perf: remove algorithmic compiler and layout cliffs

[alixander]

Uses github.com/d2lang/dagro v0.1.1, released from d2lang/dagro#2.

What

• add ordered IR field/edge indexes while preserving safe public mutation semantics
• replace global lazy-glob rescans with a deterministic incremental worklist
• cache cloned import AST and context-safe IR templates
• resolve D2 Dagre endpoints and parallel-edge counts once, cache container topology, and batch independent nested extraction
• optimize Dagro's deep-compound ordering and layer-graph construction
• replace ELK adapter/postprocessing rescans with graph indexes, subtree intervals, and an exact mutation-aware spatial broad phase
• reuse layout geometry for style-only scenarios and steps under an explicit stable-resolver opt-in
• document the performance work in the next changelog

Why

Removing the JavaScript runtimes exposed several older source-compatible algorithms that repeatedly scanned or copied the entire graph/IR. Large globs, repeated imports, nested containers, style-only boards, and bend-heavy layouts could therefore grow quadratically or worse.

Impact

Apple M4, one CPU, synthetic stress fixtures unless noted:

• leading glob, 1,000 fields: 1.59 s → 12 ms (~133×)
• repeated 100-field import, 100 uses: 54 ms → 12 ms (~4.6×)
• 4,000 flat fields: 67 ms → 7 ms (~9.6×)
• 3,200 distinct edges: 200 ms → 19 ms (~10.6×)
• 32-node style-only diagram with 8 scenarios + 8 steps: 40 ms → 7.3 ms (~5.5×), ~81% fewer allocated bytes
• nested D2 Dagre stress: ~32% faster
• dense ELK 250-node/1,000-edge stress: ~30% faster
• Dagro depth-100 compound stress: 3.60 s → 0.90 s (~4×), 73% fewer allocated bytes

These are targeted/pathological benchmarks rather than a corpus-weighted claim; ordinary diagrams primarily benefit by avoiding the specific cliffs.

Compatibility

• exact D2 E2E SVG/layout corpus remains unchanged with reuse enabled
• geometry reuse defaults off for arbitrary public d2lib layout resolvers; CLI/WASM opt in only for stable built-in resolution
• sequence, grid, and constant-near layouts conservatively bypass reuse
• public IR lookup/mutation APIs remain linear and safe for direct exported-slice mutation; only compiler-private paths use derived indexes
• ELK keeps the original exact collision predicates and mutation order, with brute-force fallback for unsafe/huge coordinates
• Dagro passes the upstream Dagre 0.8.5 differential corpus
• IR fixture changes only remove duplicate internal source-reference records; compiled semantics and rendered output are unchanged

Validation

• go test ./e2etests -run '^TestE2E$' -count=1 -parallel=1
• go test -race ./... (all packages except three baseline-confirmed GIF CLI race/timing failures; exact E2E passes under race)
• go vet ./...
• GOOS=js GOARCH=wasm go build -trimpath ./d2js
• focused wasm32 ELK compile
• go mod tidy -diff
• go mod verify
• independent compiler and layout correctness reviews