gps-denied-onboard/_docs/02_document/architecture.md

GPS-Denied Onboard Pose Estimation — Architecture

Date: 2026-05-09 (Plan Phase 2a — initial draft). Inputs: _docs/00_problem/{problem,acceptance_criteria,restrictions}.md, _docs/00_problem/input_data/*, _docs/01_solution/solution.md, _docs/02_document/glossary.md, _docs/02_document/tests/*.

Architecture Vision

User-confirmed in Plan Phase 2a.0 (2026-05-09). This section is the spine of the document; nothing below it may contradict it without a recorded ADR.

The system is a Jetson Orin Nano Super-hosted onboard companion that delivers a GPS-equivalent WGS84 position (with honest 6×6 covariance and provenance label {satellite_anchored | visual_propagated | dead_reckoned}) to a fixed-wing UAV's flight controller in GPS-denied or GPS-spoofed environments. It runs as a single Python-with-C++-extensions monolithic process per binary track on the companion PC, fusing pre-flight-cached satellite tiles served by the parent-suite satellite-provider with live nav-camera frames (3 Hz) and FC-supplied IMU/attitude (100–200 Hz). A canonical hierarchical pipeline VIO → retrieval → re-rank → matching → AdHoP-conditional refinement → pose → fusion drives the per-frame loop within a 400 ms p95 latency budget. Cross-component coupling routes through a shared GTSAM substrate so posterior covariance is recovered natively (D-C5-5 = (c)). The companion is read-only against satellite-provider while airborne — both the pre-flight tile download and the post-landing tile upload run from the operator-side Tile Manager (C11), a separate binary that is excluded from the airborne CMake target so the companion image cannot load either code path even via reflection or config error (process-level isolation, AC-8.4).

Components — intent-level (formal decomposition belongs to Step 3)

• C1 — Visual / Visual-Inertial Odometry: pluggable VioStrategy (Okvis2 default, VinsMono in research builds only, KltRansac mandatory simple-baseline), config-selected at startup, not hot-swappable mid-flight.
• C2 — Visual Place Recognition: pre-cached satellite-tile retrieval (UltraVPR primary, MegaLoc secondary, MixVPR / SelaVPR / EigenPlaces / NetVLAD / SALAD additional candidates), all behind a single VprStrategy interface; concrete implementation chosen by config at startup.
• C2.5 — Top-N inlier-based re-rank: re-ranks the top-K=10 VPR candidates by single-pair LightGlue inlier count down to top-N=3.
• C3 — Cross-domain matcher: DISK+LightGlue (D-C3-1 = (a)) over the N=3 retained candidates; ALIKED+LightGlue secondary; XFeat alternate.
• C3.5 — AdHoP-conditional refinement: invoked only when initial reprojection residual exceeds threshold; bypassed otherwise to preserve AC-4.1.
• C4 — Pose estimation: OpenCV ≥4.12.0 solvePnPRansac (IPPE) wrapped in GTSAM Marginals for native 6×6 covariance recovery (D-C4-2 = (b); auto-degrades to Jacobian-based covariance D-C4-2 = (a) under thermal throttle per D-CROSS-LATENCY-1).
• C5 — State estimator: GTSAM iSAM2 + CombinedImuFactor + IncrementalFixedLagSmoother (K=10–20 keyframes, D-C5-3); native posterior covariance via Marginals; AC-4.5 = internal smoothing only, not FC retroactive correction.
• C6 — Tile cache + spatial index: PostgreSQL btree spatial index over filesystem ./tiles/{zoomLevel}/{x}/{y}.jpg mirroring satellite-provider's on-disk layout, plus FAISS HNSW index for VPR descriptors (.index written via faiss.write_index + atomicwrites + SHA-256 content-hash gate, D-C10-3).
• C7 — On-Jetson inference runtime: TensorRT 10.3 engines (Polygraphy / trtexec / IBuilderConfig hybrid orchestration), JetPack 6.2, SM 87; ONNX Runtime + TRT EP fallback; pure PyTorch FP16 baseline.
• C8 — Flight-Controller adapter: pymavlink GPS_INPUT for ArduPilot Plane (MAVLink 2.0 message signing on the companion ↔ AP wired channel, D-C8-9 = (d)) and YAMSPy / INAV-Toolkit MSP2_SENSOR_GPS for iNav (signing-gap accepted residual risk).
• C10 — Pre-flight cache provisioning: builds the model-derived cache artifacts (descriptor generation, engine compilation, manifest + content-hash) on top of an already-populated tile store; F2 takeoff verifier (D-C10-1, D-C10-3, D-C10-6, D-C10-7). C10 does NOT touch satellite-provider — tile network I/O lives in C11.
• C11 — Tile Manager (operator-side, distinct binary/image, ADR-004 process-isolated): owns operator-side network I/O against satellite-provider in both directions. TileDownloader interface fetches tiles into C6 during F1 (TLS + service-internal API key); TileUploader interface pushes mid-flight tiles to satellite-provider's ingest endpoint (D-PROJ-2 contract; not yet implemented service-side). C11 carries no flight-state gating of its own (Batch 44 SRP refactor) — the post-landing safety check lives in C12 (single source of truth). The component bundles both interfaces because they share auth, HTTP client, deployment unit, and the airborne-exclusion property.
• C12 — Operator Pre-flight Orchestrator (operator-side, same image as C11): orchestrates the operator-side workflows that C11 implements. Hosts the pre-flight cache provisioning UI, sector classification (active-conflict vs stable rear), the Flight resolver (FlightsApiClient → bbox + takeoff origin), the post-landing upload orchestrator (gates TileUploader on the flight_footer FDR record's clean_shutdown field — AZ-329), and the operator re-localization service (AC-3.4 visual-loss hint dispatched to the airborne companion over the GCS link via the OperatorCommandTransport Protocol; concrete pymavlink-backed impl is an E-C8 deliverable — AZ-330). The C12 ⇄ C11 boundary is a thin Protocol cut (TileUploaderCut) so C12 does not import C11 directly (AZ-507).
• C13 — Flight Data Recorder (FDR): per-flight ≤64 GB NVM record of estimates + IMU traces + emitted MAVLink + system health + mid-flight tiles + ≤0.1 Hz failed-tile thumbnails; raw nav/AI-cam frames excluded (AC-8.5, AC-NEW-3).
• External: satellite-provider (parent-suite .NET 8 service): tile producer pre-flight; tile sink post-landing (D-PROJ-2). Treated as a planned external dependency on the upload + voting paths.

Architectural principles / non-negotiables

1. Camera-specific math enters only via a Camera calibration artifact JSON (intrinsics + distortion + body-to-camera extrinsics + acquisition method factory_sheet | checkerboard_refined | hybrid). No hard-coded camera math anywhere; test fixtures (adti26) and production deployments (adti20) load different artifacts on the same code path.
2. VioStrategy is selected at startup via config; not hot-swappable mid-flight.
3. Build-time exclusion of unused Strategy implementations. A given binary links only the implementations it actually uses at runtime. The default deployment binary links the production-default strategies (e.g. OKVIS2 on C1) plus the engine-rule-mandatory simple-baseline (KltRansac on C1); the IT-12 comparative-study binary links all C1 implementations side-by-side. The mechanism is per-component CMake BUILD_* flags (BUILD_VINS_MONO, BUILD_SALAD, …) plus the per-binary composition root choosing among the linked implementations at startup. Justification is technical — binary size on the 8 GB shared Jetson, boot/load time inside the AC-NEW-1 30 s budget, deployed dependency / attack surface, and accidental-selection risk reduction (a binary with only OKVIS2 + KltRansac linked cannot be misconfigured into running VINS-Mono). Component licenses do not drive this decision — see ADR-002. CI emits both the deployment binary and the research binary on every PR.
4. In-air network I/O against satellite-provider is forbidden — in BOTH directions. Enforced primarily by process-level isolation — the Tile Manager (C11), which carries both the TileDownloader and the TileUploader interfaces, is not loaded in the airborne companion image. The defense-in-depth software guard is a C12-side flight_footer.clean_shutdown == True check (read by PostLandingUploadOrchestrator from the post-flight FDR via FdrFooterReader); C11 itself no longer gates (Batch 44 SRP refactor). The companion is read-only against C6 in flight; both pre-flight tile fetching and post-landing tile upload happen on the operator workstation.
5. All persistent imagery is in satellite-provider's on-disk tile format (./tiles/{zoomLevel}/{x}/{y}.jpg + matching metadata) so post-landing upload is byte-identical. No raw frames on disk except the AC-8.5 forensic ≤0.1 Hz failed-tile thumbnail log inside FDR.
6. Honest 6×6 posterior covariance via GTSAM Marginals is the safety floor for AC-NEW-4 and AC-NEW-7. Under-reported horiz_accuracy is a defect, not a tuning knob.
7. MAVLink 2.0 message signing on the companion ↔ ArduPilot wired channel, with per-flight key rotation (D-C8-9 = (d)). iNav has no signing equivalent — accepted residual risk, Plan-phase carryforward proposes an iNav firmware feature request.
8. D-CROSS-LATENCY-1 hybrid: K=3 baseline auto-degrades to K=2 + Jacobian covariance under Jetson thermal throttle, preserving AC-4.1 at +50 °C ambient at the cost of ~5–10 % accuracy loss (still inside AC-NEW-4).
9. Two execution tiers (Tier-1 workstation Docker = fast/cheap; Tier-2 Jetson hardware = AC-bound) appear in the deployment plan and CI matrix per finding F6.
10. Camera intrinsics and full-altitude footage are calibration prerequisites, not implementation gaps. Production accuracy claims are gated on D-PROJ-1 closure (hybrid factory + checkerboard refinement). Test fixtures use adti26 calibration sourced from public/factory references.
11. Spoofed GPS never re-enters the estimator unless the FC GPS report passes a three-part gate (AC-NEW-8 + AZ-490 follow-up): (a) FC GPS health stable + non-spoofed for ≥ 10 s, (b) a visual/satellite consistency check has succeeded on the next anchor frame, AND (c) the FC's reported position is within ≤ 200 m of the companion's last emitted PoseEstimate. The third clause is the mid-flight bounded-delta gate — even a "stable, non-spoofed" GPS frame is rejected if it disagrees with the companion's posterior by more than the configurable budget. Real GPS that passes the gate is fused via add_pose_anchor with the FC's covariance (treated as one more anchor source, never overriding the visual pipeline without the gate).
12. Operator-planned mission is the primary cold-start trust anchor, not the FC EKF (AZ-490 follow-up). The operator authors the route in the parent-suite Mission Planner UI (suite/ui), the route persists in the parent-suite flights REST service (suite/flights), and C12 (operator tooling) reads the Flight from that service to: (a) derive the cache bbox as the envelope of the waypoint lat/lon plus a configurable buffer, (b) extract the first-ordered waypoint as the takeoff origin (lat / lon / alt), and (c) bake the takeoff origin into the C10 Manifest so the airborne C5 can warm-start from it via set_takeoff_origin(origin, sigma_horiz_m, sigma_vert_m) before any FC IMU / VIO sample arrives. This unblocks the GPS-jammed-at-takeoff scenario the FC-EKF-only cold-start path (AZ-419 today) cannot handle. The FC EKF's last valid GPS becomes a secondary cold-start input — used only when the operator origin is missing from the Manifest OR when the FC EKF reading passes the same bounded-delta consistency check against the operator origin.
13. AC-4.5 is internal smoothing only. GTSAM iSAM2 retroactively refines past keyframes onboard and emits the corrected current frame; the FC log is forward-time only — neither ArduPilot nor iNav supports FC-side retroactive correction (Mode B Fact #107).
14. Interface-first components with constructor-injected dependencies. Every component is defined as an interface (Python Protocol or ABC) before any concrete implementation exists, lives in its own folder under src/components/<component>/, and is wired together via constructor injection at a single composition root. Components never reach out to a global registry, a singleton, or import a sibling component's concrete class directly — they receive their collaborators as __init__ arguments typed against the sibling's interface. Multiple interchangeable implementations of the same interface MUST be supported by design (e.g., C1 has three VioStrategy implementations; C2 has UltraVPR + MegaLoc + MixVPR + … behind a single VprStrategy; C8 has two FC-adapter implementations behind a single FcAdapter). Selection happens once, at startup, by config; the composition root resolves config → concrete implementation → wires the graph; the rest of the runtime sees only interfaces. Side benefit (NOTE): this design also gives the project packaging optionality — different combinations of BUILD_* flags can produce binaries tailored to specific deployment targets, customer bundles, or (if/when relevant later) end-product licensing strategies, without any source-level change in application code. That optionality is a consequence of the interface-first design, not a driver — the architectural decisions in this document are made on technical grounds; component licenses do not influence them. See ADR-002 § Consequences and ADR-009.

