Implement the production-default InferenceRuntime strategy on JetPack
6.2 + TensorRT 10.3 (per D-C7-9). The runtime owns the full TRT
lifecycle: compile_engine via the Polygraphy + trtexec + IBuilderConfig
hybrid (FP16 / INT8 / Mixed precision), deserialize_engine with
EngineGate-first ordering and a pre-allocation GPU memory budget gate,
infer via H2D -> enqueueV3 -> D2H -> stream sync on the owned CUDA
stream, idempotent release_engine, and an injected
ThermalStatePublisher delegation for thermal_state.
INT8 calibration cache trust (D-C10-6, AC-2/3/4) is enforced by a
.calib_cache.sha256 file-integrity sidecar (AZ-280) plus a new
.calib_cache.dataset_sha256 sidecar that records the dataset content
hash at compile time; reuse only when both agree, rebuild silently on
dataset hash mismatch, raise CalibrationCacheError on corrupt sidecar
(never silently overwritten).
GPU memory budget (NFT-LIM-01, default 4 GiB) is checked BEFORE any
TRT call beyond the gate (AC-6); a pre-allocation refusal raises
OutOfMemoryError and leaves the resident state unchanged.
TensorRT 10.3 / Polygraphy / PyCUDA are lazy-imported inside the
methods that need them so the module loads cleanly on Tier-0 hosts.
A standalone CLI entry (python -m
gps_denied_onboard.components.c7_inference.tensorrt_runtime compile
<onnx> <build_config.json>) is wired for C10 CacheProvisioner
(AZ-321) to invoke pre-flight without holding a runtime instance.
C7InferenceConfig gains gpu_memory_budget_bytes (default 4 GiB) and
trtexec_timeout_s (default 600 s, Risk 4 mitigation), both validated
in __post_init__.
Tests: 26 active + 6 Tier-2-gated skips; AC-1 / AC-3 / AC-4 / AC-5
/ AC-6 / AC-7 / AC-10 + NFR-reliability fully covered on Tier-1
via fake CUDA / TRT modules; AC-2 / AC-8 / AC-9 / NFR-perf-deserialize
placeholders skip with prerequisite reason and live in the AZ-298
Tier-2 microbench harness. Code review verdict
PASS_WITH_WARNINGS (1 Medium hot-path hoist fix auto-applied).
Co-authored-by: Cursor <cursoragent@cursor.com>
Land F1+F2+F3 from the PASS_WITH_WARNINGS cumulative review of
batches 28-30 (AZ-305 / AZ-307 / AZ-308) before continuing to
batch 31. All three are bounded by the c6_tile_cache component;
no public API contract change beyond the new error re-export.
F1 (Medium / Architecture):
Re-export CacheBudgetExhaustedError from c6_tile_cache package
__init__ so consumers can catch the AZ-308 budget-exhaustion
variant without widening to TileCacheError (which drops the
needed_bytes / available_bytes / evicted_count diagnostics).
F2 (Medium / Architecture):
Refresh the c6_tile_cache section of module-layout.md so the
Public API line and the Internal-files list reflect what is
actually on disk after batches 28-30 (drop the stale
Tile / TileRecord / connection.py entries; add the AZ-305
postgres_filesystem_store + tools.py, AZ-307 freshness_gate,
AZ-308 cache_budget_enforcer entries; pivot the Public API
bullet to the __init__.__all__ as canonical, mirroring the
c7_inference section format).
F3 (Low / Maintainability):
Promote the triplicate intra-module _iso_ts_now() helper into
a single c6_tile_cache._timestamp.iso_ts_now and import it
from postgres_filesystem_store, freshness_gate, and
cache_budget_enforcer. FDR record envelope ts format now has
one source of truth.
Test impact:
tests/unit/c6_tile_cache: 105 passed, 57 skipped (pre-existing
Docker-compose skip markers). No new tests required for F1/F2
(re-export + doc) and F3 (pure refactor; existing tests assert
FDR record shape, not the helper symbol).
Autodev state advanced to awaiting-invocation; next session
resumes greenfield Step 7 at batch 31 (AZ-298 TensorrtRuntime).
Co-authored-by: Cursor <cursoragent@cursor.com>
Cumulative review of batches 23-27 (cycle 1) surfaced three Medium
documentation-drift findings on module-layout.md. All three fixed
inline per user direction:
F1: c7_inference Internal list expanded with architecture_registry,
config, engine_gate, errors, manifest, thermal_publisher (added
across AZ-300/301/302).
F2: c6_tile_cache `connection.py` re-attributed from AZ-304 (which
deferred it) to AZ-305 with a "planned, not landed yet" tag.
