Graph extension list is written three times with no drift guard (anchor fixes two; extension_map needs a test)

[NoopDog]

Problem

The graph extension set is written out three times, and a fourth place holds a deliberate subset. Nothing enforces that any of them agree.

where	contents	kind
rules/unified_rules.yaml:218-225 — extension_map	8 keys mapping to pangenome	source of truth for "is this a graph file"
rules/unified_rules.yaml:492 — pangenome_graph.when.extensions	the same 8, byte-identical	rule guard
rules/unified_rules.yaml:626 — pangenome_reference_mc.when.extensions	the same 8, byte-identical	rule guard
src/meta_disco/header_classifier.py — GRAPH_TEXT_EXTENSIONS	4 of the 8	deliberate subset: the text formats the fetcher can parse (GFA_CONFIG.extensions reuses it)

Add a ninth graph format (.gaf, odgi .og) and you must edit three lists. Miss extension_map and the file has no category; miss pangenome_graph and it produces no base claim (not_classified); miss pangenome_reference_mc and a new-format -mc- graph silently stays pangenome instead of pangenome.reference. Nothing fails loudly in any of those cases.

Fix, in two parts — and the obvious half doesn't work

1. The two rule lists can be collapsed with a YAML anchor. Verified: yaml.safe_load resolves anchors, and both rules live in the same document (document #1; the file is loaded with safe_load_all and the first --- is at line 238). So this works with zero code change:

- id: pangenome_graph
  when:
    extensions: &graph_exts [".gfa", ".gfa.gz", ".rgfa", ".rgfa.gz", ".gbz", ".vg", ".gbwt", ".xg"]
...
- id: pangenome_reference_mc
  when:
    extensions: *graph_exts

2. An anchor CANNOT reach extension_map. It is in document #0, and YAML aliases do not cross document boundaries — confirmed:

yaml.safe_load_all('a: &x [1]\n---\nb: *x\n')  ->  ComposerError

So the extension_map ↔ rules relationship needs a test, not an anchor. That is the more valuable half anyway, because it is the edge where a miss produces a silently unclassified file.

Proposed guard

A drift test in the spirit of test_orchestration.py and test_stub_payloads_cover_all_file_types:

def test_graph_extension_lists_agree():
    graph_keys = {k for k, v in rules.extension_map.items() if v == "pangenome"}
    for rule_id in ("pangenome_graph", "pangenome_reference_mc"):
        assert set(rule_extensions(rule_id)) == graph_keys

def test_gfa_config_extensions_are_graph_extensions():
    """A deliberate subset — the text formats the fetcher can parse — but it must
    not contain an extension the graph rules do not know."""
    assert set(GFA_CONFIG.extensions) <= graph_keys

Both pass today (verified): extension_map pangenome keys are exactly the 8, and GFA_CONFIG.extensions is the 4-element text subset.

The GFA_CONFIG subset relationship matters in a way that is easy to miss: a text extension added to GFA_CONFIG but not to the graph rules would be fetched and parsed, produce no base pangenome claim, and then — if the content carries rGFA tags — receive a data_type: pangenome.reference claim on a record whose data_modality is not_classified. That is precisely the incoherent record #151's review found by another route.

Generalization (optional, larger)

when.file_format matches the extension string, and UnifiedRules.get_file_type() (which returns the extension_map category) has zero callers repo-wide. If _rule_matches resolved extension → category and rules could say when: {file_type: pangenome}, both rule lists would disappear and extension_map would become the single source of truth it already looks like. That is a rule-engine change touching every file type; the drift test is the cheap 90%.

Tasks

• Collapse pangenome_graph / pangenome_reference_mc extension lists with a YAML anchor.
• Add the extension_map ↔ rules drift test (anchors cannot express it).
• Add the GFA_CONFIG.extensions ⊆ graph extensions subset test.
• Consider when: {file_type: <category>} and giving get_file_type() its first caller.

Refs: #148, #151 (added GRAPH_TEXT_EXTENSIONS and deduplicated the Python side), #154 (fact model — extension becomes a fact, not a filename artifact).