Open architectural items (tracked, NOT blocking Phase 2a)

• D-PROJ-1 (camera calibration acquisition): CLOSED in this Plan cycle as hybrid factory + checkerboard refinement (~1 day per deployed unit). No physical hardware available this cycle, so production calibration is documented as instructions only; runtime path uses test-fixture calibration for adti26 images.
• D-PROJ-2 (parent-suite satellite-provider ingest endpoint + multi-flight voting layer): open, parent-suite work, tracked in _docs/_process_leftovers/2026-05-09_satellite-provider-design-tasks.md. Onboard-side proceeds against the real satellite-provider — and uses an e2e-test-only mock-suite-sat-service fixture (under tests/fixtures/) to stand in for the not-yet-shipped POST contract during integration tests.
• D-PROJ-3 (multi-flight fixture acquisition for AC-NEW-4 / AC-NEW-7 statistical headroom): not pursued this cycle; AC-text was relaxed 2026-05-09 to Monte-Carlo-over-current-data with stated 95 % CI; multi-flight statistical headroom is residual risk in the Step 4 risk register.
• D-C8-2 runtime gate (companion-driven MAV_CMD_SET_EKF_SOURCE_SET switch): pattern is firmware-supported but not deployed-precedent. ArduPilot Plane SITL validation (IT-3) is the lock gate; D-C8-2-FALLBACK options recorded.
• D-C2-12 (DINOv2-feature-based matcher evaluation): carryforward research item; potentially closes D-C3-1 retrain cost.

---

1. System Context

Problem being solved: a fixed-wing UAV operating in eastern/southern Ukraine must continue to navigate and report position to its flight controller when GPS is denied (no fix) or spoofed (false fix). The onboard system replaces real GPS with a WGS84 position estimate derived from pre-cached satellite tiles + live nav-camera frames + FC IMU/attitude, with honest covariance and a provenance label. Mission profile: 8 h flights, ~60 km/h cruise, ≤1 km AGL, ≤400 km² total cached area.

System boundaries (inside vs outside):

Inside the system (this project)	Outside the system
Companion PC runtime (Jetson Orin Nano Super, JetPack 6.2)	Flight controller firmware (ArduPilot Plane, iNav)
All onboard pose-estimation logic (C1–C8, C13)	Parent-suite satellite-provider (.NET 8 REST microservice)
Pre-flight cache artifact build (C10 — engines + descriptors + manifest)	Parent-suite flights REST service (.NET 8; owns the Flight + Waypoint DTOs)
Operator-side Tile Manager (C11 — pre-flight download + post-landing upload)	Parent-suite Mission Planner UI (suite/ui — where operators plan the route)
Operator Pre-flight Orchestrator (C12)	GCS (QGroundControl)
FDR writer (C13)	Nav camera hardware (adti20); AI-camera hardware
Camera calibration artifact format + loader	UAV airframe / FC IMU / sensors
	Operator's workstation OS / authentication
	The act of calibration itself (operator runs checkerboard rig)

External systems:

System	Integration Type	Direction	Purpose
satellite-provider (parent-suite .NET 8)	REST + filesystem (read), REST (post-landing write, D-PROJ-2)	Both	Pre-flight tile source; post-landing tile sink (planned)
flights REST service (parent-suite .NET 8)	REST (read) over HTTPS	Inbound to C12	Source of the operator-planned Flight (waypoints, ordering, altitudes). C12 derives bbox + takeoff origin from the Flight. Operator workstation only — never reached from the airborne companion
Mission Planner UI (suite/ui)	Indirect via flights REST	Inbound (mediated)	Where the operator authors the route before C12 consumes it. Out of scope for this project, but the API contract it produces IS in scope
ArduPilot Plane FC	MAVLink 2.0 over UART/USB (signed)	Both	Inbound: external position via GPS_INPUT. Outbound: IMU, attitude, GPS health, EKF source-set commands
iNav FC	MSP2 over UART (unsigned), MAVLink outbound	Both	Inbound: external position via MSP2_SENSOR_GPS (companion is sole GPS source on iNav). Outbound: IMU/attitude/telemetry
QGroundControl (GCS)	MAVLink 2.0 (link-bandwidth-limited)	Both	1–2 Hz downsampled summary out (AC-6.1); operator commands in (AC-6.2)
Nav camera (USB/MIPI-CSI/GigE)	Camera SDK / V4L2	Inbound	3 Hz nadir frames at 5472×3648 px
AI camera	Camera SDK + gimbal/zoom telemetry	Inbound	AC-7.x object localization (deferred to follow-up cycle)
Operator workstation	Filesystem + USB/Ethernet	Both	Pre-flight: stages cache + calibration onto companion. Post-flight: triggers upload tool, reads FDR

---

2. Technology Stack

Layer	Technology	Version	Rationale
Language (host)	Python	3.10 (JetPack 6.2 default)	Glue layer for GTSAM/FAISS/OpenCV/pymavlink/YAMSPy; matches every selected library's primary binding
Language (perf-critical)	C++	C++17	OKVIS2, VINS-Mono, GTSAM core, OpenCV, FAISS native; Python wrappers cross the boundary
Inference runtime	TensorRT	10.3 (JetPack 6.2 pin)	Pinned per D-C7-9; fallback ONNX Runtime + TRT EP; pure PyTorch FP16 baseline for mandatory simple-baseline track
Visual matching	DISK + LightGlue	upstream HEAD pinned per Plan-phase	D-C3-1 = (a); replaces SuperPoint+SuperGlue (Magic Leap noncommercial canonical)
VPR (primary)	UltraVPR	RAL 2025 / ICRA 2026 (cbbhuxx/UltraVPR)	Documentary Lead PRIMARY; rotation-invariant, unsupervised aerial pretrain (multi-heading aerial flight + closes D-C2-1 retrain cost)
VPR (secondary)	MegaLoc, MixVPR, SelaVPR, EigenPlaces, NetVLAD	upstream HEAD pinned per Plan-phase	Mode B Fact #110/#113 + mandatory simple-baseline (NetVLAD/MixVPR)
State estimator	GTSAM + gtsam_unstable.IncrementalFixedLagSmoother	per Plan-phase pin (no published CVE at audit time)	Native 6×6 covariance; D-C5-5 = (c) PriorFactorPose3 only
Image / pose math	OpenCV (Python+C++)	≥ 4.12.0	CVE-2025-53644 mitigation (Mode B Fact #112); IPPE flags for D-C4-1 = (b)
VPR descriptor index	FAISS HNSW	upstream HEAD pinned per Plan-phase	faiss.write_index + atomicwrites + SHA-256 content-hash gate (D-C10-3)
FC adapter (ArduPilot)	pymavlink + MAVLink 2.0 signing	bundled unmodified per D-C8-3	Verified Source #4; ArduPilot canonical signing per Source #128
FC adapter (iNav)	YAMSPy + INAV-Toolkit MSP2	MIT throughout	iNav has no inbound MAVLink ext-positioning handler (SQ6)
VIO (production)	OKVIS2 (BSD-3-Clause)	upstream HEAD pinned per Plan-phase	D-C1-1-SUB-A = (a) production-default
VIO (research / IT-12)	VINS-Mono	upstream HEAD pinned per Plan-phase	Research binary only (BUILD_VINS_MONO=ON) for IT-12 comparative study; build-time exclusion from deployment binary per ADR-002
VIO (mandatory baseline)	KLT+RANSAC over OpenCV	OpenCV ≥ 4.12.0	Engine-rule-required mandatory simple-baseline
Tile cache backend	PostgreSQL + filesystem	PostgreSQL 16 (mirror of satellite-provider)	C6 mirrors satellite-provider's on-disk and table layout so C11 TileUploader's post-landing payload is byte-identical to what the parent suite already serves
Container runtime	Docker (Tier-1) + bare JetPack (Tier-2)	Docker 27.x; JetPack 6.2	Tier-1 workstation Docker; Tier-2 Jetson native (no Docker — direct JetPack to keep INT8 calibration cache trustworthy per D-C10-6)
Build system	CMake + Python pyproject.toml	CMake ≥ 3.27	CMake option(BUILD_VINS_MONO ...) D-C1-1-SUB-A; Python wheels built per Jetson via cibuildwheel-equivalent recipe
CI/CD	GitHub Actions (Tier-1) + self-hosted Jetson runner (Tier-2)	latest pinned action versions	Two-binary emit on every PR (production + research); Tier-2 runs are AC-bound jobs only
Configuration	YAML (per-flight) + Camera calibration JSON	n/a	Single config root; the only camera-specific entry point is the calibration JSON

Key constraints from restrictions.md and how they shape the stack:

• Hardware pinned to Jetson Orin Nano Super (8 GB shared, 25 W) → forces TensorRT engine compilation on-device + INT8/FP16 mix per D-C7-1; rules out heavy multi-process stacks (D-C1-1-SUB-A = (b) was rejected on latency budget).
• Python is the host language but ROS-bound C++ is unavoidable for VIO → both production and research binaries are CMake projects that produce a Python-importable .so per VioStrategy; the rest of the runtime is pure Python.
• PX4 is out of scope, ArduPilot Plane + iNav both required → C8 must split per FC, with no single message contract spanning both.
• Build-time exclusion of unused Strategy implementations (ADR-002) → CMake BUILD_* flags (BUILD_VINS_MONO, BUILD_SALAD, …) determine which implementations are linked into each binary; the deployment binary links the production-default + the mandatory simple-baseline; the IT-12 research binary links all strategies. Justification is technical (binary size on 8 GB shared Jetson, AC-NEW-1 boot budget, dependency surface, accidental-selection risk). Component licenses do not influence this decision.
• MAVLink message-signing posture asymmetry → pymavlink signing handshake is part of takeoff load on the AP path; iNav unsigned link is documented as accepted residual risk in security_analysis.md carryforward.
• No raw-frame storage (AC-8.5) → all camera ingestion is streaming; the only persistence path for frame imagery is via tile orthorectification (AC-8.4).
• 8 h continuous duty cycle at 25 W up to +50 °C ambient → the auto-degrade hybrid (D-CROSS-LATENCY-1) is a first-class concern of every latency-sensitive component, not an afterthought.

