feat(catalog): bump redis PECL 6.2.0 → 6.3.0 (restores PHP 8.5)

[phramz]

Summary

• Bumps catalog/extensions/redis.yaml from redis 6.2.0 → 6.3.0 and drops the exclude: [{ php: "8.5" }] block.
• Restores redis in the three PHP 8.5 compat-harness input-shape fixtures (multi-ext-85, none-reset-85, single-ext-85), which were temporarily swapped to apcu in feat: enable PHP 8.5 core + PECL rebuild (phase 2) #37. Also restores redis in multi-ext-top14-arm64-85.
• Syncs the duplicate inline runtime catalog at cmd/phpup/main.go (extracted to a new runtimeExtensionSpecs() function) with the YAML source of truth.
• Adds TestRuntimeExtensionSpecsMatchYAML to guard against future drift between the inline runtime catalog and catalog/extensions/*.yaml (mirrors the pattern of TestCatalogBundledMatchesCompat).

Closes #38.

Why 6.3.0

redis 6.2.0's common.h includes ext/standard/php_smart_string.h, an internal PHP header removed in the 8.5 source tree — compilation fails with fatal error: ext/standard/php_smart_string.h: No such file or directory. Upstream redis 6.3.0 drops the include.

Implementation discovery: duplicate catalog source of truth

Local CI reproduction (make ci-cell OS=jammy ARCH=x86_64 PHP=8.5) after the YAML-only change failed with:

resolve: extension redis 6.2.0 for PHP 8.5 linux/x86_64/nts not found in lockfile

Root cause: cmd/phpup/main.go carried a hardcoded inline catalog.Catalog{...} literal used by the runtime install path, duplicating the version pins in catalog/extensions/*.yaml. The YAML change alone was insufficient.

Fix: extract the inline spec into runtimeExtensionSpecs() in a new cmd/phpup/runtime_catalog.go, bump redis to 6.3.0 there, and add TestRuntimeExtensionSpecsMatchYAML to enforce inline↔YAML version/kind parity going forward. Further unification (making the runtime load from YAML directly) is a bigger refactor outside this slice's scope; the drift test makes the current duplication safe.

Upstream 6.2.0 → 6.3.0 behavior review

Feature-additive release; no user-facing BC breaks.

• Added: hash field expiration commands (HEXPIRE/HEXPIREAT/etc.), Valkey DELIFEQ, Redis vector-set commands, GEOSEARCH[STORE] BYPOLYGON, HGETEX/HSETEX/HGETDEL, Redis::serverName / Redis::serverVersion, constructor database option, connection maxRetries, INI knob to restore 5.x legacy session-lock-failure behavior.
• Changed: internal refactors; Redis::execHello visibility changed to protected (affects only subclasses that called it directly — rare).
• Removed: the ext/standard/php_smart_string.h include (the reason for this bump).
• Deprecated: none new.

Upstream notes: https://github.com/phpredis/phpredis/releases/tag/6.3.0

PR self-publish flow

Per the bundle-rollout design, PR CI will rebuild all 10 redis bundles (5 PHP ABIs × 2 arches), push them to GHCR, commit the refreshed bundles.lock back to this PR head, then run the compat harness against the freshly published bundles. The PR merges atomically with code + bundle content + lockfile.

Local pre-push verification

• make check — green.
• make ci-cell OS=jammy ARCH=x86_64 PHP=8.5 — green (11/11 8.5 fixtures pass including the newly redis-bearing multi-ext-85/none-reset-85/single-ext-85).

Test plan

• CI: planner emits 10 redis rebuild cells (5 PHP × 2 arch).
• CI: build-extension.yml builds redis 6.3.0 for all 10 cells and pushes to GHCR.
• CI: lockfile-update commits refreshed bundles.lock back to this PR head; lockfile contains ext:redis:6.3.0:{8.1..8.5}:linux:{x86_64,aarch64}:nts and no ext:redis:6.2.0:* entries.
• CI: compat harness green for every redis-bearing fixture (8.1–8.4 input-shape, 8.5 input-shape, multi-ext-top14-arm64-85, jumbo fixtures).
• CI: make check passes; TestRuntimeExtensionSpecsMatchYAML is green.

Spec: docs/superpowers/specs/2026-04-24-redis-6.3.0-bump-design.md

[phramz]

CI status note

All 20 `pipeline` cells (4 OS × arch × 5 PHP) pass — redis 6.3.0 builds cleanly across the matrix and the harness fixtures are green, including the newly redis-bearing 8.5 fixtures (`multi-ext-85`, `none-reset-85`, `single-ext-85`).

The 4 `smoke-action` cells fail with:

```
resolve: extension redis 6.3.0 for PHP 8.4 linux/x86_64/nts not found in lockfile
```

This is structural, not a regression in this PR: `smoke-action` resolves bundles through the committed `bundles.lock` + GHCR, and the `publish` job (which writes `ext:redis:6.3.0:*` entries to `bundles.lock` and pushes the bundles to GHCR) is gated on `if: github.ref == 'refs/heads/main'`. Any PR that bumps a bundle version hits this; merging the PR triggers `publish` on main, which auto-commits `chore(lock): update bundles.lock from pipeline ` and unblocks GHCR — matches the recent main-history pattern (`2135981`, `5ef5c4f`, `de13af7`).

Filed as #83 — proposes making `smoke-action` consume the `pipeline`-built oci-layout artifacts via `PHPUP_REGISTRY`/`PHPUP_LOCKFILE` env so future bump PRs can validate end-to-end without GHCR roundtrip (matches the design intent in `docs/superpowers/specs/2026-04-23-local-ci-unification-design.md`).