feat: distill learned rules from human feedback

[SimonOyaneder]

Summary

• Closes the feedback loop: accumulated human feedback on Canon findings is distilled into concise team preference rules via LLM
• Rules are stored per-repo and injected into the system prompt on subsequent reviews, so Canon learns from team corrections
• Distillation is event-driven (triggered when feedback count crosses a threshold, default 5) and fire-and-forget (non-blocking)

Changes

• Migration 0003: New repo_learned_rules table (one row per repo, upserted)
• src/db/learned-rules.ts: DB access functions (get/store rules, fetch feedback for distillation, count helpers)
• src/intelligence/distiller.ts: Sends feedback to LLM, returns max 10 bullet-point rules
• src/config/index.ts: New distillThreshold config (env DISTILL_THRESHOLD, default 5)
• src/review/prompt-builder.ts: New learnedRules field in PromptContext, injected after CANON.md section
• src/github/pr-handler.ts: Fetches learned rules from DB and passes to prompt context
• src/github/feedback-handler.ts: Triggers distillation after feedback storage when threshold is reached

Test plan

• npm run lint — type-check passes
• npm test — all 330 tests pass (20 new tests added)
• Integration test with real DB (testcontainers)
• Manual verification: simulate feedback → verify distillation → verify rules in prompt

🤖 Generated with Claude Code

[canon-review]

Found 2 pattern deviations: 🟡 2 medium

[canon-review]

🟡 Naming Convention · medium · confidence: 95%

DB row interface uses snake_case property names, while existing TypeScript interfaces in the repo use camelCase and map snake_case at query-read time.

Existing pattern: Use camelCase in TS interfaces/types even when SQL columns are snake_case (e.g., map row.file_path -> path, row.content_hash -> contentHash).
Reference: src/intelligence/embeddings.ts

Suggested change

	existing_pattern: string;
	export interface FeedbackRow {
	category: string;
	severity: string;
	finding: string;
	existingPattern: string;
	replyBody: string;
	replyAuthor: string;
	}
	// In query:
	SELECT
	f.category,
	f.severity,
	f.finding,
	f.existing_pattern AS "existingPattern",
	rf.reply_body AS "replyBody",
	rf.reply_author AS "replyAuthor"

[canon-review]

🟡 Pattern Violation · medium · confidence: 93%

Uses snake_case fields from FeedbackRow directly in app-layer code, deviating from the repo’s camelCase object-property pattern.

Existing pattern: Application-layer object properties are camelCase; SQL snake_case is translated at DB boundary.
Reference: src/intelligence/embeddings.ts

Suggested change

	`Existing pattern referenced: ${row.existing_pattern}`,
	`Existing pattern referenced: ${row.existingPattern}`,
	`Developer reply (${row.replyAuthor}): ${row.replyBody}`,

[SimonOyaneder]

/canon review

[canon-review]

Found 1 pattern deviations: ⚪ 1 low

[canon-review]

⚪ Style Inconsistency · low · confidence: 95%

Import list exceeds the repository’s established multiline import formatting pattern used for long named imports.

Existing pattern: Long named imports are split across multiple lines with one symbol per line for readability.
Reference: src/github/init-handler.ts

Suggested change

	import { getFeedbackCount, getStoredFeedbackCount, getFeedbackForDistillation, storeLearnedRules } from "../db/learned-rules";
	import {
	getFeedbackCount,
	getStoredFeedbackCount,
	getFeedbackForDistillation,
	storeLearnedRules,
	} from "../db/learned-rules";