Adopt morton decimal shard ids + hive-partitioned output layout (manifest, commit stamp)

[espg]

🤖 from Claude

Combined layout migration (per espg's call: one issue, phased — both phases change S3 key shapes, so landing them together is one wire-format break instead of two). Design record: docs/design/sparse_coverage.md §2–§3 (D1–D6, D12), tracked on #198; convention spec frozen mortie-side (espg/mortie#48 → espg/mortie#62).

Blocked by espg/mortie#104 — needs the decimal emit / scalar / hive_path surface, released to PyPI (we pin mortie>=0.8.5; this will bump).

Phase 1 — decimal shard-id strings (existing flat layout)

Shard ids surface externally as raw packed-word integers today; route them through the decimal repr:

• CSR subgroup paths: {group_path}/{field}/{shard_key}/… (processing/write.py:583-598)
• Async status objects: <store>.status/<run>/{shard_key}.json (runner.py:1403)
• Worker/dispatcher logging (worker.py:208 etc.)
• --morton-cell / catalog stringified-key round-trips (runner.py:899-926)

Hard break, no migration shims (pre-1.0; existing test stores become debris).

Phase 2 — hive tree behind a config flag

• {store_root}/{sign+base}/{d1}/…/{full_id}.zarr — one digit per level to shard order, full-id leaf (D2/D3).
• morton_hive.json written once at template time: spec: morton-hive/1, dataset identity, cell/shard orders, split schedule, pyramid declaration (D6).
• Commit stamp: the shard's final write is a root group.attrs.update(...) (completion flag + cell count + timestamp + granule count). A .zarr/ prefix without the stamp is debris, safe to overwrite on retry (D4). This is not consolidated metadata — one tiny PUT on an object that exists anyway.
• Node invariant enforced in the writer: below root, only [1-4]/ children and *.zarr objects (D5 — zero metadata above the leaf; see Sharded CSR writes issue a per-array metadata-op storm — upstream ask: a no-probe / batch array-creation path in zarr-python #189/Parallelize the opt-in metadata-consolidation path (deferred from #191/PR #193) #194 for why).
• Tests: walker termination (no digit children ⇒ nothing finer), debris/retry semantics, torn-write simulation.

Acceptance

• Write-throughput validation under real concurrency (the CONUS burst estimate that motivated this — pairs with Shuffle shard dispatch order before fan-out to spread S3 prefix load #197's shuffle): a benchmark or scaled run demonstrating spread across prefixes, as a checklist item before close.
• User docs for the layout in the same PR (repo convention).

Related: #197 (dispatch shuffle, orthogonal), #198 (design doc), follow-on MOC write issue (blocked by this).

[espg]

🤖 from Claude

Scope adjustment per espg: the write-throughput validation acceptance item in the body above moves to #202 (the benchmarks issue — full-AOI NEON runs + CONUS cost model), where it belongs with the measurement machinery. This PR's acceptance stays: tests for walker termination, debris/retry semantics, torn-write simulation, and layout docs in the same PR.

[espg]

mortie 0.9.0 released with espg/mortie#104 fixes, this is unblocked.

Please mark #200 as unblocked after all implementation phases here are finished