F-0b: four residual false-reference syntax roles excluded structurally; type-reference channel unioned at every wall consumer; sixth admission shrink

[gunbai-bot]

F-0b of the NAMESPACE-XL chain — the second half authorized by F-0's (#9698) item-5 census: the wave-wall syntax-role matrix re-run against current main found four false-reference roles surviving #9430/#9504, plus a binding-channel asymmetry. Both are real defects, fixed here with discriminating fixtures. Also carries the time-critical sixth admission shrink (#9698's two rows went stale the moment it merged and were refusing every PR's namespace phase).

The four roles

The collector's own doc named these as deliberately-not-excluded, each needing a structural discriminator derived against its own fixture (a guessed parent-kind rule would blind the wall to genuine references sharing the position). Each now has one, in collect_reference_occurrences:

role	discriminator	genuine reference kept
record TYPE decl field label	the field_to_child_node shape: declared type parked in inferred, no children/params, NoExprData, NoConnective, authored ident	a refinement base type (leaf_type_node, inferred: None) stays collected
named call-argument label	parent kind ExprCall, same rule as the existing ExprRecordLit	argument values walked; callee spelling on the call node itself
parameter binder	the params edge passes the binder flag, consumed at one level	the param's declared type is children[0], still collected
coproduct variant names in the declaration	parent Connective::Disj (set only by the two coproduct item builders)	variant payload fields still walked; the names remain on the export surface via variants

The binding-channel asymmetry

membership_bound_through already unioned authored_type_references (the parser-stamped channel that reaches declared types parked in inferred, invisible to the tree walk) — but direct_membership and binding_rows did not. So a module whose only reach into another was a declared field/param/payload type contributed no membership edge (closure and blast radius under-reported) and no binding row (a cut repointing a declared type moved nothing on either side of the wall). Both now take the same union. The role exclusions above make this union more load-bearing, not less: the genuine reference beside each suppressed label lives exactly in that channel.

Evidence — every pair executed both directions

Eight new test pairs in tests/namespace_wave_admission.rs, each the #9106 shape: the same deletion of the same spelling, reached once from the label/binder (must carry no delta) and once from a genuine reference in the same syntactic neighbourhood (must stay NewUnresolvedness). All 41 suite tests green with the fix; against the unfixed collector the five discriminating label/union arms go RED (BuildBuddy, both arms in one dispatch, byte-restore between). Two earlier drafts of the variant arm were measured as decorations and discarded — a same-module supplier only ever produced the self-candidate, an unimported cross-module supplier never entered the candidate set — before the blanket-import fixture made that arm's RED reachable; the falsifier run proving red-without/green-with is in the receipts.

The declared-field-type and variant-payload-type twins discriminate two repairs at once: they go quiet if either the label exclusion over-widens or binding_rows stops unioning the parser channel.

Sixth admission shrink

#9698 merged, so its two RequiredCiLane rows match no producible delta and report stale — refusing every PR, the same PR-only-but-universal shape as the four prior shrinks. Removed by the trigger they were authored with; the roster is empty again, and empty is not permissive.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GpYEk9gSYmMkwL1489D79U