ci: make compat-matrix engine-job failures fail the run

[bpamiri]

What

Deletes continue-on-error: true from the tests (engine) job in .github/workflows/compat-matrix.yml. The rustcfml job keeps its flag — it is an informational lane by design.

Why now — the gate is cleared

The issue's own sequencing rule was: burn down the pre-existing leg debt first, then flip. That happened today:

1. PR fix(compat): burn down four compat-matrix root causes (#3302) #3365 (debt burn-down) merged as 6bff05441.
2. Confirming dispatch 30976993285 on develop: all 5 engine jobs green, aggregate 132,170 pass with only the 6 oracle soft-fails (non-blocking via SOFT_FAIL_DBS, tracked under ci: Oracle datasource broken across all engines — Invalid URL, missing DBMS_LOCK, constraint cleanup #2663).
3. The safe-slice hardening (zero-test guard, neutral soft-fail check, honest summary grid) merged earlier via ci: compat-matrix safe-slice hardening — zero-test guard, neutral soft-fail check, honest summary #3366 and validated live on dispatch runs.

With this flip, a failing engine leg fails the run — the weekly banner becomes a real alarm instead of decoration.

Fixes #3302

🤖 Generated with Claude Code

[wheels-bot]

Wheels Bot — Reviewer

TL;DR — This PR removes continue-on-error: true from the tests (engine) job in .github/workflows/compat-matrix.yml, so a failing engine leg now fails the compat-matrix run instead of being silently swallowed. The change is minimal, correct, and well-justified by the linked debt burn-down (#3365 / 6bff05441). One minor docs nit keeps this at comment rather than approve; nothing blocking.

Correctness

The one-line deletion does exactly what the title claims. Verified the surrounding mechanics hold up:

• The job-internal gating already distinguishes soft-fails from real failures — SOFT_FAIL_DBS="oracle" only warns (.github/workflows/compat-matrix.yml:476-480), and OVERALL_STATUS drives the final exit 1 (:502-505). So removing the blanket continue-on-error promotes only genuine, non-soft-fail leg failures to run failures. Correct.
• fail-fast: false is retained (:19), so one failing leg still lets the other four engines finish — you get the full matrix picture on a failure, not a truncated one.
• Both downstream jobs guard with if: always() (publish-results at :631, test-matrix-summary at :668), so a now-failing tests job does not suppress the JUnit publish, PR matrix comment, or artifact upload. No regression there.
• The rustcfml lane correctly keeps its own continue-on-error: true (:605), consistent with the in-file comment describing it as a never-gating informational lane (:594-601).

No blocking findings.

Docs

• .github/workflows/compat-matrix.yml:1-2 — the file-header comment still reads Non-blocking — informational only. That is now stale for the engine legs: after this PR a failing engine job does block the run. Consider tightening it to note that only the rustcfml lane (and oracle soft-fail DBs) remain informational. Non-blocking nit.

Commits

ci: make compat-matrix engine-job failures fail the run conforms to commitlint.config.js — valid type ci, no scope, subject is 47 chars and not ALL-CAPS. Good.

No test changes are expected or needed for a CI-config-only diff.