[codex] add E2E and runtime test matrix

[Jonham]

What changed

• add real FFmpeg E2E coverage for executable discovery, probing, transcoding, streams, screenshots, events, and error settlement
• add a clean npm tarball consumer test and tighten published-package exclusions
• add a local nvm-based Node.js matrix runner with isolated dependency caches
• pair Node.js 18/20 with pnpm 10.34.4 and Node.js 22/24 with pnpm 11.9.0
• support the same matrix in Docker through environment variables
• update CI and release workflows for the explicit Node.js/pnpm pairings

Why

Dependency upgrades need a repeatable compatibility gate that executes real media workflows. Older Node.js releases cannot run the latest pnpm, so each runtime now installs and tests with a compatible pnpm version instead of reusing dependencies installed by Node.js 24.

Impact

Local and CI test runs can validate supported Node.js versions independently. Each local runtime keeps its own node_modules and pnpm store cache under .cache/node-matrix.

Validation

• Node.js 18 + pnpm 10.34.4: 8 E2E tests passed
• Node.js 20 + pnpm 10.34.4: 8 E2E tests passed
• Node.js 22 + pnpm 11.9.0: 8 E2E tests passed
• Node.js 24 + pnpm 11.9.0: 8 E2E tests passed
• packed package consumer test passed
• shell syntax and staged diff checks passed

Stack

This PR is based on codex/native-async-support-spec and should be reviewed after PR #33.

[Jonham]

Node matrix execution report

I ran every Node.js version independently in local nvm and Docker modes. The full raw logs and Markdown reports are retained locally under tmp/node-matrix-reports/.

Environment	Node.js	pnpm	Legacy tests	Functional E2E	Package E2E	Result
Local nvm	18	10.34.4	243 pass, 7 pending, 18 fail	8 pass	1 pass	Legacy baseline failure
Local nvm	20	10.34.4	243 pass, 7 pending, 18 fail	8 pass	1 pass	Legacy baseline failure
Local nvm	22	11.9.0	243 pass, 7 pending, 18 fail	8 pass	1 pass	Legacy baseline failure
Local nvm	24	11.9.0	243 pass, 7 pending, 18 fail	8 pass	1 pass	Legacy baseline failure
Docker	18	10.34.4	Not run	Not run	Not run	Docker Hub pull blocked
Docker	20	10.34.4	Not run	Not run	Not run	Docker Hub pull blocked
Docker	22	11.9.0	Not run	Not run	Not run	Docker Hub pull blocked
Docker	24	11.9.0	Not run	Not run	Not run	Docker Hub pull blocked

Docker Desktop was available (client 29.4.0, server 29.2.1), but every node:<version>-bookworm pull failed against Docker Hub with EOF. A direct pull retry and direct host requests to Docker Hub failed with the same connection reset, so this is an external registry/network blocker rather than a test failure.

The run found and fixed a macOS Bash 3.2 empty-array issue in the Docker argument handling. It also changed the default matrix command so build, legacy, functional E2E, and package-consumer E2E phases all run and report their own status even when an earlier phase fails.

Verified local cache layout

.cache/node-matrix/local/node-v18/project/node_modules
.cache/node-matrix/local/node-v18/pnpm-store-v10.34.4
.cache/node-matrix/local/node-v20/project/node_modules
.cache/node-matrix/local/node-v20/pnpm-store-v10.34.4
.cache/node-matrix/local/node-v22/project/node_modules
.cache/node-matrix/local/node-v22/pnpm-store-v11.9.0
.cache/node-matrix/local/node-v24/project/node_modules
.cache/node-matrix/local/node-v24/pnpm-store-v11.9.0

Each version uses approximately 88 MB and has an independent node_modules and pnpm store. Cache reuse was verified on the second run (dependency checks completed in roughly 145–196 ms).

Report files

tmp/node-matrix-reports/summary.md
tmp/node-matrix-reports/local/node-{18,20,22,24}.{log,md}
tmp/node-matrix-reports/docker/node-{18,20,22,24}.{log,md}

Reproduction

# Full local nvm matrix
bash scripts/test-node-matrix.sh

# One local runtime
NODE_MATRIX_VERSIONS="18" bash scripts/test-node-matrix.sh

# Full Docker matrix
NODE_MATRIX_USE_DOCKER=1 bash scripts/test-node-matrix.sh

# One Docker runtime
NODE_MATRIX_USE_DOCKER=1 NODE_MATRIX_VERSIONS="18" bash scripts/test-node-matrix.sh