F3: c7_inference Public API description rewritten by category
(Protocol + DTOs + component services + config + error family)
with a pointer to __init__.py's __all__ for the canonical list.
Cumulative review report: _docs/03_implementation/cumulative_review_
batches_23-27_cycle1_report.md (PASS_WITH_WARNINGS).
Autodev state moved to status: paused_user_requested per user
choice; /autodev will resume at greenfield Step 7 (next batch
selection) on next invocation.
Co-authored-by: Cursor <cursoragent@cursor.com>
Strictly additive Alembic migration on the AZ-263 baseline (data_model
.md § 6.1 / § 6.3): six new tiles columns (tile_uuid UNIQUE,
location_hash, content_sha256, disk_bytes, accessed_at, uploaded_at),
four new btree indices, one UNIQUE expression index over the
COALESCE-zero-uuid natural key, CHECK widening of
ck_tiles_freshness_status to the AZ-263 + AZ-303 vocabulary UNION,
four NULLable bbox columns on sector_classifications, and a new
tile_freshness_rules table seeded with the two default thresholds.
Pinned UUIDv5 namespace (TILE_NAMESPACE_UUID =
5b8d0c2e-1a4f-4b3a-8c9d-e7f6a3b2c1d0) + derive_tile_id /
derive_location_hash helpers cross-coordinated with
satellite-provider. Migration runner apply_migrations(config) drives
Alembic command.upgrade("head") against the AZ-263 env with one
retry on PG SQLSTATE 40001 and structured INFO logs on apply / no-op.
Contract bump tile_metadata_store.md v1.1.0 -> v1.2.0 adds
TileMetadata.location_hash: UUID | None = None (non-breaking).
module-layout.md updated so c6_tile_cache explicitly Owns
db/migrations/**.
Tier-1 tests: UUIDv5 determinism + locked vectors + DSN resolution +
retry mocked DBAPIError -> 1180 passed, 32 skipped. Tier-2 docker
schema tests gated by @pytest.mark.docker run against the existing
docker-compose.test.yml db service.
Co-authored-by: Cursor <cursoragent@cursor.com>
Implements the production-default ReRankStrategy: K=10 → N=3 by
single-pair LightGlue inlier count, with strict drop-and-continue
(INV-8) on per-candidate TileFetch / backbone / zero-inlier failures
and RerankAllCandidatesFailedError on zero survivors. Composition
root injects the shared LightGlueRuntime + Clock + the new
FeatureExtractor helper (an L1 placeholder OpenCvOrbExtractor that
unblocks AZ-343 and future C3 strategies — task scope expansion).
Architectural notes:
- Cross-component imports stay banned; tile_store types as `object`
and the C6 TileCacheError family is duck-typed by class module
prefix (same workaround AZ-348 adopted for c7_inference; proper
fix is to relocate TileCacheError to _types/ in a follow-up).
- Clock injection follows the replay contract (AZ-398 Invariant 2);
reranked_at is sourced from clock.monotonic_ns().
- AZ-342 factory grew `feature_extractor` + `clock` + `fdr_client`
parameters; existing AZ-342 conformance tests updated.
Tests: 19 new AC-1..AC-12 + mixed-failure scenarios in
test_inlier_count_reranker.py; existing AZ-342 suite (26) still
green. Full repo sweep 1093 passed / 2 skipped (cmake/actionlint
not on PATH).
Co-authored-by: Cursor <cursoragent@cursor.com>
Defines the public `ConditionalRefiner` Protocol (PEP 544
@runtime_checkable, two methods: `refine_if_needed` +
`was_invoked`), extends `MatchResult` in-place with two
default-valued refinement fields (`refinement_label`,
`refinement_added_latency_ms`), defines the `RefinerError` family
(`RefinerBackboneError`, `RefinerConfigError`), and ships the
trivial `PassthroughRefiner` reference impl.
Both refiner strategies are linked unconditionally — no
`BUILD_REFINER_*` flag (NOT ADR-002 territory). Runtime selection
only per ADR-001. `PassthroughRefiner` returns the input
`MatchResult` by reference (bit-identical correspondences per
contract INV-5) and always reports `was_invoked() is False`.
Documentation: renames `module-layout.md` `c3_5_adhop` Public API
symbol from `AdHoPRefinementStrategy` to `ConditionalRefiner`
(AC-14) so the doc agrees with `description.md` and the contract.
AC-9 (single-thread binding) deferred to AZ-270 runtime-root
composition, mirroring AZ-336 / AZ-342 / AZ-344 Risk-4 precedent.
AC-7 for the `"adhop"` strategy stops at `ModuleNotFoundError`
because the AdHoP backbone is owned by AZ-349. All other ACs +
NFRs covered by 36 new conformance tests.
