Derived-node identity step 2, repair (a): peel_nominal_alias_identity keeps a type parameter's mark (census 37 -> 0; plan specimens unmoved)

[gunbai-bot]

Derived-node identity step 2, repair (a) only: v1.compiler.infer_resolve peel_nominal_alias_identity keeps a type parameter's mark. Plan: docs/plans/derived-node-identity-design.md §6 step 2.

What it fixes

peel_nominal_alias_identity looked a type-parameter reference up by name and returned whatever the environment bound to that spelling. Where a module binds the same spelling (a record Box<Q> beside fn f<Q>), it replaced the parameter leaf and dropped its own mark. The repair reads the mark instead: a node carrying a TypeParameter declaration is returned as given (std.decl_ref declaration_ref_is_type_parameter).

• Measured: the generic-identity census (gunbc test //gunbc/instruments:generic-identity-census, Derived-node identity step 2: the generic-identity census (gunbc test //gunbc/instruments:generic-identity-census) #13290) goes from 37 rows of this shape to 0, and no other row moves.
• Gating RED: control conformance_keeps_the_parameter_mark over fixtures/generic_identity_census/a.dag (same_spelling_as_a_record_parameter). It fails before the repair and holds after.
• The branch also carries the step-2 bet readings (bet 2: substitute_generics_apply keeps the child declaration; the plan-field reading). They are readings in the census label, not gates, run over nodes the real compiler emitted.

Evidence at 69d839c (gunbc / claim_executor built from this tree)

• Next generation: claim_executor --regen-round-cost gives convergence_stages=0, rebuild_packages=0, and leaves the mirrors unchanged.
• src/v2/compiler/00_compile.dag resolves (NoSuchFunction). The four fold_list rows an earlier attempt newly refused do not refuse.
• Census label: STANDING held, all 7 controls held.

The plan's specimens: both UNMOVED, and not claimed

Both are shapes of gunbc.recurring_failure_mode generic_identity_decided_by_spelling, and both stay with the rest of step 2 (carry origins, switch the three sites to identity). The receipts below come from gunbc built at main ed68b5573b and at this head:

specimen	main ed68b55	this head 69d839c
srv3 test.claim.srv3_websocat_sequence_witness the_readback_never_runs_after_a_failed_step (shape 1, valid code refused)	refused: function 'join' not found in scope (122:34)	refused, identically
take_s(s: head_of(xs: [1], d: 2)), take_s(s: String) (shape 2, wrong code accepted)	Accepted, returns Int 1	Accepted, returns 1
control: the same call with String arguments	returns a	returns a

Shape 2 is silent wrongness, and it is the next priority in step 2. This PR does not touch it.

🤖 Generated with Claude Code

[briansrls]

LAND. This is the correct repair-(a) cut. peel_nominal_alias_identity is an alias operation, so a node already carrying a DeclField::TypeParameter identity is not eligible for name-based nominal lookup; returning that node unchanged preserves the existing declaration authority and copy law rather than adding another inference rule. The predicate lives beside DeclarationRef, and the old path remains unchanged for every non-parameter reference.

The fixture is discriminating at the affected carrier: a function parameter Q shares a spelling with the record parameter Box<Q>, and formal_declaration_bound_conformance must retain the function parameter's exact (owner, TypeParameter) mark. The reported 37 -> 0 movement confined to that carrier, fixed-point regeneration, and successful 00_compile resolution support the claimed blast radius.

The PR also states the ceiling honestly: neither the srv3 join refusal nor the accepted head_of<Int> into String specimen moves, and this review credits neither as repaired. The supplied-node bet readings are observations, not gates or production behavior.

GitHub's live status currently still shows the exact-head witnesses run in progress despite the request saying green; treat this approval as the code/modeling verdict and queue only after the exact-head required run completes successfully.