---

3. Deployment Model

Environments:

Environment	Purpose	Hardware
dev-tier1	Fast iterative development; unit + most integration tests	Workstation (any Linux x86_64 + NVIDIA GPU optional); Docker
dev-tier2	Hardware-bound development checks	Jetson Orin Nano Super dev kit (developer's desk)
staging-tier1	CI runs that don't require Jetson hardware	GitHub-hosted runner (x86_64); Docker
staging-tier2	CI runs that require Jetson (AC-bound jobs only)	Self-hosted Jetson runner; bare JetPack (no Docker)
production	Deployed companion image on a UAV	Jetson Orin Nano Super (pinned); bare JetPack; no inbound network listening (defense-in-depth, NFT-SEC-05)
production-operator-workstation	Operator-side workflows orchestrated by C12: pre-flight tile download (C11 TileDownloader), cache artifact build (C10), post-landing tile upload (C12 PostLandingUploadOrchestrator → C11 TileUploader), AC-3.4 re-loc hint dispatch (C12 OperatorReLocService), FDR retrieval	Operator's Linux workstation; Docker for satellite-provider mirror

Infrastructure:

• No cloud orchestration. The companion is an embedded edge device; the operator's workstation is a single host that runs the operator tooling (C11 Tile Manager + C12 Operator Pre-flight Orchestrator) and a local satellite-provider mirror or VPN-reaches the lab satellite-provider.
• Two airborne binaries shipped on every PR (ADR-002): deployment-binary (links the production-default strategy on each component + the mandatory simple-baseline; CMake BUILD_VINS_MONO=OFF, BUILD_SALAD=OFF, …) and research-binary (links every available strategy on every component; all BUILD_* flags ON, used for the IT-12 comparative study). The deployment binary is what installs onto an operational Jetson; the research binary runs on dev/lab Jetson hardware for the comparative-study report. The same code base produces both — ADR-002 mechanism scales to additional binary variants later if packaging strategy requires it. Replay is not a separate binary (ADR-011): the deployment-binary runs both live and replay modes from the same image, swapping FrameSource / FcAdapter / MavlinkTransport strategies at startup based on config.mode. A third binary — operator-orchestrator (C10 + C11 + C12) — ships from the same source tree for the operator workstation; the airborne deployment-binary does NOT contain the operator-orchestrator components (ADR-004 process isolation).
• Container scope: Tier-1 uses Docker (docker compose for the developer setup including a mock-suite-sat-service container, the operator-orchestrator container, and a Postgres for C6). Tier-2 (Jetson) does NOT use Docker — TensorRT INT8 calibration caches and jetson-stats thermal telemetry are most reliable without a container layer, per D-C7-9 + D-C10-6. The deployed image on the Jetson is a JetPack-based system image with the deployment binary preinstalled.
• Scaling: not applicable (per-UAV, single companion). Failover is per-airframe (the FC's IMU-only fallback at AC-5.2 is the system's "scale-out").

Environment-specific configuration:

Config	dev-tier1	staging-tier2	production
satellite-provider host	local Docker (satellite-provider:5100)	real satellite-provider Docker (download path; existing) + e2e-test mock-suite-sat-service fixture (POST/upload only, until D-PROJ-2 lands)	operator workstation (pre-flight only)
Camera calibration source	test-fixture artifact (adti26.json)	test-fixture artifact	adti20.json (D-PROJ-1 hybrid output)
Logging sink	console (DEBUG)	journald + FDR	FDR (per-flight, ≤ 64 GB rolling)
MAVLink signing key	dev key (committed to test fixtures)	per-flight key from test config	per-flight key generated at takeoff load, rotated per flight
Inference engine source	pre-built engines OR on-the-fly compile	pre-built (Tier-2 cache)	pre-built (verified content-hash gate)
BUILD_VINS_MONO (binary track)	both (developer's choice)	both	OFF (production-only)
Network egress	unrestricted	locked to test endpoints	none in flight (DNS blackhole + iptables OUTPUT REJECT, NFT-SEC-05)

Image / artifact pipeline:

source repo
   ├─→ CI matrix
   │     ├─ tier1 lint + unit + most integration → Docker
   │     ├─ tier1 build production-binary + research-binary (CMake split)
   │     ├─ tier1 SBOM diff (production must NOT include vins_mono)
   │     └─ tier2 (self-hosted Jetson) AC-bound suite (NFT-PERF-*, NFT-LIM-*, IT-12)
   │
   ├─→ release artifacts:
   │     ├─ deployment-binary tarball (production-default strategies + mandatory baselines + replay strategies, ADR-002 + ADR-011; runs both live and replay modes from a single image)
   │     ├─ research-binary tarball (all strategies linked; for IT-12 comparative study; also includes replay strategies)
   │     ├─ JetPack image (deployment-binary preinstalled)
   │     └─ operator-orchestrator tarball (C11 + C12 + e2e-test mock-suite-sat-service compose for offline integration testing)
   │
   └─→ deploy paths:
         ├─ Jetson operational deploy: JetPack image flash (deployment-binary)
         ├─ Lab/research deploy: research-binary install on dev Jetson
         └─ Operator workstation: Docker compose for C11+C12+local satellite-provider mirror

---

4. Data Model Overview

Detailed per-component data models live in component specs (Step 3); per-entity migration strategies live in data_model.md (Phase 2b).

Core entities:

Entity	Description	Owned by component
NavCameraFrame	5472×3648 px nadir RGB frame + capture timestamp + camera ID	Camera ingest → C1, C2
ImuSample / ImuWindow	IMU sample (accel + gyro + timestamp) at 100–200 Hz; windowed view sent to C1	FC adapter (C8 inbound side)
VioOutput	Per-frame relative pose SE(3) + 6×6 covariance + IMU bias estimate + feature quality	C1
VprQuery	Image embedding (UltraVPR/MegaLoc/etc)	C2
VprResult	Top-K=10 candidate tile IDs ranked by descriptor distance	C2
RerankResult	Top-N=3 candidate tiles ranked by inlier count	C2.5
MatchResult	2D-3D correspondences with RANSAC inliers from C3 / C3.5	C3, C3.5
CameraCalibration	Intrinsics K + distortion + body-to-camera extrinsics + acquisition method	Loaded once at startup; consumed by C1, C3, C4
PoseEstimate	WGS84 position + 6×6 covariance + provenance label + last_satellite_anchor_age_ms	C4 → C5
Tile	JPEG body + center lat/lon + zoomLevel + tile_size_meters/pixels + capture_timestamp + source + freshness flag + (mid-flight only) quality_metadata	C6
TileQualityMetadata	estimator_label, 2×2 covariance sub-matrix, last_anchor_age_ms, MRE, IMU bias norm — sufficient for D-PROJ-2 voting	C6 (write side from C5/C4 outputs)
EmittedExternalPosition	WGS84 + honest horiz_accuracy + per-FC encoding (MAVLink GPS_INPUT for AP, MSP2 MSP2_SENSOR_GPS for iNav)	C8
FlightStateSignal	`IN_AIR	ON_GROUNDboolean derived from FCMAV_STATE`
FlightFooterRecord	{flight_id, clean_shutdown, total_records, segment_count, …} — single FDR record written by C13 on clean shutdown	C13 (writer) → C12 PostLandingUploadOrchestrator (reader, via FdrFooterReader)
PostLandingUploadRequest	{flight_id, satellite_provider_url, api_key, batch_size}	C12 CLI → C12 PostLandingUploadOrchestrator
ReLocHint	Operator-supplied position hint for AC-3.4 visual-loss re-localization: {approximate_position_wgs84: LatLonAlt, confidence_radius_m, reason}; validated at construction (lat ∈ [-90,90]; lon ∈ (-180,180]; radius > 0; reason non-empty); emitted to airborne companion via OperatorCommandTransport Protocol (E-C8 concrete)	Operator CLI → C12 OperatorReLocService → (GCS link) airborne companion
FdrRecord	Estimates + IMU traces + emitted MAVLink + system health + tiles + thumbnails (≤ 64 GB / flight)	C13
Manifest	Hash of (model + calibration + corpus + sector classification + takeoff origin) for D-C10-1 idempotence	C10
EngineCacheEntry	TRT engine + INT8 calibration cache keyed by SM/JP/TRT/precision tuple (D-C10-7)	C10, C7
SectorClassification	`active_conflict	stable_rear` per area, drives freshness threshold
Flight	Operator-planned mission: ordered Waypoint list + metadata, persisted in the parent-suite flights REST service. Read by C12 via FlightsApiClient; never reached from the airborne companion	External (suite/flights) → C12
Waypoint	Ordered (lat, lon, alt, objective, source) entry inside a Flight. C12 envelopes waypoint lat/lon → bbox; first-ordered waypoint → takeoff origin	External (suite/flights) → C12
TakeoffOrigin	LatLonAlt carried in the C10 Manifest; baked in by C12 at build time from Flight.waypoints[0]; consumed at boot by C5 via set_takeoff_origin(origin, sigma_horiz_m, sigma_vert_m) (AZ-490)	C12 → C10 Manifest → C5

Key relationships:

• NavCameraFrame → VioOutput (via C1) and VprQuery (via C2): same frame, two consumers.
• VprResult.tileIds ⊆ Tile.id (FK into the tile cache).
• MatchResult references both NavCameraFrame.id and Tile.id (cross-domain pair).
• PoseEstimate aggregates MatchResult + VioOutput + ImuWindow through C4 + C5.
• EmittedExternalPosition is a per-FC projection of PoseEstimate; the projection rule lives in C8 (per-FC unit conversion D-C8-8 = (b)).
• Tile (mid-flight) is produced from NavCameraFrame + PoseEstimate via orthorectification; carries TileQualityMetadata referencing the PoseEstimate it was emitted from.
• FdrRecord is the union of all emitted streams + all inputs (excluding raw nav/AI-cam frames); rollover policy = oldest segment dropped first.

