Redesign CI skip: map each label to its required stacks

[Yicong-Huang]

Task Summary

Redesign the precheck's CI-skip computation so each label has an explicit list of stacks it requires, and the run set is the union across all labels on the PR. Replace the current ad-hoc rule cascade.

Current state

required-checks.yml's precheck applies hand-written rules:

• labels ⊆ {docs, dev} → skip all stacks
• only frontend (with optional docs/dev) → skip every non-frontend stack
• no frontend label → skip frontend
• otherwise → run all

This works for the cases we have but does not generalize: every new shape (e.g. "python-only PR", "no engine label") needs another bespoke branch and another rule to reason about.

Proposed design

Define a static mapping from each labeler-applied label to the stacks it requires:

label	frontend	scala	python	agent-service
frontend	✓
python		✓	✓
engine		✓	✓
service		✓		✓
common		✓
dependencies
ddl-change		✓
ci	✓	✓	✓	✓
docs
dev
feature / fix / refactor (branch-name labels)	(no stack opinion)

precheck reads PR labels and outputs run_<stack> = stack ∈ ⋃ map[label]. Empty union → all skipped. release/* labels keep their existing role driving the backport matrix and contribute no stacks themselves.

Open questions for the design review

• python triggers the scala stack as well because the scala backend tests shell out to PyAmber operators; a python-side regression in amber/src/main/python/** can break scala integration tests, so we run both whenever python changes.
• service currently matches every *-service/** directory, which includes both the scala services and agent-service. The mapping above forces both stacks; if we want finer granularity we'd add a dedicated agent-service label (split *-service/** against agent-service/**).
• dependencies is a no-op for stack selection. Every dependency manifest the labeler matches lives under a component directory and is already covered by that component's label (frontend/package.json → frontend, *-service/build.sbt → service, amber/requirements.txt → engine, etc.). The only manifests outside a top-level component are amber/{requirements,operator-requirements}.txt, which still sit under amber/** and therefore pick up the engine label that already runs both scala and python.
• ci deliberately fans out to all stacks — workflow edits can affect anything, so it's the safety valve.
• feature / fix / refactor are branch-name driven and orthogonal to file paths; suggest treating them as no-op for stack selection.

Migration

• Replace the current rule cascade with a LABEL_STACKS map and Object.values(...).flat() style union.
• Keep the docs/dev "skip all" outcome as an emergent result (their rows are empty), not a special case.
• Keep the labeler wait already in precheck so labels are settled before the union is computed.

Priority

P2 – Medium

Task Type

• Code Implementation
• DevOps / Deployment