gps-denied-onboard/_docs/03_implementation/batch_78_report.md

Batch 78 Report — FT-P-01 vertical slice (cycle 1, batch 12 of test phase)

Batch: 78 Date: 2026-05-17 Context: Test implementation (greenfield Step 10 — Implement Tests) Tasks: AZ-598 (5 cp) — 1 task (FT-P-01 vertical slice) Cycle: 1 Verdict: COMPLETE — PASS (self-reviewed + cumulative-reviewed; see reviews/batch_78_review.md + reviews/cumulative_76_78_review.md)

Summary

Two distinct concerns shipped under one ticket because they unblock each other:

1. Observer extension — sitl_observer._FdrReplayObserver.wait_for_outbound was missing despite being called by FT-P-01 + FT-P-05. The b78 audit caught this; the implementation adds cursor-based replay from outbound_messages_<fc_kind>_<host>.json plus an OutboundMessage dataclass + the two scenario kwarg fixes (fc_adapter= → fc_kind=).
2. FT-P-01 fixture builder — a vertical-slice tool that produces the two fixture files (outbound_messages_* + observer_*) for the FT-P-01 scenario. Pivoted from the original "live SITL docker capture" design (would have needed ~13+ cp of new SUT-side frame-ingestion code) to a "drive gps-denied-replay against a 1 fps MP4 + synthetic stationary tlog" approach that reuses the existing production ReplayInputAdapter. No new SUT code; one subprocess call instead of a multi-container compose.

Direction-correction surfaced mid-batch

During b78 scoping I told the user incorrectly that the "upload-tlog+video" feature wasn't implemented. Discovery during scope analysis showed src/gps_denied_onboard/replay_input/ exists exactly for that use case (CLI = gps-denied-replay, coordinator = ReplayInputAdapter, auto-sync = AZ-405). I corrected the user immediately, surfaced the direction options, and the user chose to stay the course on b78 (FT-P-01 vertical slice). The discovery also enabled the pivot from live-SITL-capture to "reuse the gps-denied-replay CLI" — turning the impossible-in-one-batch phase 2 into a tractable one.

AZ-598 — observer extension + FT-P-01 builder (5 cp)

Phase 1 — observer extension

• e2e/runner/helpers/sitl_observer.py (extended):
  • New OutboundMessage(lat_deg, lon_deg, image_id=None) frozen dataclass.
  • _FdrReplayObserver unfrozen (cursor state is now meaningful); _outbound_cursor: int = 0 + lazy _outbound_messages cache.
  • New wait_for_outbound(timeout_s: float | None = None) method with three outcomes: OutboundMessage / TimeoutError / RuntimeError.
  • New module-level _load_outbound_messages(fc_kind, host) helper that validates the entire messages list at first read.
• e2e/tests/positive/test_ft_p_01_still_image_accuracy.py: get_observer(fc_adapter=...) → get_observer(fc_kind=...).
• e2e/tests/positive/test_ft_p_05_sat_anchor.py: same kwarg fix.
• e2e/_unit_tests/helpers/test_sitl_observer.py: +11 tests covering cursor advance, timeout, exhaustion, missing file, missing env, malformed schema (list/object/keys), optional image_id, and cursor independence between observers.

Phase 2 — FT-P-01 fixture builder

• e2e/fixtures/sitl_replay_builder/__init__.py (new): minimal package docstring; deliberately no symbol re-exports (avoid the build_p01_fixtures function/submodule name-shadow pitfall — documented in the docstring).
• e2e/fixtures/sitl_replay_builder/build_p01_fixtures.py (new):
  • BuilderConfig frozen dataclass.
  • encode_stills_to_mp4(image_paths, output, fps=1.0) — OpenCV; accepts _video_writer_factory / _imread for testability.
  • generate_stationary_tlog(output, duration_s=120, hz=200) — pymavlink; writes zero-motion RAW_IMU + ATTITUDE pairs.
  • run_gps_denied_replay(video, tlog, fdr_out, time_offset_ms=0, extra_args=...) — subprocess to the production CLI; bypasses auto-sync because the synthetic tlog has no take-off.
  • parse_fdr_for_outbound_estimates(fdr_path, fdr_kind=..., lat_key=..., lon_key=...) — JSONL walk; configurable record-kind + field-key projection.
  • write_outbound_messages_fixture(output, image_ids, estimates) — schema writer; preserves None → JSON null for timeouts.
  • write_observer_fixture(output) — minimal observer config so get_observer succeeds.
  • build_p01_fixtures(cfg, *, _runner=None, ...) — orchestrator.
  • _main(argv=None) -> int — argparse CLI entry point.
• e2e/fixtures/sitl_replay_builder/README.md (new): strategy, usage, output structure, limitations.
• e2e/_unit_tests/fixtures/test_sitl_replay_builder.py (new): +24 tests — 3 for encode_stills_to_mp4, 4 for generate_stationary_tlog (incl. one real-pymavlink round-trip), 3 for run_gps_denied_replay, 6 for parse_fdr_for_outbound_estimates, 3 for write_outbound_messages_fixture, 1 for write_observer_fixture, 4 for end-to-end orchestration.
• e2e/_unit_tests/test_directory_layout.py: registers the three new files in the layout invariant.

Out of scope (deferred)

• Live capture EXECUTION — the builder runs gps-denied-replay as a subprocess; that subprocess requires pip install -e . at repo root plus access to the input images. Not executed in this batch; documented as a manual operator step. A future ticket can add a CI job that runs the live capture + commits the resulting fixtures.
• Other scenarios (FT-P-02 through FT-N-04) — each needs its own fixture-builder flow (continuous video + IMU CSV replay, blackout/spoof setup, etc.).
• iNav adapter — only ArduPilot supported in this batch.
• fc_kind ↔ fc_adapter naming convergence — kwarg-fix only; a future cleanup ticket should converge the vocabulary.

Test Results

• New unit tests: 35 (11 wait_for_outbound + 24 builder).
• Full e2e/_unit_tests suite: 664 passed in 137 s (previous cumulative: 637 → +27 net).
• No new linter errors.
• grep raise NotImplementedError under e2e/tests/ returns zero matches (b77 invariant preserved).

State

• Spec moved: _docs/02_tasks/todo/AZ-598_ft_p_01_vertical_slice.md → _docs/02_tasks/done/.
• _docs/_autodev_state.md advanced to last_completed_batch: 78, last_cumulative_review: batches_76-78 (K=3 cumulative shipped alongside the batch review).