[mgovorcin]

What

Since October 2026, CMR's granules.umm_json results are unreliable for the NISAR GSLC and GUNW collections:

• Records counted in CMR-Hits are left out of the results.
• Under some concept ids, UMM-JSON serves another granule's record. That granule then appears twice and the real one is missing. For example, G4246696294-ASF is NISAR_L2_PR_GUNW_023_055_D_066_... in atom-JSON and from search/concepts/, but UMM-JSON serves the record of ..._023_059_A_038_... under it.

On 2026-10-21, UMM-JSON served the wrong record for 13,769 of 155,910 provisional GSLCs and 12,607 of 89,932 GUNWs, and left out a further 198 and 6,024. The atom-JSON granules.json listing of the same query has every granule under its own name. The concepts endpoint returns the right record; echo10 does not.

This PR:

• Adds fetch_cmr_pages(..., verify=True). It walks the same query in atom-JSON and drops UMM-JSON records whose GranuleUR disagrees with the atom title. It then re-fetches the missing and mislabelled ones: first from UMM-JSON by concept_id[], keeping only matching names, then from search/concepts/<id>. The result holds one record per granule name.
• Adds CmrSearchReport, optional, which records the left-out, mislabelled, duplicate-name and unrecovered concept ids, for reporting them to CMR.
• Retries CMR timeouts, dropped connections, 429 and 5xx with backoff, and gives every request a timeout. Previously requests had none.
• Turns verification on for nisar.search and nisar.search_gunw. The DISP-S1 search is unchanged, because a sample of the newest 20,000 OPERA_L3_DISP-S1_V1 granules came back clean in both formats.

The same fix runs in opera-adt/nisar_db (#63, #65), which publishes the NISAR frame viewer.

Cost

Verification adds one atom-JSON page walk, plus one concepts-endpoint request per bad record. A California GUNW search (1,128 granules, 458 bad) took 138 s.

Verification

• Live: search_gunw over California (-125, 32, -114, 42) returns all 1,128 granules. Without verification, UMM-JSON returned 1,096, and 426 of them carried another granule's record.
• tests/test_cmr.py uses a fake CMR that leaves out, mislabels and duplicates records. It covers recovery, the report, keeping one record per name, retries, and non-retried 400s. The existing NISAR search tests switch the atom-JSON listing off with one fixture, since they check parsing and filtering.
• pytest tests/test_cmr.py tests/test_nisar_search.py tests/test_disp_search.py: 42 passed; pre-commit clean. In my local pixi test environment, 13 unrelated tests fail identically on main without this change. They cover the projection library, stitching, the version check and aiohttp.

🤖 Generated with Claude Code