Fix zero-geometry generation; optimize scoped input preparation (#172, #132)

benchmark/fragmented_roi_profile.py

@@ -0,0 +1,81 @@
+"""Bounded provider-free profile of the declared ROI extraction/ZIP owners."""
+
+from __future__ import annotations
+
+import argparse
+import cProfile
+import json
+import pstats
+from collections import Counter
+from pathlib import Path
+from time import perf_counter
+from zipfile import ZipFile
+
+import numpy as np
+from polystore.disk import DiskStorageBackend
+from polystore.roi import extract_rois_from_labeled_mask
+from polystore.roi_converters import FijiROIConverter, NapariShapeMetadata
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "output_dir", type=Path, help="Owned disposable scratch directory."
+    )
+    args = parser.parse_args()
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    labels = np.zeros((128, 128), dtype=np.int32)
+    labels[4:124:4, 4:124:4] = 188
+    labels[1:3, 1:3] = 7
+    labels[125:127, 125:127] = 42
+    start = perf_counter()
+    extract_rois_from_labeled_mask(labels, min_area=1)
+    cold_extraction_seconds = perf_counter() - start
+    profiler = cProfile.Profile()
+    start = perf_counter()
+    rois = profiler.runcall(extract_rois_from_labeled_mask, labels, min_area=1)
+    warm_extraction_seconds = perf_counter() - start
+    stats = pstats.Stats(profiler)
+    bbox_scan_calls = sum(
+        value[1] for key, value in stats.stats.items() if key[2] == "find_objects"
+    )
+    start = perf_counter()
+    members = FijiROIConverter.rois_to_imagej_members(rois)
+    encoding_seconds = perf_counter() - start
+    archive = args.output_dir / "fragmented.roi.zip"
+    start = perf_counter()
+    DiskStorageBackend().save(rois, archive)
+    archive_seconds = perf_counter() - start
+    with ZipFile(archive) as zipped:
+        member_count = sum(name.endswith(".roi") for name in zipped.namelist())
+    print(
+        json.dumps(
+            {
+                "shape": list(labels.shape),
+                "parent_labels": [
+                    NapariShapeMetadata.from_metadata(roi.metadata).label
+                    for roi in rois
+                ],
+                "fragments_by_parent": dict(
+                    Counter(
+                        NapariShapeMetadata.from_metadata(member.metadata).label
+                        for member in members
+                    )
+                ),
+                "zip_roi_members": member_count,
+                "cold_extraction_seconds": cold_extraction_seconds,
+                "warm_extraction_seconds": warm_extraction_seconds,
+                "bbox_scan_calls": bbox_scan_calls,
+                "member_encoding_seconds": encoding_seconds,
+                "archive_encoding_and_write_seconds": archive_seconds,
+            },
+            sort_keys=True,
+        )
+    )
+    # The archive writer includes its own required conversion; this does not
+    # imply production converts twice. The standalone conversion isolates cost.
+    stats.strip_dirs().sort_stats("cumulative").print_stats(15)
+
+
+if __name__ == "__main__":
+    main()