Data flow summary (one-line each; full sequences in system-flows.md):

• Pre-flight: satellite-provider → C11 TileDownloader → Tile cache (C6) → C10 → EngineCacheEntry + Manifest + descriptor .index (atomic write + content-hash gate).
• Takeoff load: Manifest content-hash verify + FAISS mmap + TRT deserialize + MAVLink signing handshake → ready.
• Per-frame runtime: NavCameraFrame + ImuWindow → C1 (VioOutput) → C2 → C2.5 → C3 → C3.5 → C4 → C5 → C8 → EmittedExternalPosition to FC.
• Mid-flight tile gen: NavCameraFrame + PoseEstimate → orthorectify → dedup → write to local C6 (no upload).
• GCS telemetry: C5 → C8 → 1–2 Hz downsampled summary to QGroundControl.
• FDR: every emitted/received stream → C13 ring with per-flight ≤ 64 GB cap.
• Post-landing: operator triggers C12 PostLandingUploadOrchestrator → reads flight_footer from FDR via FdrFooterReader → on clean_shutdown == True invokes C11 TileUploader (via TileUploaderCut Protocol) → reads C6 → uploads to satellite-provider ingest endpoint (D-PROJ-2 contract). Refusal modes (footer_missing, unclean_shutdown, flight_id_not_found, fdr_unreadable) raise FlightStateNotConfirmedError with operator-actionable remediation text and a distinct CLI exit code per mode.
• Operator re-loc (AC-3.4 visual-loss path): operator submits ReLocHint via the reloc-confirm CLI → C12 OperatorReLocService validates the DTO → forwards to airborne companion via OperatorCommandTransport (E-C8 concrete) → records c12.reloc.requested FDR record (outcome ∈ {sent, failed}). Live log redaction (lat/lon rounded to 5 decimals; reason truncated to 200 chars); FDR record persists the full hint un-redacted for post-flight forensics.

---

5. Integration Points

Internal Communication

All in-process Python calls; the system is a single host process per binary track. "Pattern" describes the interaction shape.

From	To	Protocol	Pattern	Notes
Camera ingest thread	C1 (VioStrategy.process_frame)	In-process queue (bounded, drop-oldest)	Producer-consumer	Frame skip is allowed under sustained load (AC-4.1 "~10% may drop")
Camera ingest thread	C2 (vpr_pipeline.query)	In-process queue (bounded, drop-oldest)	Producer-consumer	Same frame fan-out, distinct queue depths
C2	C2.5	Direct call	Function call	C2.5 wraps C3 matcher; no queue
C2.5	C3 / C3.5	Direct call	Function call	C3.5 invoked iff MatchResult.reprojection_residual > threshold
C3 / C3.5	C4	Direct call	Function call	MatchResult passed as DTO
C1 + C4	C5	In-process queue (timestamp-ordered merge)	Pub/sub	C5 holds the GTSAM iSAM2 state; one writer thread
C5	C8 (FC outbound)	In-process queue (per-FC encoder)	Pub/sub	One encoder per active FC profile; selected at startup
C8 (FC inbound)	C1 (ImuWindow), C5 (FC IMU/attitude prior)	In-process pub/sub (timestamp-aligned)	Pub/sub	Single source of truth for FC IMU; both consumers see the same window
C8 (FC inbound)	flight-state guard (process boundary)	In-process pub/sub	Event	Used by FDR + GCS heartbeat; airborne companion does not load C11 at all
C5 → orthorectifier → C6	C6 (write-only while airborne)	In-process function call	Command	Write path is in-process; the in-air image has no upload code path
All components	C13 (FDR writer)	In-process queue (lossy on overrun)	Pub/sub	Overrun = logged rollover, never silent drop (AC-NEW-3)

External Integrations

External system	Protocol	Auth	Rate limits	Failure mode
ArduPilot Plane FC	MAVLink 2.0 (GPS_INPUT 5 Hz; MAV_CMD_SET_EKF_SOURCE_SET; STATUSTEXT / NAMED_VALUE_FLOAT) over UART/USB	MAVLink 2.0 message signing, per-flight key (D-C8-9 = (d))	5 Hz periodic emit; signing handshake at takeoff load (≤ 5 s, AC-NEW-1)	Signing handshake fail → companion refuses takeoff; mid-flight signing key compromise → FC ignores unsigned messages, AC-5.2 takes over
iNav FC	MSP2 MSP2_SENSOR_GPS over UART; MAVLink outbound for telemetry	None (iNav has no signing) — accepted residual risk per Mode B Source #129	5 Hz periodic emit	Mid-flight bad-frame → iNav mspGPSReceiveNewData() receives only the latest frame; honest hPosAccuracy is the only safety net
QGroundControl (GCS)	MAVLink 2.0 (STATUSTEXT, NAMED_VALUE_FLOAT, GPS_RAW_INT)	Same MAVLink 2.0 signing as the AP path (AP profile); no signing on iNav profile	1–2 Hz downsampled (AC-6.1); operator commands are best-effort	GCS link drop → companion continues; no mid-flight reconfiguration is required from GCS
satellite-provider (pre-flight)	REST over HTTP, OpenAPI at /swagger; filesystem access if co-located	TLS + service-internal API key (operator workstation only); the companion never reaches satellite-provider directly while airborne	Off-line pre-flight; not time-critical	Cache miss → C11 TileDownloader fails fast pre-flight; C10 build is blocked downstream; takeoff blocked
satellite-provider (post-landing ingest, D-PROJ-2, planned)	REST POST /api/satellite/tiles/ingest (multipart)	Per-flight onboard signing key (carried with each tile); rate-limited	Bursty post-landing	Endpoint not yet implemented service-side → C11 keeps batches queued locally; never blocks the pre-flight cycle
Operator workstation (pre-flight stage)	Filesystem (USB / Ethernet)	OS-level (operator login)	Not time-critical	Bad-stage detection via Manifest content-hash gate (D-C10-3)
Nav camera	USB / MIPI-CSI / GigE (lens-module dependent)	n/a	3 Hz	Frame drop / hardware fault → "VISUAL_BLACKOUT" path (AC-3.5, AC-NEW-8)

satellite-provider upload contract (per D-PROJ-2 carryforward)

The onboard side of D-PROJ-2 is fully specified in _docs/_process_leftovers/2026-05-09_satellite-provider-design-tasks.md. From this architecture's standpoint:

• Tile writes are append-only and idempotent (the same (zoomLevel, lat, lon, capture_timestamp, companion_id, flight_id) tuple is the dedup key).
• Quality metadata is mandatory on every uploaded tile so the planned voting layer can promote pending → trusted without re-deriving statistics on the service side.
• Onboard tiles never claim the trusted status; they are uploaded as pending and the parent-suite voting layer (D-PROJ-2 design task #2) decides promotion.
• Test substitute: mock-suite-sat-service is an e2e-test-only fixture (under tests/fixtures/mock-suite-sat-service/) that implements the upload contract for NFT-SEC-01 / FT-P-17 / IT runs until D-PROJ-2 lands service-side. It is not a component in the architectural sense — the production architectural counterparty for both download and upload is the real satellite-provider. The fixture is retired the moment the real ingest endpoint ships.

---

6. Non-Functional Requirements

Targets are taken verbatim from acceptance_criteria.md and tests/traceability-matrix.md. The tests column points to the canonical tests/ files where each NFR is exercised.

Requirement	Target	Measurement	Priority	Tests
End-to-end latency (AC-4.1)	p95 ≤ 400 ms (steady-state and thermal-throttle hybrid)	NFT-PERF-01 (Tier-2); D-CROSS-LATENCY-1 partition	High	tests/performance-tests.md
Tail latency under thermal stress (AC-NEW-5 + AC-4.1)	p99 ≤ 600 ms; p95 ≤ 400 ms at +50 °C 8 h	NFT-9 hot-soak	High	tests/performance-tests.md
Memory cap (AC-4.2)	< 8 GB shared (CPU + GPU) on Jetson Orin Nano Super	NFT-LIM-01 8 h replay	High	tests/resource-limit-tests.md
Cold-start TTFF (AC-NEW-1)	p95 < 30 s from companion boot to first valid frame	NFT-PERF-03 (50× cold boot)	High	tests/performance-tests.md
Spoofing-promotion latency (AC-NEW-2)	p95 < 3 s on each FC	NFT-PERF-04 (SITL on AP + iNav)	High	tests/performance-tests.md
FDR storage (AC-NEW-3)	≤ 64 GB / flight; no silent drops	NFT-LIM-02 8 h synthetic	Medium	tests/resource-limit-tests.md
False-position safety (AC-NEW-4)	P(err > 500 m) < 0.1 %; P(err > 1 km) < 0.01 %, with stated 95 % CI over current corpus	NFT-RES-03 Monte Carlo	High	tests/resilience-tests.md
Operating envelope (AC-NEW-5)	−20 °C to +50 °C; 25 W; 8 h no throttle	NFT-LIM-04 workstation baseline (chamber deferred)	High	tests/resource-limit-tests.md
Imagery freshness (AC-NEW-6, AC-8.2)	Reject/downgrade tiles violating 6 mo / 12 mo thresholds	FT-N-05 / FT-N-06	High	tests/blackbox-tests.md
Cache-poisoning safety (AC-NEW-7)	Onboard-side: P(misalign > 30 m) < 1 %, P(> 100 m) < 0.1 %, with stated 95 % CI	NFT-SEC-01 onboard Monte Carlo + synthetic over-confidence injection	High	tests/security-tests.md
Visual blackout failsafe (AC-NEW-8)	Mode transition ≤ 400 ms; covariance grows monotonically; spoofed GPS never re-promoted without 10 s + visual consistency gate	FT-N-04 + NFT-RES-04	High	tests/resilience-tests.md + tests/blackbox-tests.md
Cross-FC covariance honesty (AC-NEW-4 cross-FC)	horiz_accuracy (m, AP) and hPosAccuracy (mm, iNav) carry mathematically equivalent values from the same 2×2 sub-matrix	IT-10 cross-FC	High	tests/blackbox-tests.md
MAVLink message-signing posture (AC-4.3 + D-C8-9)	Signing enabled on AP wired channel; per-flight key rotation logged to FDR; iNav documented residual risk	NFT-8 + NFT-SEC-03	High	tests/security-tests.md
Dependency CVE pinning (D-CROSS-CVE-1)	OpenCV ≥ 4.12.0; SBOM clean of unpatched CVEs at audit time; monthly re-scan	NFT-10 SBOM CVE audit	High	tests/security-tests.md
GCS bandwidth budget (AC-6.1)	1–2 Hz downsampled summary	FT-P-12	Medium	tests/blackbox-tests.md
Frame-by-frame streaming (AC-4.4)	No batching/delay; estimates emitted per frame	NFT-PERF-02	High	tests/performance-tests.md
Smoothing-loop look-back (AC-4.5, Mode B Fact #107)	FDR contains smoothed past-frame estimates; smoothing horizon converges within X m of ground truth at K = 10–20 keyframes	IT-11	Medium	tests/blackbox-tests.md

