Fix care-plan request_changes loop to actually revise the plan

[khaledsulayman]

Fixes #163. Follow-up to #147.

Problem

After #147, requesting changes on a care plan correctly loops instead of writing — but the regenerated plan comes back byte-for-byte identical. The clinician's feedback has no effect.

Root cause

1. Wrong loop target — request_changes routed to GenerateBundle, but goals/activities live in the planning brief produced by ComposePlan. GenerateBundle only serializes an existing brief; it can't change clinical content.
2. Feedback dropped — GenerateBundle only receives the automated fhir_review_feedback, never the clinician's .careplanReview.comment.

Changes

• Workflow (acp-writer-workflow.yaml): request_changes → ComposePlan (not GenerateBundle); ComposePlan now receives careplan_feedback from .careplanReview.comment.
• Compose service (llm_reasoning.py): seed the composer's brief_review_feedback from careplan_feedback via _seed_feedback(), so plan_composer addresses it on the first pass (it already renders feedback under "Reviewer Feedback (address these issues)").
• BFF mapper (sonataflow_client.py): infer_current_state distinguishes the ComposePlan re-run from the subsequent GenerateBundle during the loop, using completed_at timestamps.

InitCarePlanReview already preserves careplanReviewCount across loops, so iteration count + previousFeedback keep working.

Testing

• Unit tests updated/added in test_bff_review_mapping.py for the loop-detection (request_changes → ComposePlan → GenerateBundle → gate). All pass.
• On-cluster end-to-end verification (regenerated plan actually reflects requested changes) to follow as a PR comment.

_{Generated by Claude Code under the supervision of Khaled Sulayman}

[khaledsulayman]

✅ Verified end-to-end on-cluster (Playwright)

Deployed fix-careplan-feedback-regeneration (current main + this fix, tag 2ab3f4b) to a personal namespace and drove the full flow in the browser with patient Thomas Walker (HTN bundle), KB = AHA-HTN-2023.

The fix works. After request_changes, the workflow looped back through ComposePlan (confirmed: run status returned to running | compose_plan) and regenerated a plan that reflects the clinician feedback. Before this change, round 2 was byte-for-byte identical.

Feedback submitted: "Add a numeric BP target <130/80; patient already on amlodipine 5mg — don't start Lisinopril, optimize existing therapy; add a follow-up recheck."

	Round 1	Round 2 (after feedback)
Goal	"Lower blood pressure to target range"	"Lower systolic blood pressure to target range"
Med activity	Start Lisinopril 10mg daily	Evaluate the adequacy of Amlodipine 5mg daily
Diet	DASH + exercise	DASH + exercise + sodium <2300 mg/day
Activities	5	4

• The UI review panel surfaced "Previously requested: …" with the submitted comment (previousFeedback populated).
• Approve → run completed, CarePlan written (careplanId set). Final CarePlan goal rationale: "Target Systolic blood pressure: < 130.0 mmHg" — the numeric target carried through.

Incidental (not this PR):

• Create path (POST /api/v1/runs) returns 202 promptly now — the old acp-writer UI: care-plan generation 504s (blocking POST vs long pipeline + route timeout) #124 504 through the route appears resolved by acp-writer: Convert Phase 1 pipeline steps to async callbacks #157 (async Phase 1).
• Cosmetic: the review panel labels the second round "round 1" (reviewIteration/careplanReviewCount reads 0 on the GET after the loop). Filing separately.

_{Generated by Claude Code under the supervision of Khaled Sulayman}