fix: intercept plugin console output to prevent TUI corruption - #19109

Issue for this PR

Closes #19108

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Plugins and their npm dependencies call console.log/warn/error which writes directly to stdout, corrupting the TUI display. The logLevel config only controls OpenCode's own slog-based logging and has zero effect on plugin console output.

This is a problem I hit daily with @rehydra/opencode (its dependency rehydra calls console.warn("Validation warnings:", ...) on every anonymization pass) and custom .opencode/plugins/ scripts. The TUI becomes unreadable with JSON objects and warning messages mixed into the chat.

The fix adds interceptConsole() in packages/opencode/src/plugin/index.ts that replaces global console methods right before plugins load. All output is routed through Log.create({ service: "plugin:console" }), so it respects logLevel and goes to the log file instead of stdout.

This works because plugins run in the same Bun process — they share the global console object. Intercepting it once catches both plugin code and their transitive dependencies (which is why adding a log property to PluginInput alone wouldn't solve this — third-party deps will always use console.*).

Related to #12931 but distinct: that issue covers uncaught errors in event handlers, this covers normal console output from working plugins.

How did you verify your code works?

1. Identified the spam sources by reading plugin code (rehydra/dist/core/anonymizer.js:315 and .opencode/plugins/memorix.js)
2. Verified that console.warn in plugins goes directly to TUI stdout (confirmed by screenshot evidence)
3. Traced the plugin loading path in plugin/index.ts — plugins are loaded via import() with no console interception
4. The Log.create() API is already used throughout the codebase (e.g. const log = Log.create({ service: "plugin" }) on the same file's line 17)

Screenshots / recordings

Before (TUI corrupted by plugin console output):

{ cValidation warnings: [tool.execute.after leak(s)",
{ cValidation warnings: "AK", potential PII leak(s)",
{ c[memorix] hook fired: session.idleal PII leak(s)",
  code: "POTENTIAL_PII_LEAK", potential PII leak(s)",
  message: "Leak scan found 2 potential PII leak(s)",
}

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR