Rescue branch bundles unrelated task IDs into three commits (high)

[thedancingdeveloper]

The local feat/mariadb-baseline branch carries three rescue commits spanning multiple task IDs (T21-T25 plus runtime, CI and docs layers).

That makes review, rollback and per-issue acceptance ambiguous: there is no commit whose revert corresponds to one item, and no way to accept T21 without also accepting T25.

Split into reviewable commits or PRs once T20 is accepted.

---

Raised from the handoff register: docs/2026-08-03-handoff-issues.md (NGMS repo).

[thedancingdeveloper]

Deliberately not done yet — it would be premature.

The register says to split "after T20 is accepted", and that is the right order: splitting commits against a baseline that might change means doing it twice.

Current state, so nothing is lost track of: 4 unpushed commits on feat/mariadb-baseline.

44887922 ci: T19 partial — conformance gate and multi-arch build   ← new, see #105
cfe276fa feat: finish the MariaDB swap in the runtime, CI and docs layers
e18a5315 fix: point the workspace-member list at stackarr-mariadb
d26682ff wip: rescue in-flight MariaDB swap from crashed agent

I committed the previously-uncommitted ci.yml change rather than leaving it in the working tree — this checkout already lost most of a day's rescue work to being uncommitted once, and a commit is recoverable where a working tree is not. Still unpushed, per the register's own precondition.

Blocked on #102.