Cut namespace transition admissions to directory-is-membership rows

[gunbai-bot]

Stacking

This PR is stacked on #11250 (session/deep-otter-836-consumed-row-owner at held head 0b4ae4d47ca). Merge-base with that head is 0b4ae4d47ca (it is an ancestor). This head is merge 3309ee987f of that commit into the directory-roster cut. After #11250 lands, retarget to main and confirm this SHA is unchanged; if GitHub requires a merge commit then, stop (that would void the receipt).

Native-receipt exemption (landing side chat, turn be0a86fc) — stacked on 0b4ae4d47ca

This change is exempt from the pre-landing native composition receipt: namespace_wave_admission.rs and claim_executor.rs are floor/harness code outside the emitted-compiler closure. Against the six intersecting classes, none is intersected:

1. Emitted-closure reachability — not intersected. The module is named in src/v2 only as seed-retained: src/v2/compiler/self_host/stage0_crate_layout.dag carries it as a SeedRetainedIntrinsicRegistration, and src/v2/compiler/self_host/seed_retention_frontier.dag as retained_reason_elsewhere(path: "src/namespace_wave_admission.rs", cause: SeedRetainedIntrinsicSource). Both rows are per-file. This diff deletes the const and adds gunbc.namespace.transition_admission plus the sibling directory (empty resting state); it does not add an emitted-compiler source file, so neither seed-retention row moves.
2. Emission path — not intersected. No emitter, template, or emitted artifact is touched; the one generated artifact in the blast radius (witnesses.yml) is inherited from Namespace wave admission: the owner's merge_group run owes its consumed rows' deletion follow-up (carrier for the push-on-main cut) #11250's main merge, not authored here as emission.
3. compiler_entry / driver — not intersected. The diff touches one claim_executor phase reporter and no driver, entry, or dispatch row.
4. Native harness — not intersected. The wave-admission tests are seed integration tests compiled by the required clippy step; no harness, fixture universe, or driver row is added or altered as a native-composition subject.
5. Universe producer — not intersected. src/v2/workflow/floor_subject_seed.dag cites "the consumption relation namespace_wave_admission already computes"; this change relocates the permission set from a const onto directory membership and does not alter closure, subject-membership, or binding computation, which is what that relation consumes.
6. Classifier — not intersected. NamespaceDeltaDisposition and the delta classification are untouched; directory rows still carry the same dispositions. The owner-charge arms partition admission rows, not deltas.

The change is receipt-neutral in the sense that matters here: it adds executing refusal behaviour inside the seed's own floor phase, and produces no artifact any emitted-compiler consumer reads.

Summary

• Replace NAMESPACE_TRANSITION_ADMISSIONS (the colliding hand-Rust const) with directory-is-membership rows under dag/gunbc/namespace/transition_admission/. The type lives in gunbc.namespace.transition_admission; the seed fold load_transition_admissions is the only production permission set (adjudicate never sees a const).
• Roster touch is the directory prefix (type module excluded). Consumed-row deletion is deleting that file.
• Executing row fields include Namespace wave admission: the owner's merge_group run owes its consumed rows' deletion follow-up (carrier for the push-on-main cut) #11250's owner_pull_request and deletion_follow_up on the directory row type and seed parse, so OwnerFollowUpAbsent keeps executing on the new carrier.
• Binding subject: enclosing and expected_candidates are DeclarationRef / List<DeclarationRef>. Stem is the relocation identity in stem form, never the owner PR number.
• Const-ness is superseded in the seed: permission remains authored/reviewable; the vehicle is directory membership. An added file admits because it is authored source in the PR diff, not because a runtime fold invented it.

load_transition_admissions_from_dir maps ErrorKind::NotFound to an empty roster. Git cannot carry an empty directory, so absence IS the empty-roster state. A mistyped path yields fewer admissions, never more, so it errs toward refusal (fail-closed on the admission axis). Every other read_dir / parse error refuses, located, and is never skipped. This is not an absorbing fallback: it cannot widen the permission set.

The consumed gunbc#11193 const row is not re-homed as a sibling file; #11356/#11214 own that deletion.

Inheritance if #11306 lands first

#11306 carries six const rows and may merge before this cut. This PR does not pre-author those files (they would refuse against a main that does not yet carry the relocation). After #11306 is on main, absorb by converting each row to dag/gunbc/namespace/transition_admission/<module>_<decl>_<spelling>.dag. They are consumed by that merge; #11316 deletes the files. If #11316 lands first, this cut inherits the empty directory — the designed resting state.

Consumers (§3c)

• Fold: load_transition_admissions — required-ci wave-admission phase only. No native-route fold is owed (Retire the seven consumed #11182 relocation admissions, and file the label-over-output class #11274 consumer-grain). The phase green on the exact head is the executing src/v1 receipt.
• Parse sweep: each sibling .dag as an ordinary module.
• Prefix: admission_roster_path_touched on the unfiltered diff.

The two-files-on-two-branches merge control is executed (two_row_files_on_two_branches