feat!: reintegrate cpm-geozarr into main - #277

Fixes the confirmed findings from a multi-agent adversarial review:

- write() now runs every argument-derivable check (target type, remote
  scheme, w- existence, pipeline routing, required 'groups') BEFORE the
  destructive mode="w" removal of an existing store. Previously a pure
  options mistake (e.g. missing groups) deleted the user's previous
  product and then failed with nothing written.
- accept compute=True (rejecting only compute=False): CPM's staged-output
  path (staging.py) injects compute=True when a remote Dask client is
  active, so the strict kwargs rejection broke the exact --stage-output
  flow the writer's own error message recommends.
- validate_write_options now applies the target checks (type, remote
  scheme), so CPM's triggering preflight rejects unsupported targets at
  parse time instead of at write time.
- routing: only S02MSIL1C/S02MSIL2A products go to the S2-optimized
  pipeline; other MSI types (S02MSIL0_, S02MSIRAW) lack the reflectance
  pyramid structure and now route generic.
- eopf_geozarr.cpm raises a clear ImportError explaining the Python
  >= 3.13 floor when eopf is missing (pip install 'eopf-geozarr[cpm]'
  silently installs nothing on 3.12).
- _validate_s2_input: cover both branches with tests; also skip
  validation for zarr-v3-backed trees (the S2 input model reads Zarr v2
  only and raised TypeError on CPM 3.x products).

Verified on Python 3.13 with eopf 3.0.0: 42 tests pass, including new
regression tests for the destroy-then-fail ordering and compute
handling.

Assisted-by: ClaudeCode:claude-fable-5

- test-cpm job: installs the cpm extra (eopf pinned via uv.lock) on
  Python 3.13 and runs the CPM writer/routing tests on every push/PR,
  so they no longer silently skip in CI.
- test-cpm-latest job: weekly scheduled run against the newest published
  eopf release to catch upstream contract drift (writer registry,
  staging, CLI entry-point loading) independently of PR failures.
- new test: invoke the real `eopf convert-geozarr` console script via
  subprocess on a synthetic S2 store, covering entry-point discovery,
  click option wiring, and convert()'s reader dispatch through to the
  geozarr engine.

Assisted-by: ClaudeCode:claude-fable-5

PR #212 (Sentinel-3 OLCI GeoZarr export) merged to main after #238 was
branched, leaving select_pipeline unable to reach it: OLCI products fell
into the generic pipeline and failed on the missing 'groups' option.

Adds S3_OLCI_PRODUCT_TYPE_PREFIXES and looks_like_sentinel3_olci to
routing.py, structural fallback included -- oa01_radiance is a data
variable of the measurements node, not a child node, unlike the S2
r{N}m pyramid groups. Wires convert_olci_optimized into the writer with
a new s3_olci_optimized flag and an output_grid option, kept additive
to the existing write() signature (s2_optimized is untouched) since
that's the EOWriter entry point other CPM code calls into.

* ci: run CI on the cpm-geozarr branch

The workflow only triggered on main, so no work on the cpm-geozarr
feature branch ever got checks -- including #247, which is 20 commits
and +29105/-3763 across 23 files. The Samples Service runs this branch
in its September GeoZarr production track, so changes reached a
production pipeline unchecked.

Closes #253

* style: wrap the resolved_spatial_chunk conditional to satisfy ruff format

Introduced by the OLCI routing commit, which renamed the local
`pipeline` variable to `selected_pipeline` and pushed the line past the
limit. It went unnoticed because the branch had no CI -- which is what
this PR fixes.

* init project setup

* adding and harmonising the chunking/sharding functionality across measuremetn and non-measurement groups - allows the contuniation of chunking behaviour from input eopf datatree instead if overwriting by force - scale_offset logic not touched

* addign nan masking for coarsening operation for multiscale calualtation (only for s2_optimising, later extend!) response to issue #246

* addign rechunking after readin to circumvent chunking at read-in (which is hardwired to eopf/store/mappers)

* adding L1C injected bands for mutliscale calculation -> with rechukning of added bands to spatial_chunk size

* readding prior launch script

* implementing changes recommended by claxn and d-v-b, mainly docs, cleanup, simplification

* adressing change requests, improving chunk workflow

* simplifing encoding retrieval for (all and) 1d arrays

* hotfix: adding stac metadata to root level was previously not inlcuded -> copying root metadata to new tree root now

* adding docstrings

* working on the tests, frustrated with test layout

* adding zarr v3 test samples

* change of the data_api tests

* removing test structure which is not adhering to s2 specs, also 'notebook' scripts referenced in tests didnt seem to exist

* changing the s2 zarr v3 structure and all associated tests. Also removing all emojis from output messages

* adapting to new zarr layout, changed function calls, and multiscale calculation changes

* removing emojis

* adding checks to add metadata to arrays if keep_scale_offset=False, modifying checks for s2 structure to align with tests, addig docs

* fixing enum to strenum to allow detection of S2type in simple == comparison

* applying changes to validate pre-commit pyright issues

* addressign requested chagnes for stAc metadata addition adn fill_value fallbacks, removing 'experimental_scale_offset_codec'functionality for encoding2

* small issue of not rechuking in tests

* removing addition of multiscales to stac asset metadata, other small commetns

* adding _FillValue to encoding to allow xarray decode_cf to correctly read zarr store in abscence of eopf-xarray reader, removed large commented section

---------

Co-authored-by: Emmanuel Mathot <[email protected]>

…ng pr247 (#267)

* adding single /measurement/refectance group to stac as requested during pr247

* adressing comments for crs reuse and proj_attrs_for_crs use

* adding test to incorporate new functionalities and expanding comments

Brings in S1 GRD RTC ingestion (#216), release 0.11.0 (#235) and dependency
bumps (#268, #269). Conflicts: tests/conftest.py keeps the cpm-geozarr v3 S2
fixtures and adds main's S1 RTC fixtures; uv.lock re-locked from main's lock.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Brings in the spatial:bbox pixel-edge fix (#275, fixes #266).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

L1C/L2A detection read dt_input.name, which is None for trees opened from
a Zarr store, so no band was injected: L2A lost b08 at r20m-r720m and L1C
kept only b01/b09/b10 from r60m. Detect the level from stac_discovery
product:type (set by every CPM S2 mapping), with the name as fallback.

inject_missing_bands also updated the source band's attrs in place, so
native r10m bands sometimes carried the injected bands' provenance attrs,
depending on write timing. Copy the attrs instead.

Restore the golden snapshot test, run on the Zarr v2 inputs, and
regenerate the snapshots for this branch's output
(REGENERATE_SNAPSHOTS=1 rewrites them).

Co-Authored-By: Claude Opus 5.5 <[email protected]>