ci: gate workspaces on the unbatched multi-start search check

[Jammy2211]

Turns on check_search_memory.py (added in #228) for every workspace that calls the reusable navigator workflow. For PyAutoLabs/PyAutoFit#1452.

Why this is the part that actually protects CI

The runtime guard in PyAutoLabs/PyAutoFit#1453 reads XLA's memory analysis, which reports 0 bytes on a CPU-only JAX build — which is what CI runs. So it helps GPU users and would not have caught the failure that cost two nightly release-integrate runs. A static AST check has no such blind spot: no JAX, no device, no compile.

It also catches the defect earlier than any runtime guard can — at PR time, before the script ever reaches the nightly.

Why it was held back from #228

Run against the real trees, the checker found eight further unguarded sites beyond the two already fixed. Turning the gate on in the same PR would have red-lined three workspaces over unrelated work. All eight now carry an explicit batch_size:

• docs: correct the VRAM batch-size guidance + make batch_size explicit autolens_workspace#469 (3 sites)
• docs: make the multi-start batch_size decision explicit autogalaxy_workspace#205 (4 sites)
• docs: make the multi-start batch_size decision explicit autofit_workspace#132 (1 site)

Merge this last, after those three.

Shape

Its own job rather than a step on the paths job, so the check name is legible in a PR's status list. Stdlib-only (ast + pathlib), so it needs no pip install and cannot be broken by dependency resolution.

The rule is "batch_size must be explicit", not "must be small" — batch_size=None passes. The point is that the memory decision is typed out where a reviewer sees it rather than arrived at by defaulting, which is precisely how the original failure happened.

Verification

YAML parses; the job's four steps mirror the existing paths job. The checker itself carries 15 unit tests (added in #228) and reports all six workspaces clean once the three PRs above land.

---

Generated by Claude Code