Port the explore-timeline server foundation to main

Why

The live site's fast merged cross-source timeline feed and the Explore over-time chart run on deploy-branch-only server code that was never merged to main: explore-timeline-substrate.ts (dual-dialect merged-timeline queries), the rs.explore.timeline operation, the GET /_ref/explore/records route, and the semantic_time column plus its idx_records_semantic_time / idx_pg_records_semantic_time expression index. On fresh origin/main none of these exist: the records table carries only emitted_at (ingest time), and the only timeline reader is GET /_ref/records/timeline (the ref-records-timeline operation), which orders by ingest time rather than authored/semantic time.

This is the server analog of the already-landed frontend foundation port. It brings the same semantic-time feed substrate the live site uses onto main, so main orders the merged timeline by when things actually happened (the manifest-declared semantic field) instead of when they were ingested.

It also unblocks the separate, later bucket-aggregate PR (the indexed over-time-chart date_trunc/strftime , COUNT(*) GROUP BY), which depends on the semantic_time column, the expression index, and the substrate's multi-stream scope plumbing — none of which exist on main today. Per the Codex gate, a migration was discovered necessary, so the foundation must land first as its own change before the bucket endpoint can build on it. This change is the foundation port; it does not add the bucket-aggregate endpoint.

What Changes

• Add an additive, idempotent records.semantic_time column (TEXT NOT NULL DEFAULT '') on both backends, with a dual-dialect migration that runs on a pre-existing records table without a bulk UPDATE.
• Add the expression keyset index idx_records_semantic_time (SQLite) / idx_pg_records_semantic_time (Postgres) on (connector_instance_id, stream, COALESCE(NULLIF(semantic_time,''), emitted_at) DESC, record_key DESC), created IF NOT EXISTS after the column is guaranteed present.
• Write semantic_time at ingest via the record upsert (ON CONFLICT ... DO UPDATE SET semantic_time = excluded.semantic_time); existing rows are read through COALESCE(NULLIF(semantic_time,''), emitted_at) so they fall back to emitted_at until a real value is written. No mass backfill in this change.
• Port explore-timeline-substrate.ts and the rs.explore.timeline operation (k-way merge across (connector_instance_id, stream) partitions, ingest-id snapshot anchor, semantic-time ordering, dual-dialect deps factory).
• Add direction=asc to the merged timeline route so oldest-first paging is a server keyset traversal from the genuinely-oldest record, with direction pinned in the cursor.
• Mount the reference-only route GET /_ref/explore/records that the operator console's Explore feed consumes.

Capabilities

Modified

• reference-implementation-architecture — the reference _ref read surface gains the explore-timeline merged-feed route; the records storage schema gains the additive semantic_time column and its expression index on both backends.

Impact

• Reference server: reference-implementation/server/db.js, postgres-storage.js, explore-timeline-substrate.ts (new), operations/rs-explore-timeline/ (new), server/routes/ref-admin.ts, queries/records/ingest/upsert-record.sql, plus dual-dialect substrate tests.
• Storage: one additive column + one expression index per dialect. No protocol surface change; _ref routes are reference-only artifacts, not core PDPP API.
• Downstream: unblocks the separate bucket-aggregate PR (not in this change).
• The bucket-aggregate / over-time-chart count endpoint is explicitly out of scope here and lands as its own follow-up change on top of this foundation.