[AZ-595] Batch 75: sitl_observer FDR-replay + scenario probe cleanup

Implement all 11 `sitl_observer` public surfaces as an offline
FDR-replay strategy (reads JSON fixtures under `${E2E_SITL_REPLAY_DIR}`
instead of live pymavlink/yamspy). Replace 12 per-scenario
`_harness_helpers_implemented` probes with one shared session-scoped
`sitl_replay_ready` fixture in `e2e/tests/conftest.py`.

Net: -636 LoC of duplicated scenario gating, +17 LoC shared fixture,
+38 new unit tests (596 total, up from 558). Includes K=3 cumulative
review for batches 73-75 (PASS).

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

_docs/03_implementation/reviews/cumulative_73_75_review.md

@@ -0,0 +1,171 @@
+# Cumulative Code Review Report
+
+**Batches**: 73, 74, 75 (AZ-424, AZ-425, AZ-426, AZ-594, AZ-595)
+**Date**: 2026-05-17
+**Verdict**: PASS
+
+## Scope
+
+Three consecutive implementation batches that together close the
+"negative-scenario evaluators + core harness stubs + offline observer
+strategy" arc:
+
+* **Batch 73 (AZ-424 / AZ-425 / AZ-426)** — added the three remaining
+  negative-scenario evaluators (outlier tolerance, outage request,
+  blackout-spoof) and their FT-N-01 / FT-N-03 / FT-N-04 scenarios.
+* **Batch 74 (AZ-594)** — turned the three core harness stubs
+  (`fdr_reader.iter_records`, `frame_source_replay.replay_video`,
+  `imu_replay.ImuReplayer.replay`) from `NotImplementedError` into
+  real implementations.
+* **Batch 75 (AZ-595)** — implemented all 11 `sitl_observer` public
+  surfaces as an offline FDR-replay observer reading from
+  `${E2E_SITL_REPLAY_DIR}`; collapsed 12 per-scenario
+  `_harness_helpers_implemented` probes into one shared
+  `sitl_replay_ready` fixture.
+
+## Cross-Batch Findings
+
+(none)
+
+## Cross-Batch Sweep
+
+### Spec Compliance
+
+Every AC across the five tickets has at least one unit test plus the
+intended scenario-side wiring. Coverage matrices are reproduced in the
+per-batch review files (`batch_73_review.md` § Phase 2,
+`batch_74_review.md` § Phase 2, `batch_75_review.md` § Phase 2). Spot
+checks across the three batches:
+
+* AC numbering is consistent within each ticket and traced via
+  `@pytest.mark.traces_to(...)` on every scenario.
+* The pure-logic ACs (evaluator math, schema parsing, FDR record
+  enumeration) are unit-tested at module level; the AC-5 / AC-6 "full
+  scenario passes under the parametrize matrix" rows are gated on the
+  shared `sitl_replay_ready` fixture and will activate the moment the
+  fixture builder lands.
+
+### Code Quality (cross-batch)
+
+* **Single responsibility holds across the three batches**: each new
+  evaluator owns exactly the AC math it claims (outlier_tolerance,
+  outage_request, blackout_spoof). Each new helper body owns exactly
+  one stream-of-records surface (fdr_reader, frame_source_replay,
+  imu_replay). The offline `sitl_observer` owns env-var resolution +
+  fixture-JSON ingestion, with one helper per fixture filename.
+* **No suppressed errors across any batch**: all parse paths raise
+  `ValueError` with a file pointer; missing inputs raise
+  `FileNotFoundError`; the `_load_optional_*` family in
+  `sitl_observer` falls back to `[]` ONLY when the file is genuinely
+  absent — a present-but-malformed file still raises.
+* **AAA comment discipline holds across the three batches**: 61 (b73)
+  + 34 (b74) + 38 (b75) = **133 new unit tests**, each tagged with
+  `# Arrange / # Act / # Assert`. No vague stubs.
+* **No code comments narrate code**; module docstrings explain
+  non-obvious design choices (wire-vs-runner schema rename, OpenCV
+  realtime/non-realtime distinction, offline replay rationale).
+* **Public boundary holds**: every new module imports only stdlib +
+  `cv2` (frame_source_replay) + pre-existing internal helpers. Zero
+  `from gps_denied_onboard ...` imports across the three batches.
+
+### Security (cross-batch)
+
+* **No new secrets, credentials, or network surfaces introduced** in
+  any of the three batches. Batch 75 in particular removes the
+  implicit requirement for a live SITL container in scenario unit
+  runs — the runner now reads from on-disk JSON fixtures via a
+  read-only env var.
+* **No `eval`, `exec`, `pickle`, `subprocess`, or
+  `yaml.load(unsafe=True)`** in any new module.
+* The wire-schema gate in `fdr_reader._parse_envelope` (b74) is the
+  cross-batch safety invariant — any SUT schema drift surfaces at
+  parse time, never as silent default-zero records.
+
+### Performance (cross-batch)
+
+* All new evaluators / helpers are O(N) in the bounded inputs they
+  consume (per-window frame count, per-fixture event count, per-CSV
