Move >30 s CI suites to label-gated jobs (ci:slow:*)

[offdev]

What this changes

Measured on the latest master run 38312355303 (2026-10-12, all 11 jobs green), every ctest entry that exceeded 30 s in a lane has been moved out of the main job into a dedicated job of its own (one job per suite per lane). The new jobs run only when gated:

• PR lane (ci-pull.yml): each slow suite has a repo label, created via the GitHub API:

  label	entry	lanes
  ci:slow:sim	laige-sim_tests	gcc clang asan tsan
  ci:slow:render	laige-render_tests	gcc clang asan tsan
  ci:slow:particle	particle_render	gcc clang tsan
  ci:slow:zero-alloc	zero_alloc	asan tsan
  ci:slow:ecs-stress	ecs_stress	asan tsan
  ci:slow:bench-50k	laige_bench_render_50k_fpx16 / _fp32	gcc clang

• Merge lane (ci.yml): the same suites are gated by workflow_dispatch checkboxes (slow-sim ... slow-bench-50k); plain pushes to master never run them.

The main jobs' ctest excludes the moved entries (-E) and the slow jobs select theirs (-R): every entry runs exactly once per gated run, never twice, never ungated.

Per-lane evidence (run 38312355303, seconds):

entry	gcc	clang	asan	tsan
laige_bench_render_50k_fpx16	577.05	578.42	–	–
laige_bench_render_50k_fp32	593.82	566.52	–	–
laige-sim_tests	59.97	53.79	124.03	159.52
laige-render_tests	53.83	44.45	43.20	60.41
particle_render	38.76	34.67	28.84 (stays)	39.56
zero_alloc	16.16 (stays)	14.39 (stays)	43.76	54.75
ecs_stress	16.34 (stays)	14.45 (stays)	42.87	54.70

No entry exceeds 30 s on windows/macOS (worst: laige-sim_tests 27.11 s on macos-intel), so those jobs keep the full suite and are unchanged.

Timeouts: the main Linux jobs revert to 10 min (the gcc/clang 45-min and tsan 30-min caps existed only for the moved suites); the slow suite jobs carry 20 min, the bench jobs 30 min.

Milestone gate (AGENTS.md §5, new subsection): a milestone-gate PR MUST carry ALL six ci:slow:* labels so every slow job runs and passes before merge; after the gate merges, a workflow_dispatch on master with all six checkboxes re-runs the slow matrix. Note: a label added after a run is created never re-triggers the run, so the labels must be on the PR before the run starts.

Docs updated in the same change: AGENTS.md §5, PRD §14 Perf row, roadmap change log. No engine code change — only the two workflows and the three docs.

CI validation (GitHub runs a PR's workflow from the merge commit, i.e. the PR's changes applied on base, so this PR's own runs validate the new PR-lane workflow):

• Run 38367352602 (no ci:slow:* labels): all 11 fast jobs green, all 19 slow jobs correctly skipped.
• Run 38369658151 (all six ci:slow:* labels, added only for this gate validation and removed afterwards): all 30 jobs green. The slow suite jobs finished in 1:46 to 5:23 (worst: TSan laige-sim_tests 5:23) and the bench jobs in 6:24 to 11:34, each well inside the 20/30-min caps. The first pass of this run had one unrelated failure: the unchanged macOS arm64 job hit its 20-min shared-runner cap (tests 1-120 of 123 passed, the job was cancelled while test 121 was starting); the rerun of that job passed in 16:08.