---

7. Security Architecture

Threat model (one-page summary; full extraction lives in carryforward security_analysis.md):

• The companion is a remote untrusted endpoint from the parent-suite's standpoint: a downed UAV's companion can be physically captured. Persistent secrets must therefore be per-flight ephemeral wherever feasible.
• The wired companion ↔ FC link is the only physical-access-required attack surface for in-flight injection. MAVLink 2.0 signing on the AP path mitigates CVE-2026-1579 (D-C8-9 = (d)). iNav has no signing — accepted residual risk.
• The GCS link is bandwidth-limited and best-effort; a hostile GCS can spoof operator commands but cannot inject pose data (the system never accepts pose from GCS).
• GPS spoofing is treated as expected, not anomalous (AC-3.5, AC-NEW-2, AC-NEW-8). The system never lets a spoofed GPS source re-enter the estimator without a 10 s + visual-consistency gate.
• Cache poisoning is the dominant cross-flight attack vector (AC-NEW-7): a compromised companion could write a misaligned tile that becomes the next flight's anchor. The mitigation has two halves: onboard (honest covariance + quality metadata) and parent-suite (D-PROJ-2 voting layer, not yet implemented).
• Pre-flight cache stage is on the operator's workstation; the SHA-256 content-hash gate (D-C10-3) detects in-place tampering between stage and takeoff.
• In-flight network egress is forbidden (defense-in-depth: DNS blackhole + iptables OUTPUT REJECT, NFT-SEC-05). The only outbound path from the companion is MAVLink to the FC and signed STATUSTEXT to the GCS.

Authentication (per integration):

Integration	Mechanism
Companion ↔ ArduPilot Plane FC	MAVLink 2.0 message signing, per-flight key rotation (D-C8-9 = (d))
Companion ↔ iNav FC	None (iNav has no signing implementation; accepted residual risk per Mode B Source #129)
Companion ↔ GCS (AP profile)	MAVLink 2.0 signing inherited from the FC channel
Operator workstation ↔ satellite-provider (pre-flight)	TLS + service-internal API key (workstation only; never on the airborne companion)
Companion ↔ satellite-provider (post-landing upload, D-PROJ-2 planned)	Per-flight onboard signing key carried with each uploaded tile; the planned ingest endpoint verifies the key
Operator workstation pre-flight stage	OS-level (operator login + workstation hardening — operator-orchestrator concern, C12)

Authorization:

• Onboard runtime: a single principal (the runtime process); no in-process privilege boundaries. The Tile Manager (C11) runs as a different principal on the operator workstation, holding the only credentials that reach satellite-provider (TLS API key for download; per-flight onboard signing key for post-landing upload). The airborne image does not contain the C11 binary at all.
• GCS: operator commands (AC-6.2) are best-effort hints; the operator cannot promote a pose, override covariance, or reach the satellite-provider write path. Operator re-loc requests (C12 OperatorReLocService → OperatorCommandTransport over the GCS link) trigger the satellite re-localization flow (F6) but do not bypass any safety gate — the airborne pipeline still validates the hint against the visual/satellite consistency check before promoting any pose.

Data protection:

• At rest: tile cache + descriptor index + FDR are written to the companion's local NVM. No application-level encryption (the threat model treats a captured companion as compromised; encryption would buy little against physical access). Operator-side satellite-provider storage is the parent-suite's concern.
• In transit: MAVLink 2.0 message signing on the AP channel; MSP2 unsigned on iNav. The post-landing upload runs over TLS to satellite-provider.
• Secrets management:
  • Per-flight MAVLink signing key: generated at takeoff load; rotated per flight; logged to FDR.
  • Per-flight onboard signing key for tile upload: generated at takeoff load; baked into mid-flight tile metadata; consumed by C11 post-landing.
  • Pre-flight service API key: stays on the operator workstation; never written to the companion image.
  • No long-lived secrets on the companion image beyond firmware-level boot signatures (out of scope).

Audit logging:

What	Where	Retention
All emitted external-position frames + covariance + provenance label	FDR (C13)	per flight (≤ 64 GB; rollover oldest-first)
All received MAVLink + MSP2 frames (raw tlog stream)	FDR	per flight
MAVLink 2.0 signing key rotation events	FDR	per flight
Spoofing-promotion / spoofing-rejection events	FDR + GCS STATUSTEXT	per flight + best-effort GCS link
VISUAL_BLACKOUT_* STATUSTEXT events (AC-3.5, AC-NEW-8)	FDR + GCS STATUSTEXT	per flight + best-effort
C10 content-hash gate fail events	FDR + companion refuses takeoff	per flight
Mid-flight tile-gen failures	≤ 0.1 Hz thumbnail log inside FDR (AC-8.5 forensic exception)	per flight
Component health (CPU/GPU/temp/throttle)	FDR	per flight
Source-set switch events (D-C8-2 EKF source-set)	FDR + GCS STATUSTEXT	per flight
Production binary SBOM provenance	release artifacts; not on the deployed companion	per release

---

8. Key Architectural Decisions

These ADRs distill the user-confirmed Mode-B locks plus this architecture's first-time choices. ADRs are also tracked in _docs/00_research/06_component_fit_matrix/MODEB_revisions.md and (for cross-component gates) 99_cross_component_gates.md. Step 4 (Risk Review) iterates on them; this section is the authoritative entry point.

ADR-001 — VioStrategy is selected at startup via config; not hot-swappable

Context: Three VIO implementations are required (OKVIS2 production-default, VINS-Mono research-only, KLT+RANSAC mandatory simple-baseline). Hot-swap mid-flight would add re-initialisation cost on every switch and would require keeping multiple solvers warm in 8 GB shared memory.

Decision: VioStrategy is selected at startup from a single config knob (vio.strategy: okvis2 | vins_mono | klt_ransac), and the choice is constant for the flight. The VioStrategy interface owns the abstraction; concrete strategies own their per-strategy concerns (OKVIS2's ROS bring-up, VINS-Mono's build flag, KLT's degraded covariance). Build-time inclusion / exclusion of individual strategies is governed separately by ADR-002.

Alternatives considered:

1. Hot-swap at runtime — rejected: re-init cost + memory footprint inside AC-4.2.
2. Single-strategy build per binary — rejected: defeats the IT-12 comparative-study objective on the research binary.

Consequences: A flight is locked to one VIO; failure of the active strategy = AC-5.2 fallback (FC IMU-only). The comparative study is a per-replay artifact, not a runtime decision.

ADR-002 — Build-time exclusion of unused Strategy implementations (D-C1-1-SUB-A = (a))

Context: The architecture deliberately requires multiple interchangeable implementations per component (three VioStrategy for C1; multiple VprStrategy for C2; two FC adapters for C8). At runtime each binary uses exactly one of them per component. Linking all implementations into every binary would inflate binary size on the 8 GB shared Jetson, increase boot/load time inside the AC-NEW-1 ≤ 30 s p95 budget, expand the deployed dependency / attack surface, and create accidental-selection risk (a misconfigured runtime accidentally booting a non-deployment-default strategy). A single binary with all strategies present is also harder to reason about for the IT-12 comparative study, which deliberately wants the opposite — every strategy present and replayed against the same footage.

This decision is made on technical grounds only. Component licenses (BSD/Apache/MIT/LGPL/GPL/etc.) do not influence which strategy is the deployment-default — that choice is the IT-12 measured-performance verdict on the project's operating context (Jetson Orin Nano Super + ADTi 20MP 20L V1 + Derkachi-class footage).

Decision:

1. Per-component CMake BUILD_* flag controls whether each implementation is linked into a given binary (BUILD_VINS_MONO, BUILD_SALAD, etc.). The default deployment binary links the production-default strategy (OKVIS2 on C1 today, pending IT-12 verdict) plus the engine-rule-mandatory simple-baseline (KltRansac on C1). The research binary links every available strategy of every component for IT-12.
2. The Strategy interface boundary makes the exclusion architectural rather than configurational: sibling components import only the Strategy interface, never a concrete implementation. The composition root (one per binary, see ADR-009) is the only place that names concrete classes, and a class whose file is not part of the CMake target cannot be named there — so a misconfigured deployment cannot accidentally pull in an unintended strategy.
3. Selection at startup (config-driven; ADR-001) picks among the linked-in strategies. A binary with only OKVIS2 + KltRansac linked exposes only those two values for vio.strategy; the config validator fails fast if asked for vins_mono.
4. CI emits both binaries on every PR (deployment + research) so the comparative-study artifact is always reproducible alongside the deployable artifact.

Alternatives considered:

1. Single binary with all strategies linked, runtime config picks one — rejected on binary size + boot time + accidental-selection risk + unnecessary dependency surface on the deployed device.
2. Process-isolation IPC for the unused strategies — rejected on latency budget conflict (D-CROSS-LATENCY-1) and operational complexity of two-process deployments on a 25 W edge device.
3. Multiple deployment-binary variants tailored to specific customer bundles — out of scope of this ADR; supported as a consequence (see Consequences NOTE) but not a driver of the decision.

Consequences:

• Two CI binaries on every PR; both must build and test green.
• Adding any new strategy to a component is a folder-add + a CMake BUILD_* flag + an entry in the relevant binary's composition root. No call-site changes anywhere.
• The deployment binary's SBOM is what it is — a consequence of which BUILD_* flags were ON, not a driver of which flags should be ON.
• NOTE — packaging optionality (deferred / non-binding). Because the exclusion is per-implementation per-CMake-flag, the same code base can produce additional binaries — for different deployment targets, different customer bundles, or different end-product licensing bundles if and when product licensing is decided later. This architecture deliberately makes no licensing decisions today: component licenses do not influence which strategy is the deployment-default, and the decision above is purely technical. When packaging strategy is finalized, the same BUILD_* flag mechanism produces the right bundle without source-level changes — that optionality is a side benefit of the interface-first design (Principle #13 + ADR-009), not a justification for it.

ADR-003 — Honest 6×6 covariance via GTSAM Marginals is the safety floor (D-C5-5 = (c))

Context: AC-NEW-4 and the cross-FC covariance honesty (IT-10) require a single, mathematically-recoverable 6×6 posterior covariance per emitted frame. ESKF-style Jacobian-based covariance is faster but loses information across the C4–C5 boundary.

Decision: C5 is GTSAM iSAM2 + CombinedImuFactor + BetweenFactorPose3 + GenericProjectionFactorCal3DS2, with Marginals.marginalCovariance(pose_key) recovering the 6×6 posterior. C4 is OpenCV solvePnPRansac wrapped in a GTSAM factor so C4 and C5 share the same substrate. D-CROSS-LATENCY-1 hybrid auto-degrades C4 covariance to Jacobian-based (D-C4-2 = (a)) under thermal throttle, but C5 stays on Marginals.

