Fix the direct-import-vs-transitive-leak binding defect (#6663 x #6686 main-red root fix) + fmt-immune crate-layout emitter

[gunbai-bot]

What this PR is now

The root fix for the #6663 × #6686 main-red (this PR began as a revert of #6686; main then moved — #6696 landed a colliding literal-flattening unbreak — so the revert/re-land pairs cancel and the net diff is the root fix alone, 6 files):

1. overlay_direct_import_exports (src/v1/04_infer.dag + regenerated v1_compiler_infer.rs): after the ancestry cache union in build_type_env, re-apply each direct import's own export surface (interface.env.str_bindings — locals only, never its ancestry), restricted to the names the import statement makes visible. Effective precedence, nearest first: locals > kernel > direct-selected > transitive union.
2. 5-arm receipt test (direct_import_precedence_over_transitive_leak_test): direct-selected beats a later transitive leak (selective + is_all); the field-presence wall still refuses the true shape's missing field; kernel names are never overridden; the binding-fork ledger still records conflicts.
3. Drift-gate deadlock fix: stage0_crate_layout_emit.dag stamps #[rustfmt::skip] on the generated consts so the wet emitter's bytes are the fixed point of cargo fmt (Wave 2 derived crate layout (GATES the wide fan-out) — make frontier.dag the single crate-organization authority: derive the emitted-module registration (pub mod list in 05_emit_rust hand_maintained_mods + regen_stage0 HAND_MAINTAINED_STAGE0_FILES) from the SelfEmitted frontier set, killing the §3 h #6677's artifact had two writers that could never agree: committed = fmt-clean/drift-red, regenerated = drift-clean/fmt-red; batch 4 first ran on this PR because main broke before ever reaching it).

The defect

union_parent_type_env_caches folds each direct import's whole flattened cache (own exports and inherited ancestry) with import-order overlay-wins, so a later import's transitively-leaked homonym silently beats an earlier import's explicitly selected export.

• Original detonation: import v2.std.algebra { Monoid } lost Monoid to dag-root std.algebra riding v2.std.node's ancestry (Ground fnv1a64 hash fork: unify v2.std.node.Hash with std ContentHash #6686's std.types edge) → v2 memory-control audit + witness-realization plan + P0 field wall (568→45 drained) #6663's presence wall demanded the wrong shape's fields at 28 sites → every main push red at plan resolve.
• Same defect, opposite direction, after §3: consolidate md_helpers + flatten v2 algebra literals onto std.algebra #6696: logic.dag's honest import std.algebra { BooleanAlgebra } loses to v2.std.algebra's tower-shaped homonym riding a later import's cache — on main today the Ground fnv1a64 hash fork: unify v2.std.node.Hash with std ContentHash #6686 leak accidentally re-shadows it back, which is why main is green only by an order-dependent accident. This fix makes §3: consolidate md_helpers + flatten v2 algebra literals onto std.algebra #6696's repointed imports bind deterministically, by declaration.

Ruled semantics preserved (2026-07-11 lane ruling): peer-direct forks keep import-order overlay-wins and stay ledgered on binding_forks; the kernel scope layer stays positionally above imports (kernel_type_set, containers List/Set/Map/Witness, Optional/Present/Absent/Unit) — builtin and algebra-profile typing (first, ends_with) grounds on kernel identities, and the v2 substrate models of those names are the known dual-representation interim. Fork observability is untouched; only which binding serves lookups changes. Dissolves with namespace Rule-1 (the union leak itself disappears).

Receipts (by execution)

• RED control: pre-fix seed binary on the re-land tree → the identical 28 missing required field errors from main run 29461044828; post-fix binary, same tree → clean. Only the binary varies.
• Whole-tree compile-clean on this merged tree (main + §3: consolidate md_helpers + flatten v2 algebra literals onto std.algebra #6696 + Ground fnv1a64 hash fork: unify v2.std.node.Hash with std ContentHash #6686 edge + fix): gunbc compile --source-root dag --source-root src/v2 --target dag → 0 diagnostics. (First fix cut had 11 errors from kernel/container shadowing — that iteration is why the kernel guard and its receipt arm exist.)
• Fork-ledger arms: existing transitive_interface_binding_test untouched and green.
• Self-host fixed point: regen_stage0 --verify → regen_divergence_count=0.
• Drift gate: wet regen (main_wet) leaves the tree byte-clean; cargo fmt --check green on the emitter's own output.

🤖 Generated with Claude Code

[gunbai-bot]

CI green on head ada3d2a: run 29466927103 (build ✓, ci floor ✓ — whole-tree compile-clean + batch-4 drift gate both pass). Merge bar met: 2 distinct dashboard approvals on this head (review 38626, review 38630), no request-changes, MERGEABLE. Ready for operator merge — landing this un-reds main (it carries both the binding-precedence root fix and the #6677 drift-gate deadlock fix).

Suite-triage note for context: the 6 local cargo-test failures on this tree are all pre-existing (bisected by execution: the 3 pipeline wire-contract tests red since #6663's field wall — they assert the pre-wall fail-at-emit contract; eval_measurement_purity + std_types_freebie red at #6685 already; claim_batch one is a local-binary environmental). None gate CI — nextest was removed from CI by the 2026-07-11 ruling (ci_spec ci_fmt_gate_note); fix-up belongs to a wall-lane follow-up, not this PR.

— sent from vivid-owl-803