[AZ-599] Batch 79: FT-P-02 Derkachi builder + _common.py extraction

- Add build_p02_fixtures.py: IMU CSV → tlog conversion (RAW_IMU +
  ATTITUDE pairs, centidegrees→radians yaw) and orchestrator that
  runs gps-denied replay against Derkachi MP4 + generated tlog,
  verifying ≥1 record_type="estimate" in the FDR archive.
- Extract run_gps_denied_replay + FDR-parent-dir helpers into
  sitl_replay_builder/_common.py; refactor build_p01_fixtures.py
  to import from _common (b78 tests preserved).
- Add 20 unit tests under e2e/_unit_tests/fixtures/test_sitl_
  replay_builder_p02.py covering AC-1..AC-5; total unit suite
  686/686 passing (regression gate AC-6).
- README updated to document FT-P-01 + FT-P-02 builders.
- Advance autodev state: last_completed_batch=79, current_batch=80;
  prune verbose detail blob.

Co-authored-by: Cursor <cursoragent@cursor.com>

_docs/03_implementation/reviews/batch_79_review.md

@@ -0,0 +1,156 @@
+# Code Review Report
+
+**Batch**: 79 — AZ-599 (FT-P-02 Derkachi fixture builder + b78 helper extraction)
+**Date**: 2026-05-17
+**Verdict**: PASS
+
+## Findings
+
+(none blocking)
+
+### Non-blocking notes
+
+* **Roll/pitch=0 simplification in synthesised ATTITUDE**: acceptable
+  for the Derkachi fixed-wing cruise data but unrealistic for
+  aggressive manoeuvres. Documented as a Limitation in the README;
+  re-visit if FT-P-04 (Derkachi F2F registration) or other
+  manoeuvre-sensitive scenarios show fusion drift from this gap.
+* **`SCALED_IMU2` → `RAW_IMU` pass-through (no unit conversion)**:
+  the builder passes scaled accelerometer (mg) and gyro (mrad/s)
+  values straight into RAW_IMU fields. Documented as a Limitation;
+  if the SUT's tlog parser rejects, a units conversion pass needs to
+  be added. Surfaced as a follow-up rather than fixed speculatively.
+* **Disk-full incident during regression gate**: not a code issue;
+  surfaced to the user, who freed space, after which the full
+  suite (686 tests) passed cleanly.
+
+## Findings Sweep
+
+### Phase 1 — Context Loading
+
+Read the FT-P-02 scenario to confirm: it (a) skips when
+`sitl_replay_ready` is false; (b) calls
+`fdr_reader.iter_records(fdr_root)` to walk the FDR archive; (c)
+filters for `record_type == "estimate"` and projects into
+`apd.FdrEstimate`. Confirmed FT-P-02 does NOT call
+`wait_for_outbound` (so the b78 schema doesn't apply). Read
+`data_imu.csv` to confirm column names + units (10 Hz; ms
+timestamps; SCALED_IMU2 in mg / mrad/s; GLOBAL_POSITION_INT.hdg
+in centidegrees). Inspected the production
+`replay_input/auto_sync.py` AC-13 contract for required tlog
+message types (RAW_IMU + ATTITUDE) — this drove the choice of
+which messages to pack per CSV row. Verified `pymavlink>=2.4`
+already in `pyproject.toml`. Read the b78
+`build_p01_fixtures.py` to identify the helper extraction
+boundary (`run_gps_denied_replay` + `write_observer_fixture` had
+no FT-P-01-specific code, ideal candidates).
+
+### Phase 2 — Spec Compliance
+
+| AC | Coverage | Status |
+|----|----------|--------|
+| AC-1 (parse every row, one pair per row) | `test_convert_imu_csv_writes_pair_per_row`, `test_convert_imu_csv_real_pymavlink_round_trip` | Covered |
+| AC-2 (ValueError on missing cols / malformed numerics / empty CSV) | `test_convert_imu_csv_empty_raises`, `test_convert_imu_csv_missing_required_column_raises`, `test_convert_imu_csv_malformed_numeric_raises`, `test_convert_imu_csv_missing_file_raises` | Covered (4 distinct error paths) |
+| AC-3 (heading centideg → rad) | `test_hdg_centideg_to_rad` (5 parametric cases: 0, π/2, π, 3π/2, near-2π) | Covered |
+| AC-4 (build_p02 invokes replay + verifies ≥1 estimate) | `test_build_p02_end_to_end_with_mocks`, `test_build_p02_propagates_verify_failure`, `test_verify_fdr_no_estimates_raises`, `test_verify_fdr_counts_estimates`, `test_verify_fdr_tolerates_malformed_lines` | Covered |
+| AC-5 (`_common.py` shared between b78 + b79) | `test_common_module_exports_used_by_b01` + b78 suite (24/24) still passes | Covered |
+| AC-6 (full unit-test suite passes) | 686/686 in 123 s (previous: 664) | Covered |
+
+### Phase 3 — Code Quality
+
+* **Single responsibility**: `convert_imu_csv_to_tlog` parses CSV +
+  packs MAVLink (one cohesive concern: CSV row → tlog pair).
+  `verify_fdr_has_estimates` does one thing. `build_p02_fixtures`
+  is the only orchestrator. Helper extraction into `_common.py`
+  removed duplicate definitions cleanly — no copy-pasted code
+  remains.
