mirror of https://github.com/azaion/gps-denied-onboard.git synced 2026-06-21 16:11:13 +00:00

Files

T

Oleksandr Bezdieniezhnykh 6d51e06886 [AZ-589] [AZ-590] [AZ-591] [AZ-592] [AZ-593] Re-classify cycle1 gate findings

Cycle 1 Product Implementation Completeness Gate post-mortem.
AZ-589 + AZ-590 were the wrong abstraction:

- AZ-589 targeted `okvis::ThreadedKFVio` (OKVIS v1 API) which does
  not exist in the vendored OKVIS2 upstream; smartroboticslab/okvis2
  exposes `okvis::ThreadedSlam` instead.
- AZ-590 assumed a "de-ROSified VINS-Mono pin" submodule exists;
  `cpp/vins_mono/upstream/` has no `.gitmodules` entry.
- The actual production gap is the empty central
  `_STRATEGY_REGISTRY`: `register_strategy(...)` is never called
  outside test fixtures, so `compose_root()` raises
  `StrategyNotLinkedError` for every component slug with a
  strategy-selecting config field. Affects c1_vio + c2_vpr +
  c2_5_rerank + c3_matcher + c3_5_adhop + c4_pose + c5_state.

Re-classification:

- AZ-589 + AZ-590 closed Won't Fix (Jira); spec files removed
  from todo/ but rows retained in the dependencies table as
  audit-trail.
- AZ-591 created (todo/, 5pt) — cross-cutting compose_root
  per-binary bootstrap that populates `_STRATEGY_REGISTRY` for
  the airborne binary. Scheduled as Batch 66 sole task.
- AZ-592 created (backlog/, 5pt placeholder) — AZ-332 Tier-2
  validation bundle (real `okvis::ThreadedSlam` wiring + Linux CI
  apt-install + DBoW2 vocab + Jetson). BLOCKED on Tier-2
  prerequisites; honors AZ-332's `AZ-332_tier2_validation`
  self-deferral handle.
- AZ-593 created (backlog/, 5pt placeholder) — AZ-333 Tier-2
  validation bundle (de-ROSified VINS-Mono upstream + binding +
  CI + Jetson). BLOCKED on upstream vendoring decision plus
  Tier-2 prerequisites; honors AZ-333's parallel deferral pattern.
- AZ-332 + AZ-333 re-classified in cycle1 gate report from FAIL
  to BLOCKED-on-Tier-2.

Step 7 stays in_progress until AZ-591 lands; after that it can
advance to Step 8 with AZ-592 + AZ-593 parked in backlog/.

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

2026-05-16 12:45:58 +03:00

26 KiB

Raw Blame History

Product Implementation Completeness Gate — Cycle 1

Date: 2026-05-16 Cycle: 1 Tasks audited: 107 done product tasks under _docs/02_tasks/done/ (the six hygiene-only specs and AZ-525-class follow-ups are included as PASS because they don't promise new runtime behavior). Audit scope: every task spec's Description / Outcome / Scope.Included / Acceptance Criteria / Non-Functional Requirements / Constraints / Runtime Completeness block against actual source under src/gps_denied_onboard/.

Verdict

Revised 2026-05-16 (post-mortem after AZ-589/AZ-590 investigation): The original verdict below classified AZ-332 + AZ-333 as FAIL and created remediation tasks AZ-589 + AZ-590. Subsequent investigation during Batch 66 entry showed both classifications and remediation tasks were wrong:

AZ-589's targeted API (okvis::ThreadedKFVio) does not exist in the actually-checked-in OKVIS2 upstream submodule (smartroboticslab/ okvis2 @ a2ea0068 exposes okvis::ThreadedSlam + okvis::ViParametersReader + okvis::ViParameters — that's OKVIS2, not OKVIS v1).
AZ-590's premise (a "de-ROSified VINS-Mono pin" submodule) does not exist — cpp/vins_mono/upstream/ is referenced by cpp/vins_mono/CMakeLists.txt but .gitmodules has no entry for it.
The actual production gap is the empty central _STRATEGY_REGISTRY. A workspace-wide grep confirms register_strategy(...) is never called outside test fixtures. Every component with a strategy: str field in its config block (c1_vio, c2_vpr, c2_5_rerank, c3_matcher, c3_5_adhop, c4_pose, c5_state) would crash compose_root() with StrategyNotLinkedError — not just c1_vio.
Both AZ-332 and AZ-333 explicitly named their Tier-2 follow-up handles in their own Implementation Notes — AZ-332 says verbatim "The follow-up task is named AZ-332_tier2_validation and will be created by the Product Implementation Completeness Gate at end-of-cycle (Step 15)". The original FAIL classification missed this explicit self-deferral.