docs/plans/h002_query_preparation_diagnosis_20260929.md

@@ -0,0 +1,112 @@
+# H002 attempt4 query preparation: timing correction and bounded diagnosis
+
+Owner: Lovelace, original206/PolyStore16 source continuation only.
+No edits to installed imagej-cache-progress source, H002 driver/receipts,
+scientific pipeline or data. No repeated query, raw pixels/notebooks/biological
+outputs, new JVM/runtime/download, install, profiling attach or process restart.
+
+## Evidence and exact timing meaning
+
+Named receipt:
+`/home/ts/wt/openhcs-h002-uncoached-20260929/output/runtime/h002-live-attempt4/receipts/013-openhcs_query_plate_files.json`.
+It records query with limit1, include_previews=false, terminal success.
+No inspection of the underlying container was performed by this worker.
+
+Actual driver source `output/mcp_session_attempt4.py:94` sets
+`started = time.monotonic()` once before stdio-client/session initialization.
+Lines113-115 calculate `elapsed_seconds = time.monotonic() - started` BEFORE
+awaiting session.call_tool. Therefore90.16998135202448 is session age at request
+admission, NOT query duration. Original90s claim is withdrawn; no new90s query
+regression established and no fix justified by that number.
+
+Receipt admission at1790718869.0149193, UTC2026-09-29T21:54:29.014919.
+Result-bearing receipt mtime1790718875.589610, UTC21:54:35.589610.
+Difference **6.574690818786621s**: admission-to-terminal-receipt-write interval,
+including dispatch/serialization/receipt IO; not an exact native stage timer.
+The driver rewrites the same receipt after the successful await. The parent's
+next-receipt96.745 estimate is a separate upper bound including author time.
+The previous cache acceptance explicitly measured0.211s differently; neither
+figure proves identical fixture/warm/cold cost or a general latency reduction.
+
+Both driver1614516 and MCP1614529 were alive when checked, started17:52:57/58
+local time. The MCP stderr fd points at this attempt's mcp-stderr.log.
+Read-only whitelist of actual MCP environ confirms:
+`POLYSTORE_IMAGEJ_CACHE_ROOT=/home/ts/.cache/polystore/imagej` and
+`POLYSTORE_IMAGEJ_ALLOW_DOWNLOAD=false`.
+No credentials/unrelated environment values printed or retained.
+
+Startup stderr reports BigStitcher override from the existing digest-addressed
+shared Fiji jars, then OMETiffReader initialization. These lines contain **no
+timestamps**. Several reader-open lines accumulated across this continuing
+session; they cannot be attributed/count-budgeted to query013 alone. File mtime
+is the most recent write, not each line's event time. Do not assign a precise
+JVM/decoder split from their order or repeated filenames.
+
+## Source-owned preparation boundary
+
+QueryPlateFilesCapability -> PlateInspectionService.query_files/_query_files
+(:997/:1114) -> _create_handler(:1869) -> existing registered detection/handler
+owners -> BioFormatsMetadataHandler.source_dataset(:64) -> registered
+SourcePlaneStoreAdapter -> BioFormatsJavaAdapter._declares_path and
+_discover_container -> BioFormatsJavaContext.ensure_initialized/open_reader
+-> FIJI_IMAGEJ_RUNTIME.initialize -> FijiArchiveDistribution.materialize.
+Limit1 bounds returned records after inventory preparation, not container/JVM
+initialization. include_previews=false prevents requested previews, not metadata
+discovery. No biological behavior is assumed.
+
+Existing ImageJRuntimePolicy.initialize separates materialization from Java:
+with a fresh JVM it first materializes the canonical bundle, then configures
+bundled Java with fetch=never and invokes imagej.init. If Java is already started,
+it reuses Java subject to declared Java/gateway compatibility checks.
+BioFormatsJavaContext additionally imports reader/metadata/format Java classes;
+open_reader builds OME metadata and invokes setId before source-plane projection.
+Its process-owned context/lifecycle lock retains initialized gateway state.
+No alternate Java helper/cache/decoder has been added.
+
+Installed false policy rejects a missing/invalid bundle before cold filesystem
+materialization/network. Actual environment + successful query + stderr pointing
+to the shared bundle are consistent with successful warm bundle admission,
+followed by cold Java/decoder preparation. They do not time either component.
+Bundle provisioning and JVM/decoder preparation are distinct;228 fixed sharing
+and inspection responsiveness, not all cold latency.
+
+Query currently does not declare inspection's progress heartbeat/worker offload,
+and its description claims no detection although _query_files can create/detect
+the handler. This is a source-backed exposure/observability concern, **not a
+newly proved90s regression or measured stage bottleneck**. Keep it within the
+existing capability/preparation ownership for a separately justified checkpoint;
+do not create another progress runner or speculate about bypassing rich metadata.
+
+## Reproduce the diagnosis without calling MCP
+
+Read this single receipt and its actual driver; calculate
+`receipt.stat().st_mtime - json.loads(receipt.read_text())["at"]`.
+Inspect only the actual stderr fd/log and whitelist cache/download keys from the
+existing process. Never replay the request to establish timing. Exact JVM,
+reader-open and source-projection costs are **unavailable retrospectively**.
+Future stage profiling belongs on ImageJRuntimePolicy/BioFormatsJavaContext and
+existing preparation owners, using before/after monotonic events on an authorized
+small synthetic run after the live science slot; do not attach/restart H002.
+
+## Source integration and checks (not live fixes)
+
+Original parent206 was6bc71ee5c, child16 was9e205fee, remote heads matched.
+Normally merged fetched OpenHCS main8b090e and PolyStore main94bf443 into ONLY
+their original isolated continuation branches. Merge heads parentb2edc4503,
+child50c1e49; conflict resolution retains merged17/228 download-policy and launch
+projection changes rather than older cache-only declarations. No rebase/reset/
+force-push or installed-tree mutation. Main's recorded c16b8fc gitlink is retained;
+the original16 worktree intentionally differs while its broader gates remain.
+
+Six scoped source/test files parsed; both diff checks and Ruff on dependency
+distribution/cache tests and parent cold-runtime tests pass. No pytest/runtime
+rerun needed to answer this diagnostic and none performed. NRA/refactor-audit
+coverage is focused source/consumer trace, not a global scan or native proof.
+BOUND-2/6: canonical distribution/runtime/context own config/lifetime; do not
+read/replace them in callers. IMPL-12/13: keep one materializer and Java lifecycle.
+The merge restores current-main behavior, not a performance fix/new owner.
+
+Remaining original206/16 **141 passed/26 failed** and **after-ROI profile not run**
+unchanged. Fresh cold stage profile, source preservation controls and broader
+Java/CZI fidelity remain pending. No performance/preservation/live claim from
+syntax checks or this receipt. Parent's installed228 acceptance remains separate.