Alternatives considered:

1. ESKF-only with Jacobian covariance — rejected: loses cross-component covariance honesty; engine-rule mandatory simple-baseline only.
2. Dual estimators (ESKF + iSAM2) — rejected: memory + complexity + the hybrid auto-degrade already covers thermal stress.

Consequences: GTSAM is a hard runtime dependency; AC-4.5 internal smoothing is for free; per-frame covariance recovery costs 30–90 ms in steady state (auto-degrades to 5–15 ms under thermal throttle).

ADR-004 — Process-level isolation for in-air upload prevention (AC-8.4 enforcement)

Context: AC-8.4 forbids in-air outbound writes to satellite-provider for drone-security reasons. The companion is also read-only against satellite-provider while airborne — there is no operational reason to fetch tiles in flight either, since the pre-flight cache is the contract. A software guard checking flight_state == ON_GROUND can be bypassed by code injection if the network I/O code path is ever loaded.

Decision: The Tile Manager (C11) is a separate binary / image that runs only on the operator's workstation; the airborne companion image does not contain the C11 binary at all — neither the TileDownloader (pre-flight) nor the TileUploader (post-landing) code paths can be reached from the airborne process. The defense-in-depth software guard is owned by C12's PostLandingUploadOrchestrator, which reads the flight_footer FDR record's clean_shutdown field before invoking C11's TileUploader (Batch 44 SRP refactor — the gate's single source of truth is the FDR footer C13 writes only on clean shutdown; C11 itself no longer gates). The local mid-flight tile format is byte-identical to satellite-provider's on-disk layout so no transformation is needed at upload time.

Why the gate moved to C12 (Batch 44): An earlier iteration placed the gate inside C11's TileUploader (consuming a live FlightStateSignal from C8). That duplicated the safety invariant on both sides of the C11/C12 boundary and coupled C11 to C8 just for the post-landing check. The current design (a) consolidates ownership on the operator-side workflow head (C12) — single responsibility per component, single source of truth for "vehicle is fully stopped" (= C13's footer write decision), and (b) collapses an arbitrary 30-second hold-down heuristic to an exact boolean (clean_shutdown). The TileUploader Protocol contract is frozen at v2.0.0 with the gate parameters removed; AZ-317 is superseded.

Enforcement gates (per R02 risk register):

1. CI SBOM diff: the build pipeline fails the airborne production-binary artifact if any symbol from c11_tilemanager/ (or any module that transitively imports c11_tilemanager) appears in the linked image. This is an extension of the per-implementation SBOM enforcement already in ADR-002.
2. Runtime self-check in runtime_root.py: at startup, before opening the FC adapter, the airborne composition root attempts importlib.util.find_spec("c11_tilemanager") and panics if the spec resolves to anything other than None. Cost: one import lookup at startup; benefit: catches a build-system regression even if SBOM diff was bypassed.
3. Network egress test (NFT-SEC-02): the airborne process is run inside a network namespace with no route to satellite-provider's host; any attempted outbound TCP connection to it is a release-blocking test failure.

Alternatives considered:

1. Single binary with software-only guard — rejected on principle: a runtime guard cannot be the primary control for an "is the system airborne?" safety property.
2. Hardware-level switch (e.g., physical write-enable jumper) — rejected: adds operations cost; software-image-isolation gives equivalent assurance for this threat model.

Consequences: Two binaries to maintain (companion image + operator-orchestrator image). CI builds and tests both. The operator workflow has an explicit post-landing step ("run the upload tool") which is itself a feature, not a bug.

ADR-005 — Two execution tiers (Tier-1 / Tier-2) are first-class architectural concerns (F6)

Context: AC-4.1 latency, AC-4.2 memory, AC-NEW-1 cold-start, AC-NEW-3 FDR storage, AC-NEW-5 thermal envelope, and AC-NEW-7 cache-poisoning all have validation locations on Jetson hardware that cannot be replicated on a workstation. Conversely, most logic, integration, and contract tests run in seconds on Tier-1 and would take orders of magnitude longer on Tier-2.

Decision: Tier-1 = workstation Docker (fast/cheap; runs lint + unit + most integration + Mock satellite-provider); Tier-2 = Jetson hardware (AC-bound jobs only; runs NFT-PERF-* + NFT-LIM-* + NFT-RES-* + IT-12). Both tiers are documented in the deployment plan and the CI matrix; failure on either tier is release-blocking. Tier-2 runner availability is itself a risk-register entry.

Alternatives considered:

1. Tier-2-only — rejected: order-of-magnitude slower iteration loop; runner-availability risk dominates.
2. Tier-1-only — rejected: AC-bound NFTs cannot pass without Jetson hardware in the loop.

Consequences: CI is split; some tests have an explicit tier: 2 annotation in tests/environment.md; release artifacts include both tier results.

ADR-006 — D-CROSS-LATENCY-1 hybrid is the AC-4.1 budget strategy

Context: At +50 °C ambient (AC-NEW-5 upper-temp), the Jetson auto-throttles, collapsing the steady-state K=3 latency budget. AC-4.1 has no thermal carve-out — the 400 ms p95 must hold across the operating envelope.

Decision: K=3 baseline (DISK+LightGlue × 3 candidates from C2.5; GTSAM Marginals 6×6 covariance recovery in C4) auto-degrades to K=2 + Jacobian-based covariance under thermal throttle. The trigger is the Jetson's thermal-throttle telemetry crossing a configurable temperature/clock threshold (set per D-C7-9 JetPack 6.2 + TensorRT 10.3 lock). NFT-9 hot-soak validates the hybrid.

Alternatives considered:

1. K=3 fixed + larger latency budget — rejected: AC-4.1 is the contract.
2. K=2 always — rejected: ~5–10 % accuracy loss at steady state hurts AC-NEW-4 headroom.

Consequences: ~5–10 % accuracy loss at the upper thermal envelope (still inside AC-NEW-4). The hybrid is part of the runtime, not a config knob; the threshold is.

ADR-007 — mock-suite-sat-service is an e2e-test fixture, not a first-class component (REVERSED 2026-05-09)

Context: D-PROJ-2 (parent-suite ingest endpoint + voting layer) is not yet implemented. NFT-SEC-01 / FT-P-17 / IT runs need a counterparty for the post-landing upload contract. An earlier iteration of this ADR promoted the mock to a first-class component boundary peer of satellite-provider, with its own description under components/ and its own deployable image — to make the contract auditable.

Decision (current): the mock is an e2e-test fixture only, scoped under tests/fixtures/mock-suite-sat-service/. The architectural counterparty for both the existing download path and the planned D-PROJ-2 upload path is the real satellite-provider. The contract sketch lives in _docs/_process_leftovers/2026-05-09_satellite-provider-design-tasks.md (the source of truth for the parent-suite work) and is mirrored in C11 Tile Manager's external API section (the onboard consumer's view). The mock implements that contract in tests; production never reaches it.

Why reversed: promoting an e2e-test fixture to a component boundary inflated the architectural surface and risked the test fixture drifting away from the real contract once D-PROJ-2 lands. The contract sketch in the leftover file is sufficient as the auditable source of truth without a separate component spec.

Alternatives considered:

1. Keep ADR-007 as originally written — rejected: see "Why reversed".
2. Wait for D-PROJ-2 service-side implementation before any tests — rejected: blocks the onboard cycle.

Consequences: The mock continues to ship in the operator-orchestrator tarball's compose file as a test-time service, but it is no longer documented under _docs/02_document/components/. Test specs and CI references treat it as a fixture. When satellite-provider ships the real endpoint, the fixture is replaced by pointing tests at the real service; no architectural changes flow from that switch.

ADR-008 — D-C8-2 source-set switch is Selected with runtime gate (Mode B Fact #111)

Context: AC-NEW-2 requires spoofing-promotion latency < 3 s. The companion-driven MAV_CMD_SET_EKF_SOURCE_SET switch (D-C8-2 = (b)) is firmware-supported but has no production-deployed precedent — the project would establish the canonical pattern.

Decision: D-C8-2 = (b) is selected with a runtime gate: ArduPilot Plane SITL validation (IT-3) is the lock gate. If IT-3 fails, D-C8-2-FALLBACK options are recorded — (a) operator-manual RC aux switch with relaxed AC-NEW-2 wording; (b) operator-warning STATUSTEXT instead of automated switch; (c) escalate to ArduPilot dev community.

Alternatives considered: see D-C8-2-FALLBACK above.

Consequences: AC-NEW-2 contractual latency is contingent on IT-3 passing. If IT-3 fails, AC-NEW-2 wording is renegotiated as part of D-C8-2-FALLBACK = (a).

ADR-009 — Interface-first components, constructor injection, one folder per component

Context: The architecture deliberately requires multiple interchangeable implementations per component (three VioStrategy for C1; UltraVPR / MegaLoc / MixVPR / SelaVPR / EigenPlaces / NetVLAD / SALAD candidates for C2; pymavlink-AP and YAMSPy-iNav adapters for C8). ADR-002 further mandates that the same logical component ship in different concrete forms across binaries (deployment binary vs IT-12 research binary; future packaging variants if/when needed). Without a strict interface boundary, sibling components import each other's concrete classes; build-time exclusion via BUILD_* flags becomes a fragile compile-time afterthought rather than an architectural property; testing each strategy in isolation requires monkey-patching; and adding a new strategy ripples into every call site. The interface-first pattern is the architectural mechanism that makes ADR-001 (runtime selection) and ADR-002 (build-time exclusion) tractable simultaneously.

Decision:

1. Interface first. Every component is specified as a Python Protocol (or abc.ABC, when concrete defaults are useful) before any concrete implementation is written. The interface is the contract; concrete implementations satisfy it. Step 3 component specs document the interface signature; concrete implementations are documented under their own header inside the component spec.

