WIP: preserve generic application identity in inhabitance checks

[gunbai-bot]

Work in progress for dashboard node adhoc-0a02fa3a-821. This draft was opened automatically from the implementation checkpoint; it is not ready for review or merge.

The callable guard mistakes generic type parameters for function parameters. Correcting it to recognize Arrow lets the existing structured mismatch judgment reject Out<Int>/Other. That shared correction also exposes valid product-alias false refusals, so the implementation carries declaration-bound owner and argument facts to the existing inhabitance judgment. These are separate causal paths: the motivating coproduct is repaired by the guard; product routing repairs collateral alias comparisons.

Current checkpoints include the additive std.type_application carrier, authored formal/field capture, bare-marker checks, and regression fixtures. The latest source represents nongeneric projection bases at arity zero while keeping conformance routing generic-only. The first bootstrap regeneration now includes these source changes; convergence and final semantic evidence remain pending. No hand-edited probe mirror is an implementation artifact.

Validation so far is exploratory, not final acceptance: ref-bound baseline 03a3b91fb4c1 admits the foreign constructor and rejects the monomorphic control. Generated optional-carrier probe 87880de7-9bd7-45db-87dd-5df73f30beae rejects the foreign constructor, accepts Acc, and distinguishes cross-module homonyms. Later throwaway probes isolate formal/field syntax loss and recursive alias identity; they do not establish the shipped artifact. The arity-zero probe 26493831-ccf5-42fb-bdc9-e064508e05b8 passes output-policy admission and retains Out/Other refusal and Acc acceptance.

Remaining before ready: complete recursive argument identity and its bounded typed refusals; consume the modeled cycle/projection outcomes; verify valid generic calls acquire no unintended advisories; complete bootstrap convergence, including byte-for-byte reproduction of the interpreted root by the compiled emitter; execute final controls, deletion mutations and relevant corpus checks on one final tree; update failure-mode rungs to match that evidence. Equivalent nested aliases still falsely refuse in the old-comparator probe.

Three independent findings are recorded in #10567 (parser limitation, general alias-head cycles, and unit-variant generic argument identity). They remain separate from this implementation. The unit-variant population is not claimed repaired here. The #10473 compensation is retained; its later deletion requires independent refusal, a combined-tree deletion mutation, and coverage of every population it currently protects.
Rebase checkpoint c2bb5a61563080d04b0c40e260260dc5306bb009 includes main 1001fd03c74. Actual archived-seed generation 802f9c68 produced the nine installed changed mirrors (158 surfaces adjudicated). The current root producer was interpreted against the emitted-population manifest, and the resulting tree built all three release binaries successfully in 23e02049. No generated Rust was hand-edited. This is bootstrap/build evidence, not convergence or final semantic acceptance. The PR remains draft.

Declared return positions remain an explicit open boundary: wrap<T>() -> Out<T> { Zed { q: 1 } } is still accepted in the measured probe. The retained return-boundary witness names that defect; the discriminating payload control uses Holder<T>.r: Out<T> instead. No generic-return wall is claimed.

[gunbai-bot]

Closing this draft. The branch session/vivid-ferret-153 is retained at 9522a1b29f and nothing here is discarded — this records the disposition so the work is findable rather than left as a stalled lane.

Why closed. The authoring session produced no turn after 2026-09-05 and is being archived. More substantially, the measurement this PR rests on was taken against a tree that no longer exists: declaration_substitution_basis stamped TypeVariable over concrete declared types whenever a local binding lookup missed, so roughly 8,400 positions never reached declared_type_inhabitance at all. #10727 repairs that at the producer, and those positions now arrive. This PR's body states that its callable-guard correction "exposes valid product-alias false refusals" and then compensates for them by carrying declaration-bound owner and argument facts into that same judgment — a compensation designed against the suppressed incoming state.

What should be salvaged, as its own small PR against post-#10727 main: the callable guard itself. type_node_is_callable testing params |> count > 0 conflates a generic declaration with a callable, against a model whose make_callable_type uses Arrow. That is a genuinely independent defect and I expect it to survive the root repair unchanged.

What should not be carried forward without re-deriving it: the owner/argument compensation, and any population figure in this body. #10727 plus its refinement closure now occupies that seam. Re-derive against the merged tree before defending any part of it — do not assume the compensation is still load-bearing, and do not assume it is dissolved either.

The std.type_application carrier is unassessed either way.

The three independent findings recorded in #10567 (parser limitation, general alias-head cycles, unit-variant generic argument identity) are unaffected by this closure.