perf: forward-only cursor in monotonicArena.Alloc

[jensneuse]

Summary

monotonicArena.Alloc walked a.buffers from index 0 on every call,
giving O(numBuffers) cost per call and O(N²) total work over the
arena's lifetime.
On the Cosmo Router workload reported in the issue (~180MB JSON
response, ~600-1200 buffers, ~29M Alloc calls per request),
this manifested as ~40s of router-side merge time.

This PR adds a forward-only cursor.
Subsequent walks start at cursor instead of 0.
Cursor advances on a later-buffer hit and on grow.
Reset and Release rewind it to 0 so a reused arena can re-fill
its early buffers from scratch.

For uniform-size allocations the per-call cost becomes O(1).
For mixed sizes the walk is bounded by buffers ahead of the cursor,
with the trade-off that any remaining free space in skipped buffers is
abandoned for the rest of the request.

Benchmarks

Controlled prefix (isolated walk cost):

Prefix	Before	After	Speedup
10	17.5 ns/op	2.7 ns/op	6.5x
100	149 ns/op	2.6 ns/op	57x
1000	1293 ns/op	2.6 ns/op	497x

Pre-fix: clean O(N) scaling.
Post-fix: flat O(1) regardless of prefix size.

Realistic growth workload (AllocCosmoLike):

Prefix	Before	After	Speedup
10	5125 ns/op	3.4 ns/op	1500x
100	5265 ns/op	4.0 ns/op	1300x
1000	2785 ns/op	4.0 ns/op	700x

The realistic workload speedup is larger because the unpatched arena
grows during the timed loop, so the prefix walk gets longer over time.
This is consistent with the reporter's measurement of ~3x end-to-end
on the full Cosmo Router request, where Alloc was the dominant cost.

What changed

• monotonic_arena.go: added cursor int field; Alloc walks from a.cursor; cursor advances on success and grow; Reset and Release rewind.
• monotonic_arena_test.go: 7 unit tests (cursor mechanics + regression test for issue monotonicArena.Alloc scales poorly on large subgraph responses in Cosmo Router #2 hot path) and 2 benchmarks.

Credit

Original analysis and patch by @thoec in #2.
This PR adapts the patch and adds the test/benchmark coverage.

Test plan

• go test -race ./... passes
• All 7 new cursor tests pass
• Benchmarks confirm O(N) → O(1) scaling

Fixes #2
FIxes ENG-9512

[ysmolski]

I have checked it against the engine integrational benchmarks. I don't see any increase in the memory consuption.

ys@work:~/d/cosmo/router-tests/protocol$ benchstat old new
goos: darwin
goarch: arm64
pkg: github.com/wundergraph/cosmo/router-tests/protocol
cpu: Apple M4 Max
                                     │     old      │                  new                  │
                                     │    sec/op    │    sec/op      vs base                │
ConfigHotReload-14                     8.989m ± 61%    8.583m ± 58%        ~ (p=0.529 n=10)
Sequential-14                          154.9µ ± 11%    149.6µ ±  5%        ~ (p=0.218 n=10)
Parallel-14                            17.67µ ±  2%    17.63µ ±  1%        ~ (p=0.353 n=10)
ParallelWithMinify-14                  94.25µ ± 12%   101.89µ ± 14%        ~ (p=0.579 n=10)
ParallelBig-14                         59.34µ ±  4%    62.30µ ±  6%        ~ (p=0.089 n=10)
SequentialBig-14                       726.7µ ± 11%    656.2µ ±  8%        ~ (p=0.105 n=10)
SequentialBigCostControl-14            733.0µ ±  9%    685.6µ ±  9%        ~ (p=0.393 n=10)
SequentialBigNotCached-14              1.301m ±  7%    1.263m ± 13%        ~ (p=0.436 n=10)
SequentialBigNotCachedCostControl-14   1.513m ±  2%    1.275m ±  4%  -15.72% (p=0.000 n=10)
ParallelCostControl-14                 63.73µ ±  3%    58.84µ ±  4%   -7.68% (p=0.000 n=10)
geomean                                313.6µ          301.1µ         -3.98%

                                     │     old      │                 new                 │
                                     │     B/op     │     B/op      vs base               │
ConfigHotReload-14                     39.62Mi ± 0%   39.63Mi ± 0%       ~ (p=0.315 n=10)
Sequential-14                          66.14Ki ± 0%   66.12Ki ± 0%       ~ (p=0.424 n=10)
Parallel-14                            32.77Ki ± 0%   32.85Ki ± 0%       ~ (p=0.052 n=10)
ParallelWithMinify-14                  204.8Ki ± 0%   204.8Ki ± 0%       ~ (p=0.393 n=10)
ParallelBig-14                         104.0Ki ± 1%   106.9Ki ± 2%  +2.73% (p=0.000 n=10)
SequentialBig-14                       460.5Ki ± 0%   460.4Ki ± 0%       ~ (p=0.353 n=10)
SequentialBigCostControl-14            462.0Ki ± 0%   461.9Ki ± 0%       ~ (p=0.699 n=10)
SequentialBigNotCached-14              949.5Ki ± 0%   950.8Ki ± 0%  +0.13% (p=0.001 n=10)
SequentialBigNotCachedCostControl-14   958.7Ki ± 0%   959.5Ki ± 0%  +0.08% (p=0.004 n=10)
ParallelCostControl-14                 104.8Ki ± 1%   106.7Ki ± 2%  +1.80% (p=0.000 n=10)
geomean                                361.4Ki        363.2Ki       +0.49%

