fix(build_merge): re-extracted node must not lose to its own stale copy

[kellinkquinn-commits]

The bug

A re-extracted node can be silently replaced by the stale version it was meant to correct.

Minimal repro (added as a regression test): a legacy unstamped node for notes.md whose source_location is "L38-41", then a semantic re-extract of the same id with a corrected, longer label. Before this change the graph keeps the old label.

[graphify] note: node 'notes_rank' was extracted twice from 'notes.md' under
different labels — keeping 'Unreconciled: book-to-rank labels off by one between
C-23 and C-27', dropping 'Resolved: the C-23/C-27 book-to-rank off-by-one, closed
by C-45 mapping to Technical Ranks 2/3/4'.

Two things combine

1. The tier guess. _is_ast_tier's shape fallback reads any unstamped item with an L<line> source_location as AST-tier. Semantic extractors do emit that shape for markdown — build.py's own dedup path already guards against it, in a comment that says so explicitly:

Strict _origin check on purpose — NOT _is_ast_tier (#2334) ... Those may carry drifted 'L<line>' source_locations

So a legacy semantic node can be read as AST-tier. The tier-scoped replace from #2333/#2336 then does not drop it on a semantic re-extract, and it survives alongside its own fresh version.

2. The tie-break. With both present, _collision_rank picks the survivor. Definer-ness and source_file tie, so it decides on len(label) and keeps the shorter one. Freshness is not a factor, so a re-extract that expands a claim loses to the text it was correcting.

The failure is silent — node counts unchanged, no dangling edges, and the only signal is a note that reads like routine dedup.

Why not fix the shape test

.md is in the AST extractor registry (extract.py: ".md": extract_markdown), so a doc legitimately carries both tiers, and _write_two_tier_graph in the existing tests encodes exactly that. A semantic node with a line-range location is therefore indistinguishable from an AST one by shape alone. Tightening the heuristic would just move the misclassification.

The fix

Key on identity instead of shape: a base node whose id is re-emitted by a new chunk for the same source_file is that node's own freshly-extracted version, never a coexisting other-tier node. Drop it whatever tier each side is read as.

Tier scoping still protects the other tier's distinct nodes, which is what #2333 was about — an AST heading layer has different ids from the semantic concepts, so it survives untouched. The COEXIST tests are unmodified and still pass.

Testing

Full suite on Windows / py3.11:

	failed	passed
pristine 00efd6e	49	3681
with this patch	49	3682

Identical failure sets — no regressions; the extra pass is the new test. The 49 are pre-existing environment failures on this platform (skillgen, terraform, ollama_retry_cap, uninstall_scope, watch), unrelated to this path.

Also verified end-to-end against the real 1451-node graph where I hit this: with the patch the re-extract produces the corrected label and zero collision notes, because the replace now fires and the collision never happens.

How I ran into it

A knowledge-base corpus where wiki/ pages carry adjudicated claims. A re-extract of one page was supposed to update "Unreconciled: ..." to "Resolved: ... closed by C-45"; the graph kept the stale text. Nothing downstream flagged it — the node count was right and no edge dangled. In a corpus whose whole point is that adjudicated answers supersede earlier ones, a silent revert to superseded text is about the worst available failure mode, which is why I chased it to the tie-break rather than just working around it locally.