feat: track api-diff baseline via Renovate and store diffs in docs/apidiffs

[zeitlinger]

Summary

Fixes the manual-bump drift in the api-diff baseline (#2170) and makes API changes visible in review.

• Single source of truth. The baseline lives only in the <api.diff.baseline.version> pom property (bumped from the already-drifted 1.5.1 to the actual latest release 1.6.1). mise.toml and api-diff.yml no longer hardcode it.
• Renovate owns the bump. A custom regex manager tracks the latest published io.prometheus:prometheus-metrics-core on Maven Central and bumps the property on the renovate/api-diff-baseline branch. Because Renovate only proposes published versions, there is no Maven Central propagation race and no post-release workflow or app token needed.
• Diffs committed to docs/apidiffs/. mise run api-diff syncs the japicmp per-module reports into docs/apidiffs/<module>.diff via .github/scripts/sync-api-diffs.sh, stripping the volatile preamble so files only churn on real API changes. The api-diff workflow fails if they are stale, so every API change shows up in the PR diff.
• Regeneration on bump. generate-api-diff-baseline.yml regenerates docs/apidiffs on the Renovate branch and pushes it back, mirroring generate-protobuf.yml.
• docs/apidiffs/** is marked linguist-generated so flint skips it.

Notes

• The seed diffs are large because 1.6.1 predates @StableApi (the bootstrap noise documented in feat: Add StableApi marker and API diff check #2168). They shrink to near-empty once a release contains the annotations.
• Like the protobuf flow, a GITHUB_TOKEN push doesn't re-trigger CI — close/reopen the Renovate bump PR to run the api-diff check after regeneration (noted in the workflow).

Validation

• mise run api-diff (generates the 25 seed diffs; verified idempotent)
• mise run lint
• renovate-config-validator --strict, actionlint, zizmor

Closes #2170

Also fixes a pre-existing compat-test break on main

micrometer-compatibility and jmx-exporter-compatibility fail on main (e.g. #2173), unrelated to this change. The compat harness installs local artifacts with -Dmaven.test.skip=true, which skips building the *:test-jar artifacts that the activeByDefault default profiles declare as test dependencies, breaking resolution (e.g. prometheus-metrics-exposition-textformats:jar:tests).

Fixed by deactivating those profiles in the compat install (-P !default), matching what the release task already does (-P release,!default). Verified locally: full mvnw install -Dmaven.test.skip=true -P !default → BUILD SUCCESS.

Reviewer note: the seed diffs show the full stable surface (expected)

The committed docs/apidiffs/current_vs_latest/*.txt list the entire @StableApi surface as additions, not a small delta. This is the documented bootstrap state from #2168, not a bug:

• The japicmp include filter is @io.prometheus.metrics.annotations.StableApi. The baseline 1.6.1 jar predates @StableApi, so its filtered surface is empty → every annotated class in current code registers as "NEW". The header genuinely compares against the 1.6.1 jar; the annotation filter is what makes everything look added.
• Verified the filter is correct: only @StableApi-annotated types appear (non-stable classes like CKMSQuantiles/Buffer are excluded; CallbackMetric shows only as an inherited NEW SUPERCLASS reference). It is the same root cause as the breaking-api-change-accepted label on this PR.

It self-corrects after the next release ships @StableApi: Renovate bumps the baseline to that release, the bump workflow regenerates current_vs_latest/ as an annotated-vs-annotated (near-empty) diff, and the archived <new>_vs_<old>/ becomes the first real release diff. So the current files are best read as a one-time record of the initial declared stable API surface.