fix(watch): evict legacy AST ghost nodes lacking the _origin marker on full rebuild

[robertgarcia]

Problem

On a large Nx monorepo, a full _rebuild_code / graphify update leaves tens of thousands of legacy AST nodes stranded as permanent zero-degree orphans. On one real repo this was ~52k ghost nodes — 35% of the graph — which inflated the community count (74k "communities" for 200k nodes) and drove average cohesion to ~0.01.

Each ghost is a duplicate of a real symbol: same source_file, same source_location, same label, but a different (older-scheme) id, and no edges. Example pair from the affected repo:

apps_aga_src_app_app_config                     label="app.config.ts"  degree=12   (canonical)
apps_aga_src_app_app_config_ts_app_app_config   label="app.config.ts"  degree=0    (ghost)

Root cause

Two reconciliation layers both miss these nodes:

1. _reconcile_existing_graph (watch.py) — the AST-ownership eviction rule gates on _origin == "ast". Nodes written by a graphify version that predates the _origin marker (graphify update leaks stale nodes: symbols removed from surviving files are never pruned #1116), or that predate a node-id-scheme change, carry no marker, so they escape eviction and are preserved.
2. build_from_json's ghost-merge (build.py) — reconciles ghosts by (basename, label). In a monorepo this key is ambiguous: apps/aga/.../solicitud.component.ts and apps/se/.../solicitud.component.ts both define SolicitudComponent, so two fresh AST nodes share (solicitud.component.ts, SolicitudComponent). The merge conservatively marks the key ambiguous (Ghost-merge picks an arbitrary AST node when two same-named files define the same symbol #1257) and refuses to merge — so the legacy ghost survives.

The result: a marker-less legacy node whose file is re-extracted on every full rebuild is never reconciled, and accumulates forever.

Fix

_reconcile_existing_graph now also evicts a preserved node when the fresh AST pass re-produced the same symbol under the canonical id, matched by exact (canonical source path, label) identity (via source_paths.identity(...), robust to relative/absolute path spelling). This is self-scoping — the fresh output only contains re-extracted files — and does not depend on the _origin marker, so legacy graphs self-heal on the next full rebuild.

Genuine semantic nodes on the same file are preserved, because they carry a distinct label that the AST never emits (a same-label collision is exactly the case build_from_json already treats as a ghost, so behavior stays consistent).

Reproduction (before the fix)

Two same-named files defeat the (basename, label) ghost-merge; a marker-less legacy node then survives a full rebuild:

rebuild ok: True
legacy ghost survived (BUG if True): True
semantic node survived (should be True): True
nodes for app_a SolicitudComponent:
  [('app_a_solicitud_component_solicitudcomponent', 3),
   ('app_a_solicitud_component_ts_solicitud_component_solicitudcomponent', 0)]   # orphan

After the fix:

legacy ghost survived (BUG if True): False
semantic node survived (should be True): True
nodes for app_a SolicitudComponent:
  [('app_a_solicitud_component_solicitudcomponent', 3)]   # canonical only

Tests

• New regression test test_rebuild_code_evicts_legacy_ast_ghost_without_origin_marker in tests/test_watch.py reproduces the monorepo collision and asserts: ghost evicted, semantic node preserved, canonical node keeps its edges.
• tests/test_watch.py — 59 passed, 2 skipped.
• tests/test_incremental.py + tests/test_build.py — 53 passed.
• Ghost/reconcile/id-related suites (test_build_merge_hyperedges_and_prune, test_semantic_id_remap_root, test_semantic_cleanup, test_obsidian_dangling_member, test_extraction_spec_ids, test_id_normalization_contract, test_community_hub_labels) — 99 passed, 1 skipped.
• ruff check clean on changed files.

Scope / safety

• No behavior change when the fresh AST output is empty (set is empty → clause is a no-op).
• Incremental updates unaffected: an unchanged file's nodes aren't in the fresh output, so they never match.
• Only touches graphify/watch.py, tests/test_watch.py, CHANGELOG.md.