                                     │     old     │                 new                  │
                                     │  allocs/op  │  allocs/op   vs base                 │
ConfigHotReload-14                     47.31k ± 0%   47.31k ± 0%       ~ (p=0.894 n=10)
Sequential-14                           908.0 ± 0%    908.0 ± 0%       ~ (p=1.000 n=10) ¹
Parallel-14                             436.0 ± 0%    436.0 ± 0%       ~ (p=0.700 n=10)
ParallelWithMinify-14                  1.976k ± 0%   1.972k ± 0%       ~ (p=0.986 n=10)
ParallelBig-14                         1.383k ± 2%   1.423k ± 2%  +2.86% (p=0.000 n=10)
SequentialBig-14                       6.603k ± 0%   6.603k ± 0%       ~ (p=1.000 n=10)
SequentialBigCostControl-14            6.620k ± 0%   6.620k ± 0%       ~ (p=0.650 n=10)
SequentialBigNotCached-14              11.04k ± 0%   11.06k ± 0%  +0.16% (p=0.000 n=10)
SequentialBigNotCachedCostControl-14   11.17k ± 0%   11.18k ± 0%  +0.16% (p=0.000 n=10)
ParallelCostControl-14                 1.390k ± 1%   1.421k ± 2%  +2.27% (p=0.000 n=10)
geomean                                3.617k        3.636k       +0.52%
¹ all samples are equal

                                     │      old      │                  new                  │
                                     │      B/s      │      B/s       vs base                │
Sequential-14                          493.2Ki ±  9%   512.7Ki ±  5%        ~ (p=0.230 n=10)
Parallel-14                            4.206Mi ±  2%   4.220Mi ±  1%        ~ (p=0.305 n=10)
ParallelWithMinify-14                  40.68Mi ± 11%   37.63Mi ± 16%        ~ (p=0.579 n=10)
ParallelBig-14                         54.40Mi ±  5%   51.85Mi ±  7%        ~ (p=0.089 n=10)
SequentialBig-14                       4.439Mi ± 13%   4.916Mi ±  7%        ~ (p=0.101 n=10)
SequentialBigCostControl-14            4.411Mi ±  9%   4.706Mi ±  8%        ~ (p=0.382 n=10)
SequentialBigNotCached-14              2.484Mi ±  7%   2.556Mi ± 12%        ~ (p=0.424 n=10)
SequentialBigNotCachedCostControl-14   2.131Mi ±  2%   2.532Mi ±  5%  +18.79% (p=0.000 n=10)
ParallelCostControl-14                 50.64Mi ±  3%   54.85Mi ±  3%   +8.32% (p=0.000 n=10)
geomean                                6.593Mi         6.865Mi         +4.12%

[ysmolski]

I have created a big-response benchmark in cosmo to measure how it performs with big responses.
Old:

    iteration_duration.............: avg=7.87s min=7.34s med=7.97s max=8.22s p(90)=8.19s p(95)=8.2s
    iterations.....................: 4      0.126957/s
    data_received..................: 2.4 GB 76 MB/s
    data_sent......................: 5.2 kB 165 B/s

New:

    iteration_duration.............: avg=4.88s min=4.27s med=4.64s max=5.52s p(90)=5.49s p(95)=5.51s
    iterations.....................: 7      0.204882/s
    data_received..................: 4.2 GB 123 MB/s
    data_sent......................: 9.1 kB 267 B/s

Peak Heap Allocated: grew from 8.1 GB to 10.5 GB. But again, when I run prolonged requests of this kind it stabilized both on old and new version of go-arena.

[ysmolski]

bench.js. It did not change picture for small-ish requests.
old:

    EXECUTION
    iteration_duration.............: avg=2.1ms  min=227.2µs med=1.87ms max=13.53ms p(90)=3.54ms p(95)=4.09ms
    iterations.....................: 1023136 20462.102045/s
    vus............................: 99      min=2            max=99
    vus_max........................: 100     min=100          max=100
    NETWORK
    data_received..................: 7.3 GB  147 MB/s
    data_sent......................: 2.8 GB  57 MB/s

new:

    EXECUTION
    iteration_duration.............: avg=2.11ms min=236.04µs med=1.89ms max=11.37ms p(90)=3.56ms p(95)=4.1ms
    iterations.....................: 1016307 20325.487958/s
    vus............................: 99      min=2            max=99
    vus_max........................: 100     min=100          max=100
    NETWORK
    data_received..................: 7.3 GB  146 MB/s
    data_sent......................: 2.8 GB  57 MB/s

Peak Heap Size grew from 134 MB to 138 MB. Means that heap stayed the same