docs: restructure DESIGN.md + surface the cross-domain examples

[emfdavid]

Docs-only. Cuts DESIGN.md from 839 lines of sprawl to a structure where each kind of open work has exactly one home, and moves measurements/diagrams to the pages that own them.

What moved

• M1.6 V2 scheduler: 175 → 84 lines. Kept the design; deleted the Phasing / B1 task list blocks (finished sprint planning — git history has them); moved the exp_c acceptance tables to docs/benchmarks.md.
• Roadmap: 333 → ~130 lines. Every milestone is intent + state + a pointer. M-W's design rationale moved up into the geometry ladder where readers look for it; its Phase-4 example plan is dropped (examples/advection/ is the record).
• M-GCS obstore-vs-gcsfs numbers → docs/benchmarks.md; event-loop-ownership diagram → docs/architecture.md (it's architecture, not roadmap).

What's new

• Known limitations & defects — the section that didn't exist, which is why several items only lived in session notes: window_factor sizes residency by span rather than the offset set; the silent edge-anchor drop; whole-pipeline (not per-variable) cache invalidation; windowed+shuffle residency; the accepted ~3.5% refcount cost; and the inherited zarr-python sharded-read cliff.
• Upstream capabilities we plan to inherit — the design dependency on zarr-python lazy indexing (#3906) and batched zero-copy out= decode (#3060/#2904), including what stays ours either way. Deliberately records the dependency, not PR status — that state goes stale in-repo.
• zarrs ZarrsCodecPipeline spike. Attributes the cost to zarr-python's pipeline rather than numcodecs (the uncompressed row isolates it — ~1.9× with no codec invoked), names the allocate-then-copy interface as the real target, ties it to #3060 as the alternative attack on the same copy, and notes our upside is well under the 3.5× headline since we already call the pipeline directly. Gated on it serving our direct-bytes path and pinning its Rust thread pool to 1.

Accuracy fixes found in review

• JAX/TF surfaces shipped in M3 — the thesis still said "planned".
• The Earth2Studio obstore store-swap is now labelled an unmeasured hypothesis with a mechanism, not a result. NVIDIA is doing the migration themselves (Migrate GFS byte-range GRIB fetches to obstore with shared helpers NVIDIA/earth2studio#964, GFS byte-range GRIB, HRRR/GEFS/CFS to follow), but neither those PRs nor we have published a cold-cache A/B.

Verification

uv run mkdocs build --strict passes. Note DESIGN.md is linked from the nav as an external GitHub URL, so the strict build does not validate it — the rendered diff here is the check.

---

Second commit — surface the cross-domain examples on the site

The site had one reference to the WeatherBench2 walkthrough and a single inline mention of microscopy; hubble/ and the three-framework advection story were invisible to anyone who never opened the repo.

• New docs/examples.md, in the nav directly under Home — a table of the four showcases (domain, store, the geometry each exercises, what it proves), run commands, and the offline synthetic mode for each. Full flags stay in examples/README.md, which it links rather than duplicates.
• Home page gains a Cross-domain examples table in headline position, and its domain-general paragraph now names weather, microscopy and astronomy instead of only microscopy.
• Replaced the two scattered one-off example links with pointers to the new page.
• Corrected the home-page status line: windowed multi-offset sampling and the arbitrary sample axis shipped, and a cross-chunk slab read is an explicit non-goal rather than an unbuilt feature.

examples/sdss is not on main yet, so it is not listed — it can land with its own PR.

mkdocs build --strict passes and every internal anchor was verified against the built HTML.