feat(build): warn on hub-node degree collapse during incremental merge (#1652b)

[TPAteeq]

Implements only sub-proposal (b) of #1652 — a god-node degree-drop ALERT in build_merge. It also doubles as the cheap safety net for the silent data-loss vector in #1651 (a hub node collapsing from many edges to ~0).

The gap this closes

build_merge REPLACES each re-extracted file's prior contribution: every node/edge whose source_file is in the re-extracted set is dropped before the merge rebuilds. If a re-extraction emits a different id for an entity that already exists as a well-connected hub, the old hub and its in-file edges are dropped and its cross-file edges are orphaned onto a bare node. The hub silently collapses (e.g. 174 edges → ~0) — and the total node count can even rise, so the existing count-based shrink guard (#479) never catches it. That guard is also gated on not dedup and not prune_sources, so it never fires on the normal CLI --update path (which always calls build_merge(dedup=True)).

What this does

• Snapshots the pre-merge graph's hub degrees (nodes with degree >= HUB_DEGREE_MIN) from the graph as loaded, before the replace-per-source filter runs.
• After the merge, compares each former hub's degree in the final graph and emits a clear stderr WARNING (styled like the existing prune warnings) when a hub vanishes or loses more than DEGREE_DROP_FRAC of its degree, reporting the label and before → after edge counts. Many drops aggregate into a concise, capped summary.
• Runs on every path, including dedup=True — unlike the count-based shrink guard.
• Warns, does not raise. A genuine large refactor can legitimately shed a hub's edges, so aborting would be wrong.

Thresholds are module-level constants in build.py: HUB_DEGREE_MIN = 20, DEGREE_DROP_FRAC = 0.5 (plus a display cap _HUB_DROP_REPORT_LIMIT = 10).

Explicitly out of scope

• Feature: build_merge dry-run, god-node degree-drop alert, and pre-write backup #1652's dry-run (a) and pre-write backup (c) are separate and not implemented here. (Note: a pre-write backup largely already exists via backup_if_protected.)
• Feature: first-class canonical entity registry (aliases/trigrams) consumed by extraction, build_merge and query #1651's canonical-entity registry — the real fix for the id-drift root cause — is out of scope. This PR is only the cheap detection net.

Tests

New tests/test_build_merge_degree_drop.py:

• a merge that collapses a hub (re-extraction emits a new id) triggers the warning on the dedup=True path;
• benign re-extraction that preserves the hub's id + edges does not warn;
• a small sub-threshold degree change does not warn;
• a graph with no prior hub does not warn;
• plus helper-level unit tests for _hub_degrees / _hub_degree_drops.

Command: uv run python -m pytest tests/test_build_merge_degree_drop.py tests/test_build.py tests/test_build_merge_hyperedges_and_prune.py tests/test_merge_graphs_cli.py -q → all passing (6 new + 46 existing).

Refs #1652