test(golden): capture/replay golden-fixture harness (#41)

[millsmillsymills]

What

A golden-fixture test harness that captures real Flipper CLI/RPC byte exchanges once and replays them offline in CI, mirroring the workspace VCR-cassette discipline.

• tests/golden/harness.py — RecordingTransport (wraps a real transport, logs every wire event), ReplayTransport (serves a fixture back offline), and the Fixture JSON format.
• tests/golden/record.py — captures fixtures from a USB-connected Flipper.
• tests/golden/test_golden_fixtures.py — replays each fixture in the default not integration suite.
• tests/golden/fixtures/*.json — six committed fixtures (four captured from real hardware, two no-wire-I/O refusals).
• Makefile — record-fixtures (hardware) and verify-fixtures (offline drift check).
• docs/agents/golden-fixtures.md — harness reference.

Replay model

Event-cursor: receive() serves the next rx event only once every tx recorded before it has been sent; otherwise it returns b"". This gates device output behind the matching request, so the time-based drain/read loops in the RPC and CLI code reproduce the recorded exchange faithfully, with no dependence on wall-clock timing or hardware. Sent payloads are recorded for assertions but not matched, so the push-integrity mismatch test can replay the same device MD5 against a corrupted local file.

Coverage (the six launch-gate behaviors)

cli_exec echo/prompt parsing, fs_push MD5 match + mismatch, RPC↔CLI mode switch, shared _io_lock non-interleave, the TX two-gate (enabled + disabled), and WiFi-CLI rejection.

Drift guard

make verify-fixtures (also in the default CI suite) replays every fixture and asserts the committed expect. A parsing/framing/gating change that reinterprets the real captured bytes fails the matching test — signal to re-record on hardware. Fixtures carry a schema_version; a mismatch is a hard load error.

Verification

• Fixtures captured from a real USB Flipper; full make record-fixtures (capture → verify) loop green.
• uv run pytest -m "not integration" — 144 passed, coverage 91.73%.
• ruff check, ruff format --check, ty check src/ all clean.

Closes #41

🤖 Generated with Claude Code