+  row count). The only O(N log N) step is `fdr_reader.iter_records`'s
+  multi-file merge-sort, deliberately accepted to give scenarios a
+  monotonic-time guarantee.
+* No I/O at module-import time across all three batches.
+* `_FdrReplayObserver` (b75) caches the parsed payload behind the
+  `get_observer` factory so repeat scenario calls for the same
+  (fc_kind, host) do not re-parse JSON.
+
+### Cross-Task Consistency
+
+* **The three batches are tightly coupled by design**:
+  - b73 scenarios introduced `_harness_helpers_implemented` probes
+    that gated on `NotImplementedError` from `fdr_reader`,
+    `frame_source_replay`, `imu_replay`, and `sitl_observer`.
+  - b74 landed three of those four helpers — the probes started
+    catching `FileNotFoundError` instead of `NotImplementedError`,
+    and (per the outer `except Exception: return False`) continued
+    to skip cleanly. No batch-73 scenario broke.
+  - b75 landed the fourth helper, then deleted every
+    `_harness_helpers_implemented` probe in favour of the shared
+    `sitl_replay_ready` fixture. The skip path is now keyed on a
+    single env var (`E2E_SITL_REPLAY_DIR`) rather than four
+    `try / except NotImplementedError` probes.
+* **Skip semantics are now uniform**: every scenario that depends on
+  the FDR-replay path emits a skip message of the form
+  `"FT-X-Y full scenario requires `E2E_SITL_REPLAY_DIR` to point at a prepared SITL replay fixture (AZ-595). Pure-logic ACs covered by <unit-test>."`
+  The grep-by-scenario inventory in `batch_75_review.md` Phase 6
+  enumerates the 12 scenarios.
+* **Dataclass field names hold across batches**: the b75
+  `EkfDivergenceEvent`, `GpsHealthSample`, `ConsistencyCheckEvent`,
+  `FcGpsState`, `MspFrameCapture`, `InavGpsState` field names match
+  exactly what the b73 evaluators (`outlier_tolerance_evaluator`,
+  `outage_request_evaluator`, `blackout_spoof_evaluator`) and b72
+  consumers (`ap_contract_evaluator`, `msp_frame_observer`) already
+  reference. No consumer required edits in b75.
+* **`FileNotFoundError` is the cross-batch convention** for missing
+  on-disk inputs — `accuracy_evaluator`, `multi_segment_evaluator`,
+  `mavproxy_tlog_reader`, `cold_start_evaluator` (pre-existing),
+  `fdr_reader` / `frame_source_replay` / `imu_replay` (b74), and
+  `sitl_observer._load_required_json` (b75) all agree.
+* **`sleep_fn` injection pattern** introduced in b74 for
+  `FrameSourceReplayer` + `ImuReplayer` is also the pattern used by
+  pre-existing helpers `tile_cache_builder`, `age_injector`. b75 did
+  not introduce any new sleep paths.
+
+### Architecture Compliance
+
+* **Module placement unchanged** across all three batches. Every new
+  helper / evaluator lives at `e2e/runner/helpers/<name>.py`; every
+  new unit test lives at `e2e/_unit_tests/helpers/test_<name>.py`.
+  Directory-layout invariant test (`test_directory_layout.py`) passes
+  unmodified across b73→b75 (the b73 additions registered the three
+  new evaluators; b74/b75 only edited existing registered files).
+* **No `src/gps_denied_onboard` imports** introduced in any of the
+  three batches. Confirmed by `grep`.
+* **Backwards-compatible scenario contract**: every public surface
+  that scenarios called pre-b73 retains the same name + return type
+  through b75. The only deletions are the local
+  `_harness_helpers_implemented` / `_NullSink` / `_NullImuEmitter`
+  helpers that lived inside the scenario files themselves — never
+  part of the public helper API.
+
+## Test Results (cumulative)
+
+| Batch | New unit tests | Cumulative pass count | Net delta |
+|-------|----------------|------------------------|-----------|
+| 72 (baseline) | — | 460 | — |
+| 73 | 61 (14 outlier + 18 outage + 29 blackout-spoof) | 527 | +67 |
+| 74 | 34 (14 fdr_reader + 10 frame_source_replay + 10 imu_replay) | 558 | +31 |
+| 75 | 38 (sitl_observer end-to-end) | **596** | +38 |
+
+* Full `e2e/_unit_tests` suite: **596 passed in 123 s** at end of b75.
+* No new linter errors at any batch boundary.
+* The pre-existing collection-time `/e2e-results/evidence` teardown
+  warning persists when scenarios are collected outside docker — it is
+  a known b67 artefact, not caused by any of these three batches.
+
+## Cumulative Verdict
+
+PASS. The three batches together (a) close the negative-scenario
+evaluator coverage, (b) land the three core harness stubs that
+unblocked every NotImplementedError-gated scenario, and (c) unify the
+scenario skip pattern behind one env var. No cross-batch
+inconsistencies, no architectural drift, no security regressions.