sdkharness: repository-owned leaf checks refuse to run outside a loopback simulator

[sophiecarreras]

Why

An independent review of the repository-owned sdkharness tests (merged in #611 and #613) found that the 33 conformance leaf checks, which are executable directly, default CONFORMANCE_TARGET to staging, map staging/production realms, and fall back to ambient B2_APPLICATION_KEY_ID / B2_APPLICATION_KEY. The dispatchers (run-conformance, run-resilience) were safe, but a developer running a leaf by hand with real credentials in the environment could create, mutate or delete resources in an external account. This contradicts .sdkharness/README.md ("never use a production B2 endpoint or real credentials").

What

• New .sdkharness/tests/lib/loopback_guard.py: refusal(level) returns a reason unless the target is simulator, <LEVEL>_SIMULATOR_URL is http://127.0.0.1:<port>, an optional <LEVEL>_SIMULATOR_HTTPS_URL is https://127.0.0.1:<port>, and (resilience) RESILIENCE_CONTROL_URL is a loopback origin.
• All 33 conformance and 16 resilience leaves call it first and print the standard FAIL (configuration -- ...) line with exit 1. Default target is now simulator; the staging/production realm entries and the ambient-credential fallbacks are removed (the fixed test-key-id / test-key credential is the only one). client.simulator loses unused realm and ambient-credential helpers (it only drives the in-process RawSimulator with the network disabled).
• README: how to run one health, one conformance and one resilience check through the dispatchers against a local simulator.

Verification

• 71 unit tests pass (test/unit/test_sdkharness_conformance_resilience_contract.py and test_sdkharness_contract.py); the new tests (54 failing cases) fail on the previous leaves and pass now.
• Ran through the dispatchers against a local simulator: files.upload, bucket.crud, enc.sse_c, client.simulator PASS; api.backoff_503 PASS; upload.retry_408 FAIL as before (the known Retry-After/408 SDK findings). A direct leaf run with CONFORMANCE_TARGET=staging and ambient keys exits 1 with FAIL (configuration ...).
• ruff check and ruff format --check clean on .sdkharness and the test file. Not run locally: the full nox matrix and integration suite (read from CI).
• Changelog fragment: +sdkharness-loopback-guard.infrastructure.md (no related issue; the contributing guide's convention for PRs without an issue).

Review requested from the maintainers who merged recently. Nothing here changes SDK code.

🤖 Generated with Claude Code

[sophiecarreras]

Follow-up commit e678272 for the independent review of this PR:

• Health leaf credential (medium): .sdkharness/tests/health-golden-path accepted any non-empty B2_TEST_APPLICATION_KEY_ID/B2_TEST_APPLICATION_KEY and passed it to authorize_account. It now refuses anything but the simulator's fixed test-key-id / test-key pair before the SDK is touched (the supplied values are never echoed), and uses the fixed pair for the call. I re-read every conformance, resilience and library file in .sdkharness/: the conformance leaves already use only the fixed pair and the resilience leaves hard-code it, so the health leaf was the only path.
• IPv4-only (consistency): the health leaf also accepted ::1; it now accepts only http://127.0.0.1:<port> like the conformance and resilience leaves (README states this).
• README startup race (low): the copy-paste sequence now waits (up to ~10 s) until the simulator has printed its three listener lines before reading its URLs. I ran the sequence verbatim against a local b2-simulator main: conformance files.upload and the health check PASS; a real-looking key is refused with configuration: only the fixed simulator credential is accepted.
• Tests: 3 new cases (real-looking key / id / both; the SDK must never be called) and an ::1 case; they fail on the previous commit (4 failures) and pass now (75 passed in test_sdkharness_*). ruff check clean on the touched files with the pinned ruff 0.8.x. (The two files ruff format --check flags, noxfile.py and doc/sqlite_account_info_schema.py, are unformatted on master too.)

Not run here: the full nox matrix. CI on this PR shows the known unrelated live-encryption failures (#614) — I have not diagnosed them.