Remove duplicate chunk discovery paths

[asinghvi17]

Task E2, the whole of Phase 4 of regrid-notes/2026-08-21-regridding-simplification-plan.md. Fifth in the stack, on top of #74 (which is on #72, on #71, on #70). Record: regrid-notes/2026-08-23-e2-delete-legacy-discovery.md.

Phase 4's gate: one query implementation defines graph edges, and neither the executor nor the interface translates that relation back through a compatibility tree. Both halves are tests, in test_chunkgraph.jl, testset one query implementation defines every edge.

Deleted, each after a repo-wide audit

• connectedchunks / connectedchunks! — all four methods. Not exported, not public, and after E1 no caller outside test_lazy.jl. The two srcindex in-place forms were the card's "compatibility source-index state": a prebuilt source index carried around outside any relation and re-queried.
• The whole chunktree bridge — the export, the declaration, the chunktree-collecting chunkextents fallback with _collectextents!, RasterFlatTree and its fifteen trait methods, and chunktree(::RasterGrid).

The chunktree verdict: remove

Nothing implemented it but the bridge itself, a TileCells forwarder, and three test toys. DGGSpace never had one — it specializes chunkextents/chunkindex/candidatechunks! directly. DGG's private _cachedchunktree is an unrelated name (it caches a HierarchicalGridCursor's caps) and is untouched, as is the local variable named chunktree in crosssystem/regrid.jl. ext/ has no reference and docs/ does not build GlobalRegridding.

So the only thing that still needed chunktree was chunktree: a space packed its caps into a flat tree so the generic chunkextents could walk it and collect the same caps straight back out. chunkextents is now a required hook with no fallback, asserted as !hasmethod(chunkextents, Tuple{RegridSpace}). RasterFlatTree went with it, closing the note at rastergrid.jl:969-971.

Kept, because the audit says they are not dead

• chunkextents — many real consumers that want the caps as values: spacestamp, _builddependencies (both sides), subspace_dependencies, the generic chunkindex, and three production scripts.
• chunkextent — after connectedchunks went it had zero callers and zero specializations, but it is public, documented, and asserted by the qualified-contract testset. The card says keep a cheap one, and the generic chunkextents(space)[chunk] was not cheap, so RasterGrid now specializes it at O(1) in the chunk count.
• The gates harness's archived :latjoin arm — there is no latitude join left in src/ at all (G2 deleted it). What remains carries an explicit sunset condition tied to G2's waiver, which has not been retired; deleting it would have destroyed both the waiver's audit trail and half of this PR's own before/after evidence.
• The generic candidatechunks! STI fallback — unreached in-repo after the above, but it is one method of the one query function rather than a second relation, and it is what makes the documented chunkindex promise true.
• Nothing to do on "post-plan graph builder" (G4 removed it; api.jl documents that chunk_dependency_graph has no plan method) or "redundant cap vectors" (E1 moved the graph's two onto the relation on purpose; in production they alias the two DGGSpaces' own arrays).

Gate: the relation is identical

benchmark/chunk_graph_gates.jl at the branch point and on this branch. 13 cases including the production copdem90-igeo7-l12 pair from a local tile list, both arms, 26 ndjson rows each, -t 8 --gcthreads=4, 5 samples.

Every relation field identical on all 26 rows — edges, demanded_pairs, demand_missing, oracle_pairs, oracle_missing, only_here, missing_here, destination_chunks, source_chunks, radius, identity_bytes, graph_allocated_bytes, graph_summarysize_bytes. graph_allocated_bytes byte-identical on every row is the direct evidence the bridge was never on the build path. Both runs print verdict: PASS on 9 oracle-checked case(s); 4 case(s) unchecked — a real geometric verdict, not NOT CHECKED. Production :indexed +1.8 %, inside the band G3/G4/E1 all recorded on unchanged code.

Values and residency: byte-identical

	before (0e3fb70)	after
LazyStats	loads 328, hits 3654, peak 2 073 600 B	identical
source readblock! calls	400	400
repeated read / slices / retiled plan	true	true
max abs(eager − lazy)	2.44249e-15	2.44249e-15
live: plan / array	82 072 B / 43 376 B	82 072 B / 43 376 B

Suites

suite	before	after	delta
lib/GlobalRegridding/test	3 999 / 1 broken / 0 fail	4 021 / 1 / 0	+22 pass
crosssystem/regrid.jl	236	236	none
crosssystem/regrid_acceptance.jl	22	22	none
scripts/copdem_policy.jl	89	89	none
scripts/copdem_source_mode.jl	7	7	none
CopernicusDEM/runtests.jl	16 258 / 3 broken / 0 fail	16 258 / 3 / 0	none

No count dropped. Every assertion that named a deleted function was rewritten onto the relation rather than removed. The +22 is itemized assertion by assertion in the record: 19 in the new Phase 4 gate testset, 2 in discovery, 1 in the wave-failure testset.

🤖 Generated with Claude Code

https://claude.ai/code/session_019DuGKTmvs5B4EynwZddKMg