Architectural note: `PassthroughRefiner.inference_runtime` is
typed as `object` because the L3→L3 import ban
(`test_az270_compose_root`) forbids c3_5_adhop from importing
c7_inference; the runtime-root factory narrows the type at
construction time.
Co-authored-by: Cursor <cursoragent@cursor.com>
Defines the public `CrossDomainMatcher` Protocol (PEP 544
@runtime_checkable, two methods: `match` + `health_snapshot`),
the three frozen+slotted DTOs (`CandidateMatchSet`, `MatchResult`,
`MatcherHealth`) in the L1 `_types/matcher.py` layer, the
`MatcherError` family (`MatcherBackboneError`,
`InsufficientInliersError`), and the composition-root
`build_matcher_strategy` factory with lazy-import +
`BUILD_MATCHER_<variant>` gating per ADR-002.
`RollingHealthWindow` accumulator (60 s, amortised O(1) update,
strict O(1) snapshot) is constructed by the factory and injected
into every concrete matcher so all backbones share window
semantics; this is what backs C5's spoof-promotion gate.
Legacy placeholder `MatchResult` removed from `_types/matching.py`;
import-only consumers (`c4_pose.interface`, `c3_5_adhop.interface`)
repointed at the new `_types/matcher.py` home — zero behavioural
change to those components.
AC-9 (single-thread binding) and AC-10 (LightGlueRuntime
identity-share with C2.5) deferred to AZ-270 runtime-root
composition, mirroring the AZ-342 Risk-4 escape clause. All other
ACs + NFRs covered by 70 new conformance tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
Foundational scaffolding for the InlierCountReRanker (AZ-343) and
the future C3 CrossDomainMatcher consumer (AZ-344). No concrete
re-ranker is implemented here.
* ReRankStrategy Protocol (single rerank(frame, vpr_result, n,
calibration) -> RerankResult method) with all 8 invariants in the
docstring — notably INV-8 drop-and-continue (per-candidate failure
NEVER propagates unless every candidate fails).
* DTOs moved to L1 _types/rerank.py — RerankCandidate, RerankResult;
frozen+slots; tuple-not-list for RerankResult.candidates; tile_id
encoded as (zoom_level, lat, lon) tuple to keep _types/ free of any
c6_tile_cache (L3) import per module-layout.md.
* Error family: RerankError + RerankBackboneError +
RerankAllCandidatesFailedError. Only RerankAllCandidatesFailedError
escapes rerank(); RerankBackboneError is caught inside the per-
candidate loop, logged ERROR, FDR-stamped, candidate dropped.
* C2_5RerankConfig (strategy enum default "inlier_count", top_n int
default 3) with strict validation at load; registered into
Config.components on c2_5_rerank import.
* build_rerank_strategy(config, *, tile_store, lightglue_runtime)
factory: 1-strategy resolution table, lazy import,
BUILD_RERANK_<variant> gate, ImportError → StrategyNotAvailableError
mapping. The shared LightGlueRuntime is constructor-injected
(R14 fix: neither C2.5 nor C3 owns its lifecycle).
Renamed the Protocol from the existing stub "RerankStrategy" to
"ReRankStrategy" to match the contract; updated module-layout.md.
Removed the legacy RerankResult shape from _types/vpr.py — the
v1.0.0 shape lives in _types/rerank.py.
Excluded per task spec:
* Concrete InlierCountReRanker (AZ-343).
* C3 matcher protocol task (AZ-344, next in batch).
* AC-9 single-thread binding + AC-10 LightGlueRuntime identity-share
between C2.5/C3 — deferred per task spec Risk 3 until the generic
compose_root thread-binding registry and the C3 factory both land.
Tests: AC-1..AC-8 + AC-11 + NFR-perf-factory in
tests/unit/c2_5_rerank/test_protocol_conformance.py. The legacy
smoke test is removed. Full sweep: 997 passed (one pre-existing
flake in test_az296_takeoff_abort, subprocess timing, unrelated to
this commit; passes in isolation).
Co-authored-by: Cursor <cursoragent@cursor.com>
- Modified the Docker Compose configuration to include an input root for replay tests and added an environment variable for enabling SITL.
- Enhanced documentation for various testing processes, including the addition of a Runtime Completeness Decomposition Gate and clarifications on internal module testing requirements.
- Updated the implementation completeness report to reflect the current state and added new test cases for performance and resilience scenarios.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep VIO package and native bridge paths backend-neutral so BASALT remains an implementation choice rather than a component boundary.
Co-authored-by: Cursor <cursoragent@cursor.com>
Oleksandr Bezdieniezhnykh