feat: carry bounded data intervals through Dagster and DLT

Goal and scope

Carry explicit time windows from scheduling into ingestion, so adjacent loads and historical replay have bounded extraction semantics. Third PR in the stack; review its diff against the dataset-constraints branch.

Change map and risk

• Add validated UTC half-open DataInterval values to runtime routing without changing the runtime-context protocol.
• Translate Dagster calendar windows and backfill/replay mode, including daylight-saving transitions.
• Add opt-in DLT interval_column and scheduled interval_lookback. Require merge writes and use stateless bounded DLT incrementals, preserving the live cursor.
• Add a runnable lakehouse example, usage/guarantee guide, and the source research behind the three changes.
• Leave Kafka checkpointing and legacy unbounded DLT state semantics unchanged.

Original verification

• Original head focused interval/runtime/Dagster suites: 59 passed.
• Final make check: 7,417 passed, four skipped, 194 integration-deselected; lint, formatting, Python/TypeScript types and repository consistency checks pass. make test-core-regression: 447 passed.
• Canonical real MinIO/Nessie storage integration: 101 passed across seven suites. Installed-wheel verification covers 40 packages, without editable installs or repository-source imports.
• Real public discovery → Dagster → DLT → Nessie/MinIO execution: boundary record 202 retained; late record 404 recovered through lookback; historical replay leaves destination IDs [101, 202, 404], without duplicates. Upper-boundary record 303 is excluded.
• Real DLT extraction/local Iceberg transaction tests inspect a persisted live cursor across scheduling, replay, rejected writes, committed-but-reply-lost failure, retry, and empty extraction. Calendar tests include 23-hour and 25-hour days.
• Reproduce against a disposable configured lakehouse: PHLO_OBSERVATORY_SETTINGS_BACKEND=memory uv run --locked python examples/data_intervals.py. See docs/guides/data-intervals.md for setup and semantics.
• An initial full baseline run hit a process-shutdown timing failure in unchanged code; the original and integrated shutdown suites both passed in isolation, and the complete final rerun passed.
• Links and generated references pass. Final production documentation build was blocked by the orb's confirmed 6 GiB memory limit, not claimed green.
• New feature: regression coverage includes timezone boundaries, source bounds, replay, late arrivals, empty extraction and cursor preservation through failures.

Review correction

Absent sources and empty bounded extracts now share no-data finalisation, including the run observation. The real DLT/Iceberg regression reopens SQLite evidence: before the fix the run remained running; after it the run is no_data with an input resource recording zero records read for the correct attempt. All 178 selected DLT tests pass. This branch also includes the quality corrections from the preceding PRs.

The original-head GitHub run did not pass: its JavaScript lane stalled downloading Ubuntu browser dependencies and was cancelled before test execution, causing the aggregate gates to fail. The updated head requires a fresh CI result.

Post-correction head: make check passes with 7,419 tests, four skipped and 194 deselected, including lint, complexity, formatting, types and repository consistency checks. make test-core-regression passes all 447 tests. The DLT suite passes 178 selected tests, with two deselected.

Recovery and maintainer decisions

Opt-in; no migration, deployment, or extra checkpoint store. Remove interval declarations before downgrading. Stable non-null keys are required for retry-safe merges. Source authors must push bounds into API/SQL queries themselves; arbitrary custom source state is outside the guarantee. No external-cursor/table atomicity, contiguous watermark, or multi-file atomic commit is promised. Inspect ambiguous writes before retrying. Maintainer review is pending; no independent reviewer or release qualification is claimed.