perf: two quadratic passes in model creation, and three smaller costs

[bvdmitri]

What

Five non-breaking fixes to model creation. The results are unchanged: the existing tests pass untouched.

change	where	measured
apply_meta! walks the context's own factor nodes, not every node of the model filtered by membership in the context	plugins/meta/meta_engine.jl	it was quadratic, 84% of creation with @algorithm: nonlinear SSM at n = 3 000, 4.51 → 0.47 s
while constraints are applied, flattened_index of a matrix variable uses prefix sums cached per array (with_flattened_index_cache, cached_flattened_index), bound around apply_constraints!	resizable_array.jl, variational_constraints*.jl	matrix-variable constraints were quadratic (19 s at 10⁵ elements): 3.07 → 0.81 s at 4·10⁴
NodeLabel == compares the Int counter first	graph_engine.jl	30 → 19 ns per lookup
the factorisation bitsets start as the full joint, built directly, and a node with no factorisation skips the partition step	variational_constraints*.jl	graph stage 10–13% faster on an HMM and an iid model
ConstraintStack keeps a Vector, read from the top as the Stack was	variational_constraints_engine.jl	DataStructures.Stack allocated a 1 024-element block per model: 12–29% of a tiny model's infer

These were measured as part of ReactiveMP's end-of-refactor performance pass (Julia 1.13, M-series Mac). The cache uses task_local_storage as a dynamic scope, since ScopedValue is Base only from Julia 1.11 and GraphPPL supports 1.10.

Tests

• New: cached_flattened_index agrees with flattened_index on a ragged 3-D array, with and without the cache, and on a 1-D one; ConstraintStack reads from the top, keeps its per-context counts and pops per context (it gains length and eltype).
• make test: 76 878 pass, 1 broken as before.
• Not reformatted: make lint also reports 12 files on 4.9.0 itself, and formatting the touched files would rewrite unrelated code.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Re8DzeAcZpjeENCE44QJjF

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.28%. Comparing base (849fd25) to head (0a3ec0a).

Additional details and impacted files

@@            Coverage Diff             @@
##            4.9.0     #333      +/-   ##
==========================================
+ Coverage   91.04%   91.28%   +0.23%     
==========================================
  Files          16       16              
  Lines        2301     2318      +17     
==========================================
+ Hits         2095     2116      +21     
+ Misses        206      202       -4     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.