README: per-backend feature tier in the feature matrix

[brendancol]

Closes #2465. Part of #2415.

The README feature matrix now shows the feature tier for each function on each backend instead of a bare availability mark.

• ✅ stable · 🔼 advanced · 🧪 experimental · 🔧 internal · 🚫 unsupported
• Blank still means no implementation on that backend.
• Cells that used to be 🔄 (CPU fallback) are now 🔼 advanced — a fallback is a documented execution mode, not native parity.

Tiers are a first-pass, evidence-based classification drawn from each module's implementation and tests, using the conservative rubric in #2415: stable needs a correctness baseline plus cross-backend parity evidence for that backend; the NumPy path can be stable while Dask/GPU paths are still advanced; a passing unit test alone doesn't justify stable. The CRS support and datum sub-tables under Reproject are untouched.

When the machine-readable registry from #2415 lands, the table can be generated from it rather than maintained by hand.

Backend coverage

Documentation only — no code changes, no test changes.

Test plan

• CRS support sub-table and datum sub-table unchanged
• No stray 🔄 in any feature-matrix row
• Legend updated to describe the five tiers
• Section anchors and column headers unchanged

[brendancol]

PR Review: README per-backend feature tier

Documentation-only change; correctness, backend-dispatch, and test items are N/A.

Blockers

• (none)

Suggestions

• README.md:538 — Pycnophylactic. The interpolate/MCDA audit found the Dask path explicitly raises NotImplementedError (iterative Laplacian smoothing needs global zone sums per step). By the tier definitions in #2415, an explicit-error combination is 🚫 unsupported, not blank. Worth flipping the Dask and Dask+CuPy cells from blank to 🚫 so the table communicates the deliberate negative answer.
• README.md:503-504, 511 — AHP Weights, Rank Weights, Sensitivity (Dask / CuPy / Dask+CuPy). Audit reports these as pure-numpy with no dask/cupy code paths. Marked 🔼 advanced on the assumption of silent CPU fallback. If a Dask or CuPy DataArray input actually errors, 🚫 unsupported is more honest. Quick check on a cupy-backed DataArray would confirm the failure mode before locking these in.
• README.md:280 — Min Observable Height. Row carries 🧪 and still has *(experimental)* in the description. The textual marker is redundant once the tier emoji says the same thing; drop the suffix.

Nits

• README.md:144 — legend separator. Five-tier line uses   separators while the section TOC just below uses ·. Using · in both would be a small consistency polish.
• Stable claims on Dask/GPU columns. Several sections (multispectral, classification, morphology, focal core, hydrology core) ended up with stable marks across all four backends. #2415 is explicit that stable on Dask/GPU requires backend-specific parity evidence, not just a green unit test. A follow-up pass downgrading Dask/CuPy cells to 🔼 where per-backend parity isn't explicitly tested would be safer. Reasonable to land this as the first-pass best guess and tighten against the machine-readable registry from #2415 later.

What looks good

• CRS support sub-table and datum sub-table left untouched.
• No stray 🔄 anywhere in the feature tables.
• Legend explains both the fallback→advanced mapping and the blank-vs-🚫 distinction.
• Column structure, section anchors, and per-table headers all preserved.

[brendancol]

Follow-up review (after 771882e)

Three suggestions and one nit from the previous pass were applied; one nit was deferred with rationale.

Disposition of original findings

• Pycnophylactic Dask/Dask+CuPy → 🚫 — Fixed. Verified dasymetric.py:502-503 and :708-709 raise NotImplementedError; the docstring (:675) also says so. Now 🚫 on both cells.
• AHP Weights, Rank Weights Dask/CuPy/Dask+CuPy → 🚫 — Fixed. Confirmed weights.py has no dask/cupy code paths; these are pure-numpy matrix/list utilities. Marked 🚫 on all three non-numpy columns since the per-backend matrix's question doesn't really apply to them.
• Sensitivity — Fixed. Confirmed sensitivity.py:131,165 handles Dask via .compute() (kept 🔼). No CuPy path and the algorithm uses numpy.random / scipy.stats, so CuPy and Dask+CuPy are now 🚫.
• Min Observable Height — drop *(experimental)* — Fixed. The 🧪 tier emoji now carries that signal.
• Legend separator nit — Fixed. Switched   to · so the tier legend matches the TOC just below it.
• Stable-on-Dask/GPU downgrade across multispectral / classification / morphology / focal core / hydrology core — Deferred. This is a larger judgement pass touching ~60 rows; the user explicitly framed this PR as a best-guess first pass to pair with #2415's machine-readable registry. Better to tighten those marks against the registry when it lands than to make a second sweeping guess now.

Spot-checks on the new state

• README.md:503-504 AHP / Rank: ✅ 🚫 🚫 🚫
• README.md:511 Sensitivity: ✅ 🔼 🚫 🚫
• README.md:538 Pycnophylactic: ✅ 🚫 🔼 🚫
• README.md:280 Min Observable Height: trailing (experimental) removed
• README.md:144 Legend uses · separators consistent with the TOC

No new findings.

[brendancol]

Follow-up: apply the GeoTIFF tier bar (171ee46)

Maintainer feedback on the first pass: too generous with stable. Agreed.

The classification agents were inferring stable from cross-backend parity tests passing. But the precedent #2415 cites is GeoTIFF, which marks only its NumPy/Dask local-file paths as stable and calls GPU experimental despite extensive tests. Stable requires a documented contract + release-gate testing, not green parity tests.

Stricter rules applied across the matrix:

• NumPy stable only for canonical, literature-backed algorithms with golden-reference / hand-computed tests. Custom / newer / utility / helper paths drop to advanced even on NumPy.
• Dask stable essentially never outside GeoTIFF — chunk-boundary semantics are tested but not formally declared as a contract.
• CuPy / Dask+CuPy stable never — no per-backend tolerance contract anywhere outside GeoTIFF.

Net effect:

tier	before	after
✅ stable	~533	122
🔼 advanced	rest	502
🧪 experimental	18	18
🚫 unsupported	10	10
blank	16	16

stable is now concentrated where it should be: GeoTIFF NumPy/Dask, plus the NumPy paths of canonical literature ops (Aspect, Slope, Hillshade, Curvature, TPI/TRI/Landforms, multispectral indices, classification core, morphology core, focal core, hydrology D8 core, proximity core, fire, flood, MCDA combine core, IDW, A*, Reproject, etc.). Everything else is advanced — works, has tests, just lacks the GeoTIFF-grade contract.

Where the agents disagreed with that bar (e.g. they reported stable for Dask + GPU on multispectral and morphology because parity tests pass), I overrode and downgraded. The contract gate, not the test result, decides stable.