test(mesh/audit): pin PSK-verified seq-seed circular-trust defence (85%->86%)

[cagataycali]

Problem

When the audit seq sidecar is unusable (missing/corrupt/symlinked), _load_seq_counters() re-seeds _SEQ_COUNTERS by walking the persisted audit log and taking max(seq) per peer. That walk is a trust boundary: it restores the monotonic-sequence floor replay protection depends on.

The circular-trust defence on that boundary is: when STRANDS_MESH_AUDIT_PSK is configured, the seed walk trusts ONLY records whose HMAC sig it can recompute and compare_digest-match. Without it, an attacker who could append a single forged record (no PSK in dev, or a PSK exfiltrated long enough to write one line then cleared) could plant seq=10**9; on the next restart with a degraded sidecar that value would silently become the per-peer floor, denying the legitimate writer ~a billion working sequence numbers.

The existing recovery/cap tests covered only the no-PSK posture, leaving this PSK-verify branch (the actual defence) unexercised.

Change

Adds tests/mesh/test_audit_seq_psk_verified_seed.py pinning the PSK-present seed walk:

• a correctly-signed record seeds the seq floor (and _next_seq advances past it),
• an unsigned record and a PSK_DEGRADED poison sentinel are skipped,
• a wrong-HMAC record fails compare_digest and is skipped (no poisoned floor),
• an over-cap but correctly-signed record is refused by the shared _MAX_SEED_SEQ cap,
• a log of only-unverified records seeds nothing (fail-closed, with the circular-trust WARNING).

Test-only; no source change.

Verification

• 3 of the 5 tests FAIL on pre-defence code (verified by neutralising the if psk is not None: verify gate to trust-every-record); all 5 pass on main.
• strands_robots/mesh/audit.py coverage 85% -> 86%; the previously-uncovered PSK-verify trust block is now exercised.
• Full tests/mesh/ suite: 1566 passed, 2 skipped.
• ruff check / ruff format --check / mypy all clean.

[yinsong1986]

Summary

Test-only PR adding tests/mesh/test_audit_seq_psk_verified_seed.py (+155/-0), which pins the PSK-verified branch of the audit-log seq-seed fallback in strands_robots/mesh/audit.py::_load_seq_counters. The five tests exercise the circular-trust defence on the seq-seed trust boundary: a correctly-signed record seeds the floor and _next_seq advances past it; unsigned records, PSK_DEGRADED/SIGN_FAILED poison sentinels, and wrong-HMAC records are skipped; an over-cap-but-signed record is refused by the shared _MAX_SEED_SEQ cap; and an all-unverified log seeds nothing (fail-closed). No source change.

The assertions line up with the actual module behaviour: _sign uses _canonical_bytes (which excludes sig) exactly as the writer does, the over-cap test's forged log-string and the all-unverified test's unverified/NOT seeded strings match the warnings at audit.py:607 and audit.py:627, and the empty-tmpdir fixture forces the no-sidecar degraded path that triggers the walk.

What's good

• Pins a previously-uncovered, reviewed security defence with a regression test (AGENTS.md > PR-6 "Pin regression tests for reviewed fixes"); the description confirms 3/5 fail on neutralised pre-defence code.
• Uses monkeypatch.setenv and tmp_path with full fixture teardown rather than mutating os.environ (AGENTS.md > PR #86 testing patterns).
• No host paths, no non-ASCII / emoji in user-facing or test strings, scope-disciplined (test-only, zero source diff).