ci(pull-request-kotlin): persist Gradle build-cache across runs (split restore/save) [NO-JIRA]

[JesperTerkelsen]

Problem

The Kotlin PR test job recompiles ~2m 30s of KSP + Kotlin every run, even for unchanged modules. Dependencies are cached, but setup-java's built-in cache: 'gradle' keys on the gradle files and only saves on a cache miss — so Gradle's build-cache (~/.gradle/caches/build-cache-1, the compiled task output) is written once and frozen, and every run recompiles from scratch. The Sonar cache (fixed key ${{ runner.os }}-sonar) has the same freeze.

Fix

Switch both caches to the split restore/save pattern and save on every run under a per-commit, branch-scoped key, so the build-cache persists and unchanged kspKotlin/compileKotlin come back FROM-CACHE.

key:          ${{ runner.os }}-gradle-${{ github.head_ref }}-${{ github.sha }}
restore-keys: ${{ runner.os }}-gradle-${{ github.head_ref }}-   # this branch's latest (tightest hit)
              ${{ runner.os }}-gradle-                           # fallback: most recent run

• save runs on always() (compiled output is cached even when tests fail) and is skipped on an exact re-run hit.
• Caches are isolated per repository automatically (GitHub + runs-on), so no ${{ github.repository }} is needed in the key.

Validated on a canary (service-feature #737)

Two runs through this branch ref — cold, then warm (identical source):

Step	Cold (run 1)	Warm (run 2)
Restore Gradle cache	0s (miss)	10s (hit, ~690 MB @ 260 MB/s)
Run linter (ktlint + compile)	3m 46s	12s
Run tests with coverage	3m 53s	13s
Upload to SonarCloud	52s	36s
Save caches	10s	15s
Total	9m 11s	1m 56s

Log confirms: Cache hit for restore-key: …-gradle-… → compileKotlin FROM-CACHE, kspKotlin FROM-CACHE, ktlint… FROM-CACHE.

Caveat on the 1m 56s: run 2 had byte-identical source (empty commit), so even the test task cache-hit. A real PR that changes code still recompiles changed files and re-runs affected tests (~2m 17s of real internal-api test execution). So the dependable win is eliminating the ~2m 30s–3m 40s cold recompile; expect a typical code PR around ~4–5 min (from ~9 min), and a docs/no-op PR the full ~1m 56s.

Notes

• Build-cache win only materialises for repos with org.gradle.caching=true (harmless no-op otherwise); dependency caching helps every repo.
• Per-commit keys mean more cache writes — fine on runs-on (no size cap, 10-day lifecycle).
• Branch isolation: the -gradle- fallback can pull another PR's cache on a cold first build of a new branch — low risk for this PR-only, ships-nothing workflow, and moot if runs-on enforces branch-scoping on the magic cache.

⚠️ Blast radius

pull-request-kotlin.yml@main is consumed by 35 repos — merging flips all of them on their next PR run. Two repos pin a SHA (service-charges, service-internal-ocpi-tooling-service) and are unaffected. Consider migrating a couple via the branch ref first as a wider canary before merging to main.

🤖 Generated with Claude Code