SSH authentication transport host regression

Run from the repository root:

CCACHE_DISABLE=1 python3 tests/ssh_auth_transport/run.py

Requires Python 3 and a C11 host compiler (cc, or CC override). The runner also sets CCACHE_DISABLE=1 for its subprocesses. Compilation uses -Wall -Wextra -Werror; compilation and execution have timeouts. Generated C and executable live in a temporary directory, not the source tree. Assertion or compiler failures propagate as a nonzero exit.

Production code under test

Following tests/ssh_management/runtime.py, run.py extracts named function definitions verbatim from src/ssh_transport.c. It extracts the actual slot definition, transport constants, session/route enums and counters declaration too, and links the actual src/ssh_auth_policy.c with its header. Missing/ambiguous function definitions fail extraction rather than silently falling back to fixture implementations.

The extracted set covers authentication callbacks and helpers, counter saturation/clear, socket acceptance and free-slot selection, start/stop, retirement, admin RX, serial RX and shared TX flushing. The fixture supplies only synchronous boundary doubles for clock, database, socket/library allocation and I/O, broker, console, publication and context/listener construction. No credential verifier, admission algorithm or callback logic is reimplemented. boot() resets test state between independent scenarios; the restart tests use actual production start/stop without resetting the policy.

Coverage

Password success, invalid password, backend failure and rejected password change; admissions versus completed attempts, principal promotion and pending-principal clearing.
Signed key authorization denial/backend error, bad signature result, stale principal, currentness backend failure and success. Admission precedes authorization, authorized signed keys await completion, and duplicate/unexpected callbacks cannot create a promotion or count again. Verification-throttled signed requests do not reach authorization or acquire a pending signature-result marker.
Close/state/pending-result fences, unsupported method handling, keyboard callback rejection with the entire prompt structure cleared and no repeated method count after close.
Unsigned authorized/rejected probes share a 12-token pool across both slots/reconnects, refill one token at five seconds and never become completed attempts. Password/signed verification shares six tokens, refilling one per ten seconds. Three completed failed password or signed-key attempts close the slot.
Actual acceptance loop: two-slot capacity, six handshake admissions, allocation failures consume admission, throttle denial before allocation, and one-token/ten-second refill.
Actual stop/start and counter clear preserve all three exhausted pools. Counter clear initialization guard, 64-bit saturating addition and per-slot 8-bit saturation.
Partial admin RX consumption (including false with consumed bytes), zero-consumption rejection and full drain wipe only consumed spans. WANT_READ, WANT_WRITE, rekey, window/channel retry, zero and error paths preserve pending TX; positive partial sends wipe only accepted admin spans. Binary serial RX/TX remains unchanged, including embedded NUL/0xff and partial broker acceptance.
Whole-slot retirement wipe including padding and buffers, exact retained generation, socket_fd == -1 sentinel, broker-disconnect failure retaining a closing slot, and repeated cleanup without double free/close. The old live fd is closed, not retained.
Source checks for context registration of authentication, type advertisement, result and rejecting keyboard callbacks, plus per-session auth/result/keyboard contexts.

Limits and deferred validation

The wolfSSH types in fixture.h are narrow host shapes, not ABI validation. Library return codes model distinct outcomes; this does not compile wolfSSH or perform cryptographic signature work. Denied authorization is verified to return a non-success result and leave no pending completion; proving the parser actually skips signature processing belongs to the separately delegated real wolfSSH parser-order regression. Do not treat these tests as its replacement.

Context construction is stubbed; callback registration is checked in production source rather than executed. Console/broker/database behavior and real socket scheduling are outside this suite. A fixture principal uses synthetic bytes to inspect lifecycle clearing; there are no real credentials.

No firmware build, device execution, upload or erase is performed. All device validation remains deferred to whole Phase 9, including real clients, timing, concurrency, recovery and physical UART/broker behavior.