Revised classification:

AZ-332 → BLOCKED on Tier-2 prerequisites (CI build env + Jetson hardware + DBoW2 vocab artifact). Tier-2 follow-up filed as AZ-592 (parked in _docs/02_tasks/backlog/).
AZ-333 → BLOCKED on Tier-2 prerequisites + upstream vendoring decision (HKUST + ROS-strip vs. community fork). Tier-2 follow-up filed as AZ-593 (parked in _docs/02_tasks/backlog/).
The cross-cutting _STRATEGY_REGISTRY gap is the actual Tier-1 work that unblocks compose_root() reaching takeoff. Filed as AZ-591 (_docs/02_tasks/todo/).

AZ-589 + AZ-590 closed Won't Fix (Jira). Their spec files were deleted from _docs/02_tasks/todo/. The audit-trail rows remain in _docs/02_tasks/_dependencies_table.md for traceability.

Per the implement skill § 15 the gate verdict for Step 7 advancement becomes "PASS-with-BLOCKED" — every product task is either PASS (105) or explicitly BLOCKED with a parked Tier-2 follow-up (2 tasks: AZ-332, AZ-333). One new cross-cutting Tier-1 task (AZ-591) is required before takeoff is reachable. Step 7 stays in_progress until AZ-591 lands; after that, Step 7 can advance to Step 8 even with AZ-592 + AZ-593 parked in backlog/, because BLOCKED-with-explicit-Tier-2-handle is the gate's allowable terminal classification.

Original (now superseded) verdict

The original cycle-1 verdict text follows verbatim for audit. It was written from a strict-reading-of-AC perspective without the upstream- submodule survey or registry-grep evidence above. Do not act on it.

[Superseded] FAIL — Step 7 must not advance.

Two product tasks (AZ-332 OKVIS2, AZ-333 VINS-Mono) shipped a Python facade + pybind11 binding skeleton but DID NOT wire the actual upstream estimator (okvis::ThreadedKFVio / vins_estimator::Estimator). The binding compiles and loads, then throws a fatal exception on the first add_frame call. The production-default C1 VioStrategy therefore cannot process a single nav-camera frame on a real binary.

Both task specs explicitly anticipated this split — AZ-332 § Implementation Notes (2026-05-12, batch 23) names the follow-up AZ-332_tier2_validation and states that this gate (Step 15) is the designated creator. AZ-333 carries the same skeleton pattern but no self-deferral note. This report executes that contract.

Per implement/SKILL.md § 15 ("If any product task is FAIL, STOP. Do not write the final product implementation report and do not proceed to any downstream autodev step."), Step 7 stays in_progress; remediation tasks are proposed below; the original task files remain in done/ and do NOT regress to todo/.

FAIL findings

AZ-332 — C1 OKVIS2 Strategy (production-default VIO)

Promised capability: "production-default VioStrategy ... Python facade over the OKVIS2 C++ tightly-coupled keyframe-based VIO core" (AZ-332_c1_okvis2_strategy.md § Description). Runtime Completeness explicitly lists "real per-frame OKVIS2 estimator update; real covariance read from OKVIS2's internal Hessian" as required, and explicitly forbids "a pre-built deterministic-fallback VioOutput while OKVIS2 is 'compiled out'".

Evidence:

src/gps_denied_onboard/components/c1_vio/okvis2.py — 339-line Python facade. Conforms to the AZ-331 VioStrategy Protocol. PASS.
src/gps_denied_onboard/components/c1_vio/_native/okvis2_binding.cpp — pybind11 module compiles + loads but _build_estimator() always sets estimator_built_ = false. _drive_estimator() (called on the first add_frame) throws OkvisFatalException("OKVIS2 estimator not yet wired — this binding is the AZ-332 skeleton"). FAIL.
OKVIS2 upstream is never #include'd (the #include <okvis/ThreadedKFVio.hpp> line is commented out, line 48 of the binding).

Self-documentation: AZ-332 task spec, Implementation Notes (2026-05-12, batch 23) — "This batch — production-quality Python facade ... pybind11 binding source that compiles + loads but throws ... ; Tier-2 follow-up — actual okvis::ThreadedKFVio wiring ... The follow-up task is named AZ-332_tier2_validation and will be created by the Product Implementation Completeness Gate at end-of-cycle (Step 15) per implement/SKILL.md."

