[JordanCoin]

Closes #130: records search, --all-sources, fetch (documentcloud, nara) and download (muckrock) ignored network.tor and went out on the real IP.

With {"network":{"tor":true}} in config.json, records.get_adapter() built every adapter with the RecordAdapter default policy (DIRECT). These commands went out on the real IP with no Tor check and no route notice:

• records search -s <source> (every source except cia)
• records search --all-sources (all sources, cia included, skipping the Tor gate)
• records fetch -s documentcloud, records fetch -s nara without --objects/--list-objects
• records download -s muckrock

What changed

• get_adapter(name, *, egress=None) passes the policy to the adapter constructor (RecordAdapter.__init__(egress=...), the same mechanism the cia/sec/nara-objects paths already used).
• Every records CLI call site takes its policy from _records_policy_or_exit() (renamed from _cia_policy_or_exit, already source-agnostic): _egress_policy_from(load_config()) plus the fail-closed _check_tor_or_exit. Each command prints the route ("via Tor, real IP hidden" / "direct, real IP") before sending. --all-sources checks Tor once, before any request.
• The cia search and nara --objects paths no longer build a second adapter; they reuse the one from get_adapter.
• No adapter changes: every adapter in openfoia/records/ builds its client via egress_client(self._egress, ...).
• Tests (tests/test_security_egress_cli.py): for search (single and all-sources), fetch (documentcloud, nara) and download (muckrock), with Tor on in config, every adapter built has a TOR policy; with Tor unreachable, exit is non-zero and zero HTTP requests are made. One test asserts --all-sources builds cia under TOR. All 11 new tests fail on main.
• Existing tests that stub get_adapter now accept **_.

Composition with #128

#128 adds get_adapter(name, api_data_gov_key=...) at the same call sites. Whichever merges second needs a small rebase: both kwargs pass through, and get_adapter should withhold api_data_gov_key when egress.is_tor (the key identifies the user).

Local: ruff check + format clean; pytest tests -q 573 passed, 2 failed (pre-existing TestLLMExtraction, need Ollama).

🤖 Generated with Claude Code

https://claude.ai/code/session_01FWQw8DDrDZ2tFzswSSEtce

[JordanCoin]

Review (Claude opus reviewer, read-only, 2026-11-06): MERGE. Enumerated every adapter construction site in cli.py, crossref.py, pipeline/web.py, agent.py, server.py, campaign.py, cases.py: all now carry the config policy or exit before building anything; no DIRECT default remains. Route notice goes to stderr after the Tor check and names the route actually taken; --all-sources checks Tor once before any request including the nara has_key probe; check_tor never raises; renamed helper is byte-identical to the old one and no stale references remain; fail-closed tests prove zero requests via a patched AsyncHTTPTransport that both the Tor and retry transports terminate in.

This branch now also carries #128 (merged in, conflicts resolved as the reviewer specified): get_adapter takes egress and api_data_gov_key, and passes the literal DEMO_KEY over Tor. New test covers that with OPENFOIA_GOVINFO_API_KEY set. Full suite 583 passed; the 2 failures are the pre-existing Ollama tests.

Below threshold, tracked separately: NARA key, OPENFOIA_MUCKROCK_TOKEN and OPENFOIA_GOVINFO_API_KEY still travel over Tor on paths outside this diff; openfoia browse ignores network.tor unless --tor is passed; the Tor-down error goes to stdout under --json.