+* **No suppressed errors**: every JSON-decode in
+  `verify_fdr_has_estimates` is wrapped in a `try / except
+  json.JSONDecodeError` that explicitly `continue`s with a comment
+  explaining the tolerance (real FDR archives can contain partial
+  writes at the tail). Every CSV parse error wraps in `ValueError`
+  with row context. `subprocess.run` uses `check=True`.
+* **AAA test discipline**: all 20 new tests use
+  `# Arrange / # Act / # Assert`; omitted when redundant.
+* **Comments**: every public function has a docstring documenting
+  contract + raises; no narrating comments in function bodies.
+  The yaw conversion uses a named helper (`_hdg_centideg_to_rad`)
+  rather than an inline magic-formula comment.
+* **Public boundary**: b79 module does NOT import any
+  `src/gps_denied_onboard` symbol (verified by grep). pymavlink
+  imports are deferred to function bodies so the test factory
+  injections don't need the real library available.
+
+### Phase 4 — Security
+
+* No new credentials / secrets / network surfaces.
+* `convert_imu_csv_to_tlog` uses `csv.DictReader` — no `eval`, no
+  shell injection. Numeric fields go through `int(float(s))` so
+  bad input raises ValueError rather than executing arbitrary code.
+* `verify_fdr_has_estimates` uses `json.loads` (safe) and silently
+  ignores malformed lines (documented as tolerance for partial
+  writes; not a security risk because we never act on the data).
+* `subprocess` invocation uses a list-argument `cmd` (no shell
+  injection surface; same pattern as b78).
+
+### Phase 5 — Performance
+
+* `convert_imu_csv_to_tlog` is O(N) over CSV rows; for the Derkachi
+  4,900-row file → ~5 s wall-clock estimate (pymavlink packing is
+  the bottleneck). Bounded by input size; no hidden quadratic
+  behaviour.
+* `verify_fdr_has_estimates` is single-pass O(N) over JSONL lines.
+* `_iter_imu_rows` validates the header once and then streams rows;
+  doesn't materialize the entire CSV in memory until the orchestrator
+  asks for `list(...)` (could be optimized later if Derkachi data
+  ever grows orders of magnitude larger; currently 4,900 rows is
+  trivial).
+
+### Phase 6 — Cross-Task Consistency
+
+* **Builder shape parity with b78**: `BuilderConfig` /
+  `P02BuilderConfig` dataclass; `build_pXX_fixtures(cfg, *, _runner=...,
+  _mavlink_writer_factory=..., _verify_fdr=...)` signature; same
+  underscore-prefixed dependency-injection convention. A future
+  contributor reading both side-by-side will see the same pattern.
+* **`_common.py` extraction validated by both**: b78's existing
+  `test_run_gps_denied_replay_builds_correct_cmd` /
+  `test_run_gps_denied_replay_creates_fdr_parent_dir` /
+  `test_run_gps_denied_replay_passes_extra_args` /
+  `test_write_observer_fixture_schema` tests pass unchanged against
+  the moved implementation. b79 adds
+  `test_common_module_exports_used_by_b01` to lock in the import
+  relationship.
+* **No drift in `observer_<fc_kind>_<host>.json` schema**: same
+  payload structure as b75/b78.
+
+### Phase 7 — Architecture Compliance
+
+* **Module placement**: new files under
+  `e2e/fixtures/sitl_replay_builder/` (existing package from b78);
+  new tests under `e2e/_unit_tests/fixtures/`. All registered in
+  `test_directory_layout.py`.
+* **No `src/gps_denied_onboard` imports**.
+* **No new top-level dependencies** — pymavlink already there from
+  b78/b76; `csv` is stdlib; `math` is stdlib.
+* **`__init__.py` not modified** — the b78 docstring already
+  documents why we don't re-export symbols on the package namespace,
+  and that's still the right choice for b79.
+* **Refactor preserves backwards compatibility**: `build_p01_fixtures`
+  re-imports `run_gps_denied_replay` + `write_observer_fixture`
+  from `_common.py`. Any caller (test or production) doing
+  `bp01.run_gps_denied_replay(...)` still works because the symbol
+  is present in the module namespace via the `from … import …`
+  re-export.
+
+## Test Results
+
+* New unit tests: **20** (5 `convert_imu_csv_to_tlog` + 5
+  `_hdg_centideg_to_rad` parametric + 4 `verify_fdr_has_estimates`
+  + 4 `build_p02_fixtures` end-to-end + 1 export-relationship
+  + 1 missing-file negative).
+* Full `e2e/_unit_tests` suite: **686 passed in 123 s** (previous:
+  664 → +22 net = +20 new + +2 directory-layout entries).
+* No new linter errors (`ReadLints` clean on all touched files).
+* No regression in the b78 builder tests (24/24 still pass after
+  the helper extraction).