fix(categories): resolve namespaced entities for HANA backfill + seed classifier descriptions

[jung-thomas]

Problem

Two independent bugs kept the Admin UI Categories facet empty on DEV/PROD:

1. Backfill dies on HANA. scripts/backfill-categories.cjs and the classifier persist() queried bare short-name entity strings (Tutorials, TutorialCategories). Those resolve on local SQLite but emit unqualified TUTORIALS / TUTORIALCATEGORIES SQL against HANA (real tables are COM_SAP_DEVELOPERS_IMS_*). Running the sanctioned backfill via cds bind --exec failed on its first query (Could not find table/view TUTORILS). Only the SQLite-based unit test ever passed, so the bug was invisible. In-service self-heal worked (handler namespace context resolves bare names) — hence the lone stray junction row.

2. seedDescription unseeded. Categories.csv ships only ID/slug/label/sortOrder, so all 8 category rows had empty seedDescription → the embedding classifier degraded to LLM-only everywhere (~1400 AI Core LLM calls, lower quality).

Fix

• Backfill: resolve ims entity objects via a linked model (resolveImsEntities, falling back to cds.linked(cds.load('*')) when cds.model is unset in the standalone runner) and SELECT with explicit .columns('ID').
• Classifier persist(): resolve the junction via cds.entities(ns)[junction] instead of a bare string (loadItemText already did this).
• Seed descriptions: baseline texts in category-seed-descriptions-defaults.js + an idempotent, non-destructive boot-seed (seed-category-descriptions.js, wired into cds.on('served'), VITEST-gated, fills only empty rows). Deliberately not in CSV — a CSV column would full-replace the admin-editable field on every deploy.

Tests

• Unit (test/unit/seed-category-descriptions.test.js): defaults integrity (keys match shipped slugs), idempotency, non-destructive (never clobbers admin edits), self-heal.
• Hybrid (test/hybrid/backfill-categories-hana-resolution.test.js, guarded by HYBRID_TESTS=true + isSafeForWrites()): the load+link resolution path reaches the namespaced HANA tables (.name === com.sap.developers.ims.<Name>, resolved-object SELECT executes) — the exact standalone branch the bare-string code broke on; and the seed runs against real HANA.

Verified against DEV HANA

• Seed: filled 8/8 seedDescription values.
• Backfill dry-run: 4174/4174 items resolved, 0 failures (was fatal on query feat: implement Admin UI (Fiori Elements + freestyle SAPUI5) #1 before the fix).
• Full backfill run + junction-row verification: reported separately once complete.

Targets DEV per branching policy.

[jung-thomas]

DEV backfill — verified against real HANA

Ran the fixed backfill via cds bind --exec against DEV HANA. Junction rows now populate (was fatal on query #1 before the entity-qualification fix). Both fixes in this PR are exercised end-to-end: the resolved-entity SELECT/DELETE/INSERT reach the namespaced COM_SAP_DEVELOPERS_IMS_* tables, and the embedding path fires (seed descriptions seeded → cosine classification, no LLM needed for the confident majority).

Coverage on DEV:

Kind	Distinct items classified	Junction rows
Missions	858 / 890 (96%)	2011
Groups	356 / 362 (98%)	966
Tutorials	2293 / 2922 (78%)	5503

Missions/groups classified almost entirely via the embedding path (0 LLM calls for missions). The remaining gap is items whose cosine score falls below HIGH_THRESHOLD/ambiguity gap and fall through to the LLM fallback, which hit sustained AI Core 429 (rate-limit) — those items are safely skipped, not wiped (pickEmbeddingResult returns null, so persist() is never called with an empty set). They self-heal incrementally: content-publish-session.js re-runs classifyAndPersist on every tutorial republish, and a re-run on a fresh AI Core quota window will close the rest.

Not a code defect — the fix is correct; the tutorial tail is bounded by AI Core LLM throughput.