[piruin]

Fixes #175 · Related #174

Summary (TH)

เปลี่ยน view ni.patient.observation และ ni.encounter.observation จาก DISTINCT ON เป็น NOT EXISTS เพื่อให้ PostgreSQL กรองด้วย id หรือ survey_response_id ได้ก่อนเลือกแถวล่าสุด ไม่ต้องสร้าง view ใหม่จากทั้งตาราง ni_observation ทุกครั้งที่อ่าน และเพิ่ม index ให้ ni_observation.survey_response_id

Changes

Module	Change
ni_observation	Both views pick the latest valued row per (patient_id, type_id) / (encounter_id, type_id) with a NOT EXISTS anti-join: highest occurrence, ties broken by highest id, the same order as the old ORDER BY. SELECT o.* keeps the column set. New tests in test_latest_observation.py. README documents the views and the index dependency.
ni_questionnaire	survey_response_id on ni.observation.abstract gets index=True. ni.observation is the only stored model that inherits it, so the index is named ni_observation_survey_response_id_index, the name already on production.

Decisions to review

• The encounter view drops the NULL-encounter group (o.encounter_id IS NOT NULL). DISTINCT ON returned one row per type for all observations without an encounter. Every caller (encounter_observation_ids, observation_latest_ids, _compute_observation_latest) filters by encounter_id, ni_encounter_ob_line_action has no menu or reference, and nirun-dop-16.0 does not use the view.
• Rows without type_id or occurrence are left out of both views. Both fields are nullable in the schema, and the anti-join cannot group NULLs the way DISTINCT ON does: every NULL row would come back. The report model already declares type_id required=True. No NOT NULL constraint was added.

How tested

• -i ni_observation,ni_questionnaire --test-enable: 0 failed, 0 errors of 149 tests. -i ni_careplan --test-enable (searches the patient view): 0 failed. ni_careplan_problem is not installed on the local DB, so it was not tested.
• New cases in both views: an older occurrence with a higher id loses, a newer row without a value is ignored, another patient/encounter does not hide the latest, search by id returns only the latest row, and an observation without an encounter is not in the encounter view. The tied-occurrence case is covered by the existing test_same_occurrence_picks_the_last_record.
• Old DISTINCT ON vs new view on a local DB (~3.8k rows): no new rows in either view. Patient view: 1 row fewer (id 4405, no type_id). Encounter view: 18 rows fewer, all without an encounter.
• EXPLAIN (with enable_seqscan = off, the table is small): a read by id uses ni_observation_pkey, a search by survey_response_id uses ni_observation_survey_response_id_index, and both anti-join through ni_observation__{patient,encounter}__ob_type__idx. The only Sort orders the result rows; there is no full scan or sort of ni_observation.
• -u ni_questionnaire on a DB that already has ni_observation_survey_response_id_index keeps the same index (same OID).
• pre-commit clean on the changed files.

After merge (production)

• ni_observation__patient__ob_type__idx on production still has the old (patient_id, type_id) definition. tools.create_index() skips existing names, so rebuild it as (patient_id, type_id, occurrence DESC, id) with CREATE INDEX CONCURRENTLY, outside the module upgrade, to get index-only anti-joins. The encounter-side index was already rebuilt on 2026-10-23.