Task: Event collector and user feedback tracking

Description

Implement the event collection layer that subscribes to opencode's event bus to capture quality-relevant signals. This includes session events, LSP diagnostics, command execution, and inferred user feedback (accept, revert, redo, edit-after-apply). The collector aggregates signals and triggers the scoring engine.

Acceptance Criteria

• packages/opencode/src/quality/collector.ts created
• Subscribes to Session.Event.Updated — captures model_id, provider_id, tokens, cost, tool results, errors per assistant message
• Subscribes to Session.Event.Diff — captures file changes (additions, deletions) for correctness signals
• Subscribes to LSPClient.Event.Diagnostics — snapshots diagnostic state before/after response for static analysis delta
• Detects revert signals from Session.Info.revert population — updates feedback score retroactively
• Detects redo signals via prompt similarity heuristic (same topic within short time window)
• Detects edit-after-apply via file change tracking (user edits files AI just modified)
• All feedback signals trigger retroactive score update in SQLite via scorer
• packages/opencode/src/quality/index.ts created — public API, initializes collector via Instance.state(), exposes Quality.init() and Quality.dispose()
• Quality system initialized during project bootstrap when quality.enabled is true
• All event subscriptions cleaned up on dispose

Technical Details

• Use Bus.subscribe() for event subscriptions following existing patterns
• Store LSP diagnostic snapshots in memory (Map<path, Diagnostic[]>) to compute deltas
• For redo detection: hash prompt topics and compare against last N prompts — simple substring/keyword overlap, not full semantic similarity
• For edit-after-apply: track which files were modified by the last AI response, watch for user modifications to same files within a time window (e.g., 60 seconds)
• Initialize in packages/opencode/src/quality/index.ts using Instance.state(init, dispose) pattern

Dependencies

• Task 001: SQLite storage layer
• Task 002: Quality configuration
• Task 003: Scoring engine

Effort Estimate

• Size: L
• Hours: 8-10
• Parallel: false (depends on 001-003)