Blast radius: the deployment binary (config.vio.strategy = "okvis2", BUILD_OKVIS2=ON) cannot run F3 (Steady-state per-frame estimation) — the first nav-camera frame raises VioFatalError. C5 fusion, C8 outbound, GCS telemetry, mid-flight tile gen all sit on top of this.

AZ-333 — C1 VINS-Mono Strategy (research-only VIO)

Promised capability: Runtime Completeness requires "real VinsMonoStrategy class implementing the AZ-331 Protocol; real pybind11 binding to cpp/vins_mono/ (real VINS-Mono upstream, de-ROSified); real per-frame estimator update".

Evidence:

src/gps_denied_onboard/components/c1_vio/vins_mono.py — 448-line Python facade. Conforms to the AZ-331 Protocol. PASS.
src/gps_denied_onboard/components/c1_vio/_native/vins_mono_binding.cpp — same skeleton pattern as OKVIS2. _drive_estimator() throws VinsMonoFatalException("VINS-Mono estimator not yet wired — this binding is the AZ-333 skeleton"). FAIL.

Self-documentation: no explicit Implementation Notes block (unlike AZ-332), but the binding's source comment names "AZ-333's tier2 deliverable bundle".

Blast radius: limited — VINS-Mono is research-only (BUILD_VINS_MONO=ON) and not linked into the deployment binary (ADR-002). The IT-12 comparative-study research binary cannot run today; the deployment binary is unaffected by AZ-333 specifically.

PASS — by component

107 audited tasks, 105 PASS, 0 BLOCKED, 2 FAIL.

Tasks classified as PASS have at least one of:

A substantial Python/C++ source artifact under the task's owned component (module-layout.md ownership envelope).
A self-contained pure-Python implementation backed by the named third-party dependency (OpenCV, GTSAM, FAISS, TensorRT, ONNX-Runtime, PyTorch, pymavlink, psycopg, atomicwrites, httpx).
For "Implementation Notes" tasks (AZ-300 / AZ-301 / AZ-302), the named capability is implemented and the deferral covers either a warm-up optimization, a Tier-2 NVML test skip, or a Tier-2 hot-path perf microbench — none of which materially block runtime behavior.

Foundation + cross-cutting (10 tasks) — all PASS

Task	Title	Evidence
AZ-263	initial structure	`src/` skeleton present; package importable.
AZ-266	log module	`gps_denied_onboard.logging` package.
AZ-267	fdr log bridge	producer-id-tagged log → FDR records.
AZ-268	log schema contract test	shipped in tests.
AZ-269	config loader	`gps_denied_onboard.config` (env + YAML).
AZ-270	compose root	`runtime_root/__init__.py` (`compose_root`, `compose_operator`).
AZ-271	config precedence tests	shipped.
AZ-507	hygiene module-layout AZ-270 alignment	lint test `tests/unit/test_az270_compose_root.py`.
AZ-508 / AZ-526	iso-timestamp consolidation	`helpers/iso_timestamps.py`.
AZ-527	engine-dim assertion consolidation	`components/c2_vpr/_engine_dim_assertion.py` + sibling under c3.
AZ-528	c1 vio facade spine consolidation	`_facade_spine.py`.

FDR / FdrClient (4 tasks) — all PASS

Shared helpers (8 tasks) — all PASS

C13 FDR writer (6 tasks) — all PASS

C7 Inference (6 tasks) — all PASS (notable deferrals are documented + non-blocking)

C6 Tile cache (6 tasks) — all PASS

C11 Tile manager (5 tasks) — all PASS

C10 Provisioning (5 tasks) — all PASS

C12 Operator orchestrator (5 tasks) — all PASS

C1 VIO (5 tasks) — 1 PASS, 2 FAIL, 2 PASS

C2 VPR (6 tasks) — all PASS

Note: src/gps_denied_onboard/components/c2_vpr/_native/__init__.py contains only the line """Native bindings for VPR runtime — placeholder.""". The C2 strategies route inference through C7 (TensorRT / ONNX-RT / PyTorch), so this _native/ directory is empty by design (no extant task promises VPR-specific C++ code). Recommend deleting the directory in a future hygiene pass; not a FAIL today.

