Remove covered derived intervals after propagation completes

[Floating-Y]

Why

When a narrow derivation arrives before a wider derivation of the same triple, both remain in the final output. A sub-property derivation of A q D [10,20) followed by a transitive derivation over [0,100) reproduces this issue.

This follows up on the residual issue recorded in the #1034 closing comment and the cleanup suggested in the #1040 review.

What changes

After propagation completes, group outputs by (predicate, subject, object) and reuse span_contains to remove intervals covered by another single output. Equal intervals retain one output; incomparable intervals remain separate without merging or union-based coverage. Existing unbounded interval semantics are preserved.

The surviving Derived retains its complete ordered premises, via, and rule. The reproduction now returns only [0,100), supported by premises 102, 103, 104, with via = q and rule = Transitive.

Cleanup preserves #1051's propagation behavior, including narrower but shorter intermediate proofs needed to finish within MAX_DEPTH. Public interfaces, depth limits, and output-cap semantics remain unchanged.

Seven regression tests cover arrival order, input permutations and UUID ordering, interval boundaries, proof preservation, equal-output cleanup, and the narrower-shorter-proof case.

How it was checked

• The narrow-first regression failed before the fix and passed afterward.
• cargo test --locked -p utopia-reason: 102 passed.
• cargo fmt --all --check: passed.
• cargo clippy --workspace --all-targets -- -D warnings: passed.
• cargo test --workspace: 1,426 passed, 0 failed, 6 ignored.
• cargo build --workspace: passed.
• Under web/, frozen-lockfile installation, 228 tests, and the build passed.

Workspace tests used a dedicated migrated PostgreSQL 16/pgvector database with UTOPIA_TEST_REQUIRE_DB=1 and UTOPIA_TEST_REQUIRE_PDFTOTEXT=1. An initial run encountered Windows proxy interference; the full suite passed after rerunning with loopback addresses in the test process's NO_PROXY.

Ignored tests, unconfigured external-service integration tests, and the opt-in benchmark were not executed.

Before review

• Every commit is signed off (git commit -s)
• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace pass

[WaylandYang]

Thanks @Floating-Y. This is the cleanup the #1040 review left open, and doing it after propagation rather than during it is the right place: a narrower output with a shorter proof still has to stay in the adjacency lists while chains are being extended, which your last test pins.

Beyond the 102 tests, I compared dev with this branch on 400 deterministic random graphs (transitive, sub-property and inverse predicates, bounded, half-open and open intervals):

	dev	this branch
Triples derived	19,028	19,028, the same set
Outputs	34,581	25,523
Outputs covered by another output of the same triple	9,058	0
Intervals on dev not covered by an output here		0

So it removes exactly the covered outputs, a quarter of them on these graphs, and loses no interval. The store's derivation tests pass against a migrated database here as well.

One thing to know, not to change: a removed narrow output has already taken a slot of the per-predicate cap, so a predicate can report as capped with fewer than 20,000 outputs left. That errs on the side of saying so. Landing it.