Issue #778: credential-safe provider diagnostics

Scope and evidence

Reviewed by Codex, with independent subagent plan review and three code-review rounds, on 2026-11-15. Source assessment: e02496fb5e845419eb7e73d18288ab44fd79cc3b.

Skill Reviewed script(s) under scripts/ Failure behavior
downtrend-duration-analyzer analyze_downtrends.py Exception class only; stable-to-v3 fallback retained
value-dividend-screener screen_dividend_stocks.py FINVIZ exception class only; empty set retained
dividend-growth-pullback-screener screen_dividend_growth_rsi.py FINVIZ exception class only; empty set retained
canslim-screener check_institutional_endpoint.py Exception class only; false return retained
parabolic-short-trade-planner check_live_apis.py Class/status diagnostics for all FMP/Alpaca/adapter failures; raw HTTP error and non-JSON bodies suppressed; malformed payload repr suppressed
theme-detector representative_stock_selector.py, finviz_performance_client.py Selector warnings carry no traceback; source failure counters/fallbacks retained; performance client already uses fixed failure reasons

Each skill has an offline scripts/tests/test_credential_diagnostics.py using synthetic credentials, including reserved URL characters. No real credential or live account is used. ConnectionError, Timeout and HTTPError messages contain both raw and URL-encoded markers; stdout/stderr must exclude both. Selector tests also inspect formatted captured logs and require exc_info is None.

The first run against the unmodified source produced 39 failures and three passes (the already-safe performance client). The corrected source passes all 52 diagnostic regressions, including ten reflected HTTP/non-JSON-body cases. All six existing skill suites pass: 78 + 217 + 196 + 434 + 327 + 556 = 1,808 tests; one optional finvizfinance test is skipped because that package is absent from the shared development environment. This is offline evidence, not live-provider acceptance or a full dependency audit.

The security collector previously confused local literal-only diagnostic exceptions with request exceptions. Its exemption now requires a plain local ValueError/RuntimeError subclass, literal-only constructors, no alias/rebinding or shadowed base, and only formatting/str/repr reads of the handler variable. Unknown/dynamic exception classes, request exceptions, logger tracebacks and mutated variables remain flagged. Fifteen new collector cases cover literal messages, unsafe subclasses/messages, aliases, rebindings and mutation. All 32 collector tests pass. This remains an advisory static heuristic, not a whole-program proof.

python scripts/backfill_security_review.py now reports no mechanical findings for any of these six skills. python scripts/check_skill_deps.py check passes for all executable skills (no warning for the six affected skills). All six .skill archives have been rebuilt and checked against their source.

Remaining acceptance gap

Acceptance criteria 1, 2 and 4 are addressed, and the collector portion of criterion 3 is addressed. The security_review: passed restoration in criterion 3 is not complete. All six entries remain not_verified, consistent with security-review-checklist.md: clean mechanical scans and targeted leakage regressions cannot certify all credential, data-exposure, path, I/O/network and dependency dimensions for the entire skill.

The targeted review verified the changed diagnostics, their existing bounded request timeouts and failure/fallback behavior, dependency declarations, and source/package parity. It did not comprehensively review all other source paths or audit third-party library behavior. In particular, theme-detector’s optional FINVIZ dependency was absent locally, so its network and logging behavior was not established by these tests. A complete per-skill checklist review remains necessary before changing these metadata values. Keep #778 open until that acceptance is separately recorded; this change must not claim full security certification or live validation.