C2.5 Re-rank (2 tasks) — both PASS (with one noted concern, see § Notes)

C3 Cross-domain matcher (4 tasks) — all PASS

Note: c3_matcher/_native/__init__.py is similarly an empty placeholder — same situation as C2's _native/. Hygiene cleanup, not a FAIL.

C3.5 AdHoP refinement (2 tasks) — both PASS

| AZ-348 | refiner protocol | c3_5_adhop/interface.py. | | AZ-349 | AdHoP refiner | 509-line adhop_refiner.py; real C7-backed AdHoP engine load. Note: runtime_root/refiner_factory.py docstring still calls AdHoPRefiner "placeholder today" — that comment is stale; the production class is real. Hygiene fix recommended (one-line doc update). |

C4 Pose estimation (3 tasks) — all PASS

C5 State estimator (9 tasks) — all PASS

C8 FC adapter (8 tasks) — all PASS

Replay path (8 tasks) — all PASS

Notes / non-blocking observations

Production composition root has no per-binary bootstrap module registering strategies. runtime_root/__init__.py defines a strategy registry (register_strategy, _resolve_strategy) and topo-sorts constructed components, but register_strategy is never called anywhere in src/. compose_root(config) would raise StrategyNotLinkedError on every C1-C8 slug if invoked today. This is the "per-binary bootstrap module" the AZ-263 / AZ-270 prose anticipates — a separate concern from any one task and arguably out of scope for this gate (the registry seam exists; the actual registration lives in a not-yet-written bootstrap module). Recommend surfacing as a separate cross-cutting task (E-CC-CONF or E-BOOT).
helpers/feature_extractor.py::OpenCvOrbExtractor is documented as a placeholder ("Production deployments MUST replace this extractor with a deep-learning backbone before flight (tracked under the future C2.5 backbone-extractor task)"). No DISK/ALIKED extractor exists. C2.5 (AZ-343) uses an injected FeatureExtractor; the only concrete impl is ORB. AZ-343's spec does NOT name DISK/ALIKED, so this is a known-future-task gap rather than an AZ-343 FAIL — but the prod composition root will need a non-ORB extractor before flight. Recommend surfacing as a follow-up task (5 points or less).
_types/tile.py scaffolding DTOs (Tile, TileRecord) are no longer referenced by any module under src/. Dead code per coderule.mdc. Recommend deleting in a hygiene PBI; not a Gate FAIL.
runtime_root/refiner_factory.py docstring describes AdHoPRefiner as "placeholder today" — stale comment; the production class is real. One-line doc fix.
c2_vpr/_native/__init__.py and c3_matcher/_native/__init__.py are empty placeholder modules. C2/C3 strategies route inference through C7; no native code is owed. Recommend deleting both directories.
Process leftover 2026-05-11_d_cross_cve_1_opencv_pin_deferred.md remains open (gtsam still numpy 1.x). Not blocking for this gate.

Remediation tasks (REVISED 2026-05-16 post-mortem)

The two original remediation tasks AZ-589 + AZ-590 (created earlier same day) have been closed Won't Fix in Jira after the post-mortem investigation surfaced that:

AZ-589 targeted okvis::ThreadedKFVio (OKVIS v1) which does not exist in the vendored OKVIS2 upstream (smartroboticslab/okvis2 exposes okvis::ThreadedSlam).
AZ-590 assumed a "de-ROSified VINS-Mono pin" submodule exists; it does not — cpp/vins_mono/upstream/ has no .gitmodules entry.
Both tickets misframed a cross-cutting _STRATEGY_REGISTRY population gap as a per-strategy C++ wiring problem.

The revised remediation set comprises three tasks:

AZ-591 — compose_root per-binary bootstrap (Tier-1; todo/)

Parent gap: cross-cutting _STRATEGY_REGISTRY is empty in production source. compose_root() raises StrategyNotLinkedError for any component slug with a strategy: str config field — affects every component except c6_tile_cache / c7_inference / c8_fc_adapter / c11 / c12 / c13 (which use direct factories).
Goal: land runtime_root/airborne_bootstrap.py + operator_bootstrap.py that call register_strategy(...) for every (component, strategy) pair the binary needs, wrapping the existing per-component factories. Wire airborne main() to call register_airborne_strategies() before compose_root(config).
Complexity: 5 points.
Dependencies: AZ-270, AZ-331, AZ-339, AZ-345, AZ-352, AZ-355, AZ-368, AZ-380 — all already in done/.
Spec: _docs/02_tasks/todo/AZ-591_compose_root_per_binary_bootstrap.md.
Scheduled: Batch 66.

AZ-592 — AZ-332 Tier-2 validation bundle (Tier-2; backlog/)

Parent BLOCKED: AZ-332 (re-classified from FAIL).
Goal: rewrite _native/okvis2_binding.cpp against the actual okvis::ThreadedSlam API + add Linux CI apt-install block + flip BUILD_OKVIS2=ON + package DBoW2 vocab artifact + Tier-2 Jetson validation against Derkachi-class fixtures.
Complexity: 5 points placeholder (likely 8+; re-size when scheduled).
Dependencies: AZ-332, AZ-276, AZ-277, AZ-591 (must land first).
Spec: _docs/02_tasks/backlog/AZ-592_AZ-332_tier2_validation.md.
Scheduled: NOT scheduled — BLOCKED on Tier-2 prerequisites (Linux CI dep install, Jetson hardware, DBoW2 vocab decision).

AZ-593 — AZ-333 Tier-2 validation bundle (Tier-2; backlog/)

Parent BLOCKED: AZ-333 (re-classified from FAIL).
Goal: pick VINS-Mono upstream (HKUST + ROS-strip vs. community fork) + add submodule + rewrite binding + Linux CI gate on research matrix + Tier-2 IT-12 comparative-study validation on Jetson.
Complexity: 5 points placeholder (likely 8+; re-size when scheduled).
Dependencies: AZ-333, AZ-276, AZ-277, AZ-591, AZ-592 (CMake / Eigen pin overlap).
Spec: _docs/02_tasks/backlog/AZ-593_AZ-333_tier2_validation.md.
Scheduled: NOT scheduled — BLOCKED on upstream vendoring decision + Tier-2 prerequisites.

Gate decision (REVISED)

Per implement/SKILL.md § 15 the strict reading is "If any product task is FAIL, STOP". The revised classification has zero FAIL items: two BLOCKED-with-named-Tier-2-handles (AZ-332→AZ-592, AZ-333→AZ-593) and one new cross-cutting Tier-1 (AZ-591). The skill's STOP clause is satisfied because:

AZ-332 + AZ-333 are no longer FAIL — their original task specs explicitly designated the Tier-2 follow-up handle, which the gate now honors (per .cursor/rules/meta-rule.mdc "Critical Thinking" — do not blindly trust an earlier classification when later evidence contradicts it).
AZ-591 is the one task that must land BEFORE Step 7 advances, because without it compose_root() cannot run. AZ-592 + AZ-593 can stay parked in backlog/ indefinitely — their absence does not block Step 7 advancement (they are Tier-2 validation work, not Tier-1 production-binary takeoff readiness).

State: Step 7 stays in_progress until AZ-591 lands as part of Batch 66. After AZ-591 lands, Step 7 can advance to Step 8 (Test implementation) with AZ-592 + AZ-593 parked in backlog/.

Original (now superseded) remediation proposals

The original remediation proposals follow verbatim for audit. They led to creation of AZ-589 (Won't Fix) and AZ-590 (Won't Fix). Do not act on them — see the revised section above.

[Superseded] Proposed task 1 — `remediate_AZ-332_okvis2_threadedkfvio_wiring`

Parent FAIL: AZ-332.
Goal: wire okvis::ThreadedKFVio inside _native/okvis2_binding.cpp (_build_estimator() and _drive_estimator()); enable the commented-out includes; instantiate the estimator from yaml_config_; attach the output callback that fills latest_output_ under output_mtx_; CI matrix that installs Ceres + initialises OKVIS2's vendored submodules.
Complexity: 5 points.
Dependencies: AZ-332, AZ-276, AZ-277.
Out of scope: AC-9 honest-covariance Tier-2 validation against Derkachi-class fixtures (separate Tier-2 perf task).

[Superseded] Proposed task 2 — `remediate_AZ-333_vins_mono_estimator_wiring`

Parent FAIL: AZ-333.
Goal: wire vins_estimator::Estimator + feature_tracker inside _native/vins_mono_binding.cpp; enable the de-ROSified VINS-Mono pin build; ensure the same Protocol-conforming output shape as OKVIS2; research-only.
Complexity: 5 points.
Dependencies: AZ-333, AZ-276, AZ-277.
Out of scope: IT-12 comparative-study harness (lives in E-BBT).

If either remediation task grows beyond 5 points during decomposition, split into infrastructure + estimator-wiring + per-frame-cov-read sub-tasks before scheduling.

End of report.

26 KiB Raw Blame History Unescape Escape