2. One folder per component. Source layout (per coderule.mdc "place source code under src/"):

   src/
     components/
       c1_vio/
         __init__.py
         interface.py                 # VioStrategy Protocol + VioOutput, VioConfig DTOs
         okvis2_strategy.py           # deployment-default (pending IT-12 verdict)
         vins_mono_strategy.py        # research-only; behind BUILD_VINS_MONO (ADR-002)
         klt_ransac_strategy.py       # engine-rule-mandatory simple-baseline
         tests/
       c2_vpr/
         __init__.py
         interface.py                 # VprStrategy Protocol
         ultra_vpr.py                 # deployment-default (Documentary Lead PRIMARY)
         mega_loc.py
         mix_vpr.py                   # mandatory simple-baseline alternate
         sela_vpr.py
         eigen_places.py
         net_vlad.py                  # mandatory simple-baseline classical
         salad.py                     # additional candidate; behind BUILD_SALAD (ADR-002)
         tests/
       c2_5_rerank/
         interface.py                 # ReRankStrategy
         inlier_count_rerank.py
         tests/
       c3_matcher/
         interface.py                 # CrossDomainMatcher
         disk_lightglue.py
         aliked_lightglue.py
         xfeat.py
         tests/
       c3_5_adhop/
         interface.py                 # ConditionalRefiner
         adhop_refiner.py
         passthrough_refiner.py       # for non-conditional baseline
         tests/
       c4_pose/
         interface.py                 # PoseEstimator
         opencv_gtsam_estimator.py
         tests/
       c5_state/
         interface.py                 # StateEstimator
         gtsam_isam2_estimator.py
         eskf_estimator.py            # mandatory simple-baseline
         tests/
       c6_tile_cache/
         interface.py                 # TileStore + TileMetadataStore + DescriptorIndex
         postgres_filesystem_store.py
         faiss_descriptor_index.py
         tests/
       c7_inference/
         interface.py                 # InferenceRuntime
         tensorrt_runtime.py
         onnx_trt_ep_runtime.py
         pytorch_fp16_runtime.py
         tests/
       c8_fc_adapter/
         interface.py                 # FcAdapter (in+out), GcsAdapter
         pymavlink_ardupilot_adapter.py
         msp2_inav_adapter.py
         qgc_telemetry_adapter.py
         tests/
      c10_cache_provisioning/
        interface.py                 # CacheProvisioner, ManifestVerifier
        provisioner.py
        tests/
      c11_tilemanager/                 # SEPARATE BINARY — never linked into airborne image
        interface.py                 # TileDownloader, TileUploader (two interfaces in one component)
        http_tile_downloader.py
        http_tile_uploader.py
        tests/
       c13_fdr/
         interface.py                 # FdrWriter
         file_fdr_writer.py
         tests/
     composition/
      runtime_root.py                  # composition root: config -> concrete graph
      tilemanager_root.py              # composition root for the C11 operator-side tool (download + upload)
      research_root.py                 # composition root for the research/dev binary
   

3. Constructor injection only. Every component class declares its collaborators as typed __init__ arguments, against the sibling's interface (not the concrete class). Example sketch:

   # src/components/c4_pose/interface.py
   from typing import Protocol
   class PoseEstimator(Protocol):
       def estimate(self, match: MatchResult, calibration: CameraCalibration) -> PoseEstimate: ...
   
   # src/components/c5_state/gtsam_isam2_estimator.py
   class GtsamIsam2StateEstimator:
       def __init__(
           self,
           *,
           pose_estimator: PoseEstimator,           # interface, not concrete
           imu_source: ImuSource,                   # interface
           fdr: FdrWriter,                          # interface
           config: StateEstimatorConfig,
       ) -> None:
           self._pose = pose_estimator
           self._imu = imu_source
           self._fdr = fdr
           self._cfg = config
   

4. Composition root (src/composition/runtime_root.py) is the only place that knows about concrete classes. It reads config, picks each concrete implementation, validates that every named implementation is actually linked into the active binary (fails fast otherwise), and wires the graph. Every other module sees only interfaces. Build-time exclusion (ADR-002) becomes architectural, not configurational: the deployment binary's composition root literally cannot wire VinsMonoVioStrategy because that file is not linked into the deployment binary (BUILD_VINS_MONO=OFF). Future packaging variants (e.g., a customer bundle with a different VprStrategy set) work the same way — a different BUILD_* flag combination + the same composition root code.

5. Python DI mechanism: hand-rolled constructor injection in the composition root is the default — it has no extra dependency, is trivially understandable, and matches the pattern of "select once at startup, never hot-swap". A heavier DI library (dependency-injector, injector, punq) is only introduced if the composition root grows past ~150 lines or test-side wiring becomes repetitive; that is a Plan-phase deferred decision (carryforward), not a current architectural commitment. Mocking in tests is via simple stub classes that satisfy the same Protocol — no monkey-patching, no unittest.mock.patch.

6. Test wiring: each component's tests/ folder owns the test composition for that component. Test composition roots wire the unit-under-test against in-memory / fake implementations of every interface dependency. Cross-component integration tests (Tier-1) compose multiple real components with a fake FcAdapter + fake TileStore + fake InferenceRuntime. End-to-end Tier-2 tests run against the real composition root.

Alternatives considered:

1. Sibling concrete imports (from c5_state.gtsam_isam2 import GtsamIsam2StateEstimator) — rejected: makes ADR-002 build-time exclusion a CMake / SBOM artifact rather than an architectural property; couples C4 to a specific C5 implementation and vice versa; defeats the per-component test wiring; ripples into every call site whenever a new strategy is added.
2. Service locator / global registry (e.g., a process-wide DI singleton accessed via get_service(VioStrategy)) — rejected: hides the dependency graph from constructors, makes test isolation harder, and re-introduces the singletons banned in coderule.mdc.
3. Function-based DI (passing factories instead of instances) — rejected as the default: more cognitive overhead than constructor injection for a startup-bound, never-hot-swapped runtime. Reserved for the few call sites where lazy construction is genuinely required (e.g., the per-flight MAVLink signing key generator).
4. Heavy DI framework (dependency-injector, injector, punq) from day one — rejected as default: introduces a runtime dependency for a problem the composition root can solve in plain Python; reserved as an opt-in if the composition root outgrows hand-rolled wiring.

Consequences:

• Step 3 component decomposition produces, for every component: an interface.py description + ≥ 1 concrete implementation description + a test composition.
• The composition root is itself a reviewable artifact (a single Python file per binary track) that documents which concrete implementations a given binary contains.
• Build-time exclusion (ADR-002) becomes architectural: the deployment composition root cannot import a strategy whose file is not part of the deployment binary's CMake target. The same property scales to any future packaging variant — including, if/when product licensing strategy is decided, license-driven bundles (Principle #13 NOTE), without any source-level change in application code.
• Per-component folders give each implementation a natural home for its own tests/, fixtures, and adapter-specific helpers — matching coderule.mdc's "logic specific to a platform, variant, or environment belongs in the class that owns that variant".
• Adding a new C2 VPR backbone (e.g., a future foundation-model retrieval backbone via D-C2-12) is a folder-add + interface-conformance change; no other component is touched.

Cross-Component Contract Surface (AZ-507)

The ADR-009 "interface, not concrete" rule has an architectural sibling: cross-component imports go through _types/*.py (DTOs + typed-error envelopes such as _types.inference_errors), never through components.X (Public API). The only exception is runtime_root/* (the composition root), which is allowed to import concrete strategies across components precisely because it is the single place that resolves Protocol parameters to concrete classes. Every other module under components/**/*.py consumes cross-component contracts via (a) shared DTOs in _types/*, and (b) consumer-side structural Protocol cuts defined locally inside the consuming component (e.g. c10_provisioning.engine_compiler.CompileEngineCallable for the narrow compile_engine surface of the C7 InferenceRuntime). This is the same architectural property as constructor-injection-against-interface, applied to the import graph rather than the call graph. The AZ-270 test_az270_compose_root.test_ac6_only_compose_root_imports_concrete_strategies lint enforces this on every components/**/*.py; AZ-507 reconciles module-layout.md with the lint so the documentation and the build gate agree.

ADR-010 — Operator-planned mission is the cold-start trust anchor; FC GPS is secondary

Context: The original cold-start design (AZ-419 / FT-P-11) assumed the FC EKF's last valid GPS fix is available at takeoff to seed C5. Field reality contradicts this: a UAV operating in a contested-EW environment may have GPS jammed before takeoff (the jamming radius reaches the launch site, the unit launches under a jammer's umbrella, etc.). In that case the FC EKF has no GPS fix to give, and the companion has nothing to anchor the initial pose to — the entire downstream pipeline (VIO bootstrap, VPR retrieval scope, satellite anchoring) collapses or runs blind. At the same time, the parent suite already requires the operator to author a route in the Mission Planner UI (suite/ui) and persist it to the flights REST service (suite/flights) before any flight runs. The waypoint ordering is operationally meaningful: waypoint[0] is the planned takeoff point. The operator therefore already declares the takeoff position with operationally relevant accuracy (typically a few tens of metres) hours before launch, in a context that has no dependency on GPS at all. This information is the natural cold-start trust anchor.

Decision:

1. Flight is read pre-flight, not in-flight. C12 (the operator-side tool, separate binary from the airborne companion — per ADR-002) calls the parent-suite flights REST service via a typed client (AZ-489 FlightsApiClient) when the operator runs gps-denied-cli build-cache --flight-id <Guid>. An offline path (--flight-file <path>) reads the same DTO shape from a JSON export so the workflow survives operator workstations that have no path to the flights service. The companion binary never depends on the flights service at runtime (Principle #9 — denied-environment operation).
2. C12 derives bbox + takeoff origin from the Flight. The bbox is the envelope of waypoint lat/lon plus a configurable buffer (default 1 km, AZ-489 AC-3). The takeoff origin is Flight.waypoints[0].(lat, lon, alt) — the operator's authored launch point.
3. Both fields are baked into the C10 Manifest. BuildRequest and Manifest carry takeoff_origin: LatLonAlt | None (AZ-323 / AZ-325 / AZ-324 amendments). The hash that drives D-C10-1 idempotence includes takeoff_origin, so a re-plan of the route produces a new cache identity and the verifier (AZ-324) rejects a mismatched cache at boot.
4. C5 consumes the origin before any sensor sample. The companion's composition root reads takeoff_origin from the cache manifest at boot and invokes set_takeoff_origin(origin, sigma_horiz_m, sigma_vert_m) on the active StateEstimator (AZ-490) before the first add_vio / add_fc_imu call. Both GtsamIsam2StateEstimator and EskfStateEstimator accept the origin as a Bayesian prior — iSAM2 attaches a PriorFactorPose3 at Pose3.Identity() (the operator origin BECOMES the local-ENU (0,0,0) anchor) with diagonal sigmas [5°, 5°, 5°, sigma_horiz_m, sigma_horiz_m, sigma_vert_m]; ESKF seeds the nominal position to (0,0,0) and writes the position block of the error covariance to diag(sigma_horiz_m², sigma_horiz_m², sigma_vert_m²). Defaults are sigma_horiz_m = 5.0 m, sigma_vert_m = 10.0 m from C5StateConfig.
5. FC GPS is a secondary, gated input. If the FC EKF later produces a GPS reading (in-flight or at takeoff), it is fused through the existing add_pose_anchor machinery only after passing the three-part gate of Principle #11 — including the ≤ 200 m bounded-delta check against the companion's last emitted PoseEstimate. Real GPS that passes the gate is one more measurement, never an override.
6. Failure modes. If the Manifest has no takeoff_origin AND the FC EKF has no usable GPS at takeoff, C5 stays in INITIALIZING and the FC adapter (C8) emits a non-fused source label; the FT-P-11 takeoff-abort policy (AZ-419 amended) applies. If the Manifest has takeoff_origin AND the FC EKF GPS is wildly inconsistent with it at takeoff (e.g., > 200 m), the operator origin wins and the FC GPS is logged as suspect — this is the GPS-spoofed-at-takeoff case and is the entire point of this ADR.

