Verify derived plan steps follow changes to their premises

[Maya-Kid]

A plan step represented by a business-rule conclusion should follow changes to its factual premises without changing unrelated conclusions. Add three real PostgreSQL store tests for that contract:

• Closing and rejecting a premise retires the affected current conclusion, preserves the appropriate historical view and leaves the unrelated step's row unchanged. Duplicate observations, occlusion and a late earlier observation are covered along that path.
• With two alternative proofs, removing one preserves the conclusion and replaces its proof; removing the last retracts it.
• With explicitly bound open statements, typed materialization followed by explicit temporal reconciliation and derivation retires the old-location step while preserving its past interval.

All three tests run normally on dev; there are no newly ignored tests and no production changes or dependencies on #884, #896 or #897. They exercise existing store APIs and explicitly supplied alignment/reconciliation, not an automatic ingestion-to-decision loop.

Validation: 8000623, based on dev@9f50e3c: Rust format, strict workspace/all-target Clippy and the PostgreSQL 16/pgvector workspace suite passed (1,081 passed, 0 failed, 5 existing ignored). UTOPIA_TEST_REQUIRE_DB=1 was set and all three new passing test names were verified. The unavailable PDF fallback utility limits that unrelated coverage.

Run against a migrated, dedicated PostgreSQL/pgvector database:

UTOPIA_TEST_REQUIRE_DB=1 cargo test --locked -p utopia-store --test store \
  a_plan_step_follows_its_premise -- --test-threads=1

Set UTOPIA_DATABASE_URL for that database. The HTTP replay and its version-specific assertions remain experimental evidence on the fork, outside this PR. The two proposed-behavior gaps and executable reproductions are tracked separately in #899 and #900.

Refs #875.

[WaylandYang]

Useful regression coverage on the real store path, and the docstring is honest that it exercises existing APIs with explicit reconciliation rather than the automatic loop (#899 and #900 remain open). Two optional suggestions: split the first test's seven phases so a failure in one does not hide the others, and note in the observe helper that it mirrors pending::confirm's ordering so drift is caught by a reader. LGTM.