feat: closed-loop engagement tuning (Phase I)

[psdjungpulzze]

Phase I — Engagement / Effectiveness Loop

The last of the six loop archetypes. Adds ENGAGE signal emission to notification mark-read and a weekly tune-engagement job that aggregates per-tenant engagement scores into `loops_calibrations`.

Depends on: #32 (Phase H — completion). Branched off Phase H.

What's in this PR

• `src/app/api/v1/notifications/[id]/read/route.ts` — emits ENGAGE on the first read (`wasUnread` guard) so we count actual engagement rather than every revisit. Notification has no `organizationId` field, so we attribute via the user's primary OrgMembership (oldest by `joinedAt`). Multi-org users get best-effort attribution.
• `src/worker/jobs/tune-engagement.ts` (new) — pure derivation `computeEngagementScale` (count/day) + per-tenant orchestration. Aggregates ENGAGE by feature, writes `engagement-rate:` calibrations to `loops_calibrations`. Skips features with no signals; updates existing rows in place.
• `src/worker/index.ts` — registers the job; cron `30 5 * * 1` (Monday 05:30, after threshold tuning at 05:00).

Cold-start safety

• `LOOPS_ENABLED !== "true"` → emit + job both no-op
• User has no org membership → no emission (best-effort attribution)
• No ENGAGE signals → no calibrations written
• Read-site consumers (callers of `getCalibration("engagement-rate:")`) get null on cold start

Test plan

• 12 new tests: 6 math (rate normalization, defensive zero/negative handling, rounding) + 6 orchestration (LOOPS_ENABLED gate, multi-tenant iteration, per-tenant fault isolation, feature aggregation, 30-day lookback, update-in-place vs create)
• `npm run typecheck` clean
• `npm test` — 456 passing vs Phase H baseline 444 (+12); same 8 pre-existing failures unchanged

Loops architecture status after this PR

Archetype	Status	Wired in
Suggestion → Acceptance	full loop	B + C + D
Prediction → Outcome Calibration	full loop (consumer pending)	G
Vocabulary / Mapping	client ready (UKB extension pending)	F
Threshold / Trigger	full loop (consumer pending)	B + H
Ranking / Routing	full loop	E
Engagement / Effectiveness	full loop (consumer pending)	B + I (this PR)

All six archetypes have producer + storage + aggregation/learning jobs wired. Phase J wires the remaining production consumers (cost forecast → `budget.ts`, threshold → auto-flag severity, engagement → digest) so every loop steers behavior end-to-end.