Alternatives considered:

1. Keep FC EKF as primary (status quo of AZ-419) — rejected: cannot survive GPS-denied takeoff, which is in scope per Principles #1 and #9. Field reports of pre-launch jamming make this a realistic, not edge-case, failure mode.
2. Operator types the origin into a CLI prompt at build-cache time — rejected: duplicates information the Mission Planner UI already captures, drifts from the canonical route, and breaks if the operator re-plans without re-typing. The Flight DTO is the single source of truth.
3. Pull Flight from the companion at runtime over a back-channel — rejected: violates Principle #9 (denied-environment operation; no egress from the companion to anything other than the FC). The flights service is an operator-workstation concern only.
4. Treat operator origin as a hard assignment instead of a prior — rejected: a hard assignment cannot be fused with a later high-quality posterior, breaks ADR-003's "honest covariance" property, and prevents the add_pose_anchor fusion path from ever correcting the origin if it was authored with imprecision.

Consequences:

• AZ-419 (FT-P-11) is amended: the primary cold-start path is operator-origin-from-manifest; FC-EKF-GPS is the fallback path with its own sub-AC.
• C10 contracts gain a takeoff_origin field in BuildRequest, Manifest, and the verifier's validation set (AZ-323 / AZ-325 / AZ-324). Contract version bumps to v1.1.0.
• C5 gains a set_takeoff_origin(origin, sigma_horiz_m, sigma_vert_m) method on the StateEstimator protocol (AZ-490). Protocol contract version bumps to v1.1.0.
• C12 gains the FlightsApiClient boundary + offline --flight-file path (AZ-489).
• Principle #11 (the spoofed-GPS gate) is extended with the bounded-delta clause; the gate now serves both takeoff and mid-flight.
• The companion binary's network surface is unchanged — only C12 (operator-side, separate binary) talks to the flights service.

ADR-011 — Replay is a configuration of the airborne binary, not a separate image (REVERSES the v1.0.0 four-binary design)

Context: The original Decompose Step 2 design for epic AZ-265 (E-DEMO-REPLAY) treated replay as a fourth Docker image (gps-denied-replay-cli) built from the same source tree with a different BUILD_* flag combination — specifically BUILD_C6=OFF, BUILD_C10=OFF, BUILD_C11=OFF, BUILD_C12=OFF, plus the new replay-only build flags ON. The justification was the same as ADR-002 for the live/research/operator split: minimize binary size, attack surface, and accidental-selection risk. An SBOM-diff CI step was specified (AZ-403) to enforce the exclusion of the four "off" components from the replay binary.

Two facts surfaced during the Step 7 (Implement) batch loop that contradicted this design:

1. The C2 (VPR) → C6 dependency cannot be honestly removed. C2 retrieves candidate tiles by querying the C6 DescriptorIndex (FAISS HNSW over pre-built per-tile descriptors). With C6 absent the index has no host, and C2's VprStrategy.lookup(c1) either returns empty (replay produces no positioning fixes, defeating epic AC-3 of ≤ 100 m for ≥ 80 % of ticks) or has to be backed by a parallel "lite" index variant (which is not the production code path and therefore destroys the epic's premise that demo confidence equals field-test confidence on the same footage). Either way the v1.0.0 design's BUILD_C6=OFF flag for replay conflicts with the v1.0.0 epic AC-3.
2. The user requirement is the opposite of binary isolation. Replay's purpose is "demo confidence equals field-test confidence on the same footage" — i.e., the demo and the real flight should run exactly the same code path. Reducing the binary's component set (even one with a sound technical justification like ADR-002) actively works against that purpose: any divergence between the replay image and the airborne image becomes a potential source of demo↔field drift that no SBOM diff can detect once the two binaries' source trees evolve independently.

Decision:

1. Replay is a configuration of the airborne binary. The airborne Docker image is the replay image. No fourth Docker image, no SBOM-diff CI step, no BUILD_C6=OFF for replay. The operator runs the same image with the same gps-denied-onboard entry point (or its sibling gps-denied-replay console-script wrapper) — only the config differs.
2. The mode-aware decision is config.mode = "live" | "replay" resolved once at startup in compose_root. The composition root branch (the single point of mode awareness in the codebase) swaps three strategies and adds one observer:
   • FrameSource: LiveCameraFrameSource ↔ VideoFileFrameSource.
   • FcAdapter: PymavlinkArdupilotAdapter / Msp2InavAdapter ↔ TlogReplayFcAdapter.
   • MavlinkTransport: SerialMavlinkTransport ↔ NoopMavlinkTransport (the outbound bytes go nowhere in replay; the C8 encoder code path is unchanged — see Invariant 5 of the replay protocol).
   • Adds JsonlReplaySink as an additional listener on C5's EstimatorOutput stream (replay-only; the UI consumes the JSONL file). The live binary's downstream sinks (C8 outbound to FC, QGC telemetry adapter, C13 FDR) are unchanged.
3. A new replay_input/ Layer-4 cross-cutting module owns (video, tlog) → (FrameSource, FcAdapter, Clock) convergence. It instantiates the replay strategies, applies the time-offset (manual or auto via AZ-405), and hands the composition root a ReplayInputBundle. The composition root sees no if mode == "replay" plumbing — it sees standard FrameSource + FcAdapter + Clock instances. This is the architectural mechanism that delivers Principle #13's interface-first promise for the replay-vs-live boundary.
4. Operator pre-flight workflow is identical between replay and live. The operator plans a route in the parent-suite Mission Planner UI (suite/ui); the route persists in the flights REST service; C12 reads the Flight, derives the bbox + takeoff origin, calls C11 TileDownloader against satellite-provider, builds the C10 cache (descriptor index + engines + manifest). The only step that differs is "go fly" → "run gps-denied-replay against video + tlog". The companion image consumes the cache identically in both modes (Invariant 12 of the replay protocol).
5. MAVLink emit destinations in replay are no-op sinks for non-UI consumers. The C8 outbound encoders (GPS_INPUT, GCS STATUSTEXT, NAMED_VALUE_FLOAT, MAV_CMD_SET_EKF_SOURCE_SET) run unchanged; their byte streams hit NoopMavlinkTransport and disappear. The user-confirmed design intent: the only position output the UI cares about in replay is the per-tick C5 EstimatorOutput, which is captured by JsonlReplaySink and tailed by the parent-suite UI. MAVLink signing key is mandatory in both modes (Invariant 11 of the replay protocol — the operator supplies a dummy key file for replay; the signing handshake runs and its bytes are dropped by the noop transport).
6. Three binaries, not four. The active build matrix returns to the ADR-002 cadence: airborne (Tier-1 + Tier-2 production; live + replay both run from this image), research (IT-12 comparative-study, mirrors airborne plus the additional VioStrategy / VprStrategy variants), operator-orchestrator (pre-flight workflows on operator workstation). The replay-cli column is removed from module-layout.md's Build-Time Exclusion Map; the replay-only BUILD_* flags (BUILD_VIDEO_FILE_FRAME_SOURCE, BUILD_TLOG_REPLAY_ADAPTER, BUILD_REPLAY_SINK_JSONL) are ON in airborne and research, OFF in operator-orchestrator.

Alternatives considered:

1. Keep the fourth gps-denied-replay-cli binary with BUILD_C6=OFF (status quo of v1.0.0) — rejected for the two reasons in the Context section: the C2→C6 dependency makes BUILD_C6=OFF incompatible with epic AC-3, and the very purpose of replay (demo↔field fidelity) is undermined by any source-tree divergence the SBOM-diff step cannot detect.
2. Keep the fourth binary but with BUILD_C6=ON — rejected: same code as airborne minus C10/C11/C12, which is exactly what airborne already is (the airborne binary already excludes C10/C11/C12 per ADR-002 / ADR-004). The fourth binary would be byte-identical to the airborne image; maintaining it as a separate CI artifact adds work for zero gain.
3. Make replay an HTTP service rather than a CLI — rejected as out-of-scope for this ADR (the parent-suite UI subprocess + JSONL tail design predates this decision and is not in scope here). The replay CLI / live entry-point split is a CLI shape concern, not an architectural concern; the airborne binary remains a long-lived process with no HTTP listener.
4. Move the JSONL sink to a different output (e.g., piped into stdout, or a unix socket) — deferred. The current results.jsonl file output is the simplest UI-tailable contract and matches the parent-suite UI's subprocess assumption. If the UI later needs streaming-without-disk, the sink Protocol allows a StdoutReplaySink or UnixSocketReplaySink strategy without any change to the composition root.

Consequences:

• _docs/02_document/contracts/replay/replay_protocol.md is at v2.0.0 (replaces v1.0.0). New invariants 5, 11, 12 codify the encoder-mode-agnosticism, the signing-key mandate, and the real-C6-cache-in-replay properties.
• module-layout.md Build-Time Exclusion Map drops the Replay-cli column; airborne column gains BUILD_VIDEO_FILE_FRAME_SOURCE=ON, BUILD_TLOG_REPLAY_ADAPTER=ON, BUILD_REPLAY_SINK_JSONL=ON. The narrative reduces "Four binaries…" to "Three binaries…".
• module-layout.md Cross-Cutting section gains a replay_input/ entry (Layer-4 coordinator, owned by AZ-405).
• AZ-403 (replay-cli Dockerfile + SBOM diff CI step) is cancelled; its task file moves to done/ with a cancellation banner pointing at this ADR. Its dependency edges (incoming from AZ-404, outgoing to nothing) are removed from _docs/02_tasks/_dependencies_table.md. The Jira ticket transition to "Cancelled" is recorded in _docs/_process_leftovers/ if the tracker MCP is unavailable at execution time.
• AZ-401 shrinks: it no longer authors a separate compose_replay function; it extends compose_root with the config.mode == "replay" branch and wires JsonlReplaySink + NoopMavlinkTransport. Complexity drops from 3 → 2 points.
• AZ-402 shrinks: it is a thin mode-config wrapper that dispatches into the live entry point, not a standalone CLI.
• AZ-405 grows slightly: it now also owns the replay_input/ coordinator (the natural home for the auto-sync logic + the time-offset application).
• AZ-404 (E2E replay test) is unchanged in scope but reworded: it asserts mode-agnosticism (Invariant 1) and runs against the unified airborne image — no fourth-image entrypoint to verify.
• C8 gains a thin MavlinkTransport Protocol seam introduced by AZ-400: SerialMavlinkTransport (live) and NoopMavlinkTransport (replay) implement it. This is a no-op restructure of the existing C8 transport code; the encoders are unchanged. The Protocol seam is the architectural mechanism for Invariant 5 (encoders are byte-identical).
• Demo↔field fidelity is now structurally guaranteed: the same binary runs in both contexts; any drift between them is a behavioural-test failure, not an SBOM-diff failure.