clusterversion: bump MinSupported from v25.3 to v25.4

[celiala]

Part of the quarterly M.4 "Bump MinSupported" task as outlined in pkg/clusterversion/README.md.

This PR bumps MinSupported from v25.3 to v25.4 in 5 commits:

Commit 1: Refactor CLAUDE.md into modular runbooks

Refactors the monolithic 4114-line CLAUDE.md into a modular structure:

• Main CLAUDE.md (113 lines): Overview, navigation, and general guidance
• Individual runbooks in pkg/clusterversion/runbooks/:
  • R1_prepare_for_beta.md
  • R2_mint_release.md
  • M1_bump_current_version.md
  • M2_enable_mixed_cluster_logic_tests.md
  • M3_enable_upgrade_tests.md
  • M4_bump_minsupported_version.md
• Updated README.md Claude Prompts to reference new runbook locations

This makes each runbook easier to read, maintain, and use independently.

Note: This commit also adds Step 7 to the M.1 runbook documenting the requirement to update pkg/storage/pebble.go with a version mapping for the new release, which was missing from the original M.1 documentation.

Commit 2: Bump MinSupported from v25.3 to v25.4

• Updated MinSupported constant in pkg/clusterversion/cockroach_versions.go from V25_3 to V25_4
• Updated pkg/storage/pebble.go:
  • Added V25_4 entry to pebbleFormatVersionMap with FormatMarkForCompactionInVersionEdit (fixes missing entry that should have been added during M.1)
  • Updated MinimumSupportedFormatVersion from FormatValueSeparation to FormatMarkForCompactionInVersionEdit

After this change, clusters running v25.3 can no longer connect to clusters running master, and direct upgrades from v25.3 to master are no longer supported.

Commit 3: Remove schema changer rules for release_25_3

• Deleted entire release_25_3 rules directory (~13,000 lines)
• Removed import and registry entry from plan.go
• Updated Bazel dependencies

These frozen schema changer rules are no longer needed since v25.3 nodes can no longer participate in mixed-version clusters.

Commit 4: Remove bootstrap data for v25.3

• Deleted 25_3_system.keys, 25_3_system.sha256
• Deleted 25_3_tenant.keys, 25_3_tenant.sha256
• Removed V25_3 entry from initialValuesFactoryByKey map
• Removed go:embed variables for v25.3 bootstrap data
• Updated BUILD.bazel to remove embedsrcs

Bootstrap data files are no longer needed since clusters can no longer start at v25.3 (below MinSupported).

Commit 5: Remove local-mixed-25.3 test configuration

• Removed local-mixed-25.3 config from logictestbase.go
• Removed from default-configs and schema-locked-disabled sets
• Deleted generated test directories for local-mixed-25.3
• Removed all skipif/onlyif references from 64 logic test files
• Regenerated Bazel BUILD files via ./dev gen bazel

The local-mixed-25.3 configuration simulates a mixed-version cluster with v25.3 nodes, which can no longer connect after bumping MinSupported.

---

Reference PR: #157767 (v25.2 → v25.3 MinSupported bump)

Epic: None
Release note: None

[blathers-crl]

Your pull request contains more than 1000 changes. It is strongly encouraged to split big PRs into smaller chunks.

_{🦉 Hoot! I am a Blathers, a bot for CockroachDB. My owner is dev-inf.}

[cockroach-teamcity]

This change is 

[celiala]

Unit tests are failing for pkg/upgrade/upgrades/first_upgrade_test.go: TestFirstUpgradeRepair

• @fqazi or @rafiss - could you recommend what to do for the TestFirstUpgradeRepair test failure?
• See also analysis and recommendation from Claude below

thank you!

# pkg/upgrade/upgrades/first_upgrade_test.go
=== RUN   TestFirstUpgradeRepair
...
first_upgrade_test.go:288: expected query 'ALTER TABLE test.public.foo RENAME TO bar'
error 'relation "foo" \(106\): invalid foreign key backreference: missing table=123456789...'
got: pq: referenced descriptor ID 123456789: looking up ID 123456789: descriptor not found

---

Claude's Analysis: TestFirstUpgradeRepair Test Failure

Issue: Error message changed from detailed FK-specific format to simpler descriptor lookup format.

TL;DR

Questions

1. Was this error message change intentional?
2. Is the detailed FK validation error valuable for debugging, or is "descriptor not found" sufficient?
3. Should we standardize on the simpler format (matching repair_test.go)?

Recommendation

Update test expectations (Option 1). The new format:

• Correctly identifies the problem
• Matches existing test patterns
• Maintains test's core purpose (validating repair functionality)
• 3-line change

If detailed FK errors are preferred, investigate and restore previous behavior (Option 2).

Root Cause

Descriptor validation now happens earlier during lookup (pkg/sql/catalog/errors.go) rather than during FK constraint validation. The simpler error surfaces before the detailed FK validation runs. This pattern matches pkg/sql/tests/repair_test.go:910 which already expects the simpler format.

Fix Options

Option 1: Update Test Expectations (Recommended)

// Change from:
const errRE = "relation \"foo\" \\(106\\): invalid foreign key backreference: missing table=123456789..."
const errReForFunction = " function \"f\" \\(107\\): referenced descriptor ID 123456789..."

// To:
const errRE = "referenced descriptor ID 123456789: looking up ID 123456789: descriptor not found"
tdb.ExpectErr(t, errRE, "ALTER TABLE test.public.foo RENAME TO bar")
tdb.ExpectErr(t, errRE, "ALTER FUNCTION test.public.f RENAME TO g")

Why: Simpler, consistent with existing tests, still validates corruption detection and repair.

Option 2: Investigate If Error Change Was Intentional

If the detailed FK error is valuable for debugging, investigate recent descriptor validation changes and consider restoring it.

Option 3: Accept Both Formats

Make test flexible to accept either error format using strings.Contains() checks.