodeQL py/incomplete-url-substring-sanitization alert on
  tests/unit/test_secrets.py:234 — a test-only false positive (a
  hardcoded-literal assert "vault.example.com" in rendered debug-repr
  check, not a URL-sanitization decision on untrusted input); suppressed
  with an inline codeql[...] comment explaining why.
• Bumped the transitive cryptography dependency (pulled in via
  pyjwt[crypto] / google-auth, no direct pyproject.toml pin) from
  49.0.0 to 50.0.0, resolving CVE-2026-69247 / GHSA-g6cj-pr64-35w5 (a
  Bleichenbacher timing oracle in PKCS#7 EnvelopedData decryption).
  main's branch protection requires the Security audit (SAST & Dependencies) check (pip-audit --strict), which this advisory was
  failing; uv.lock only — verified with the same uv export --no-dev --no-emit-project --format requirements-txt + pip-audit --strict --disable-pip invocation CI runs, and confirmed no JWT/OIDC test
  regressions (tests/unit/test_oidc.py, test_http_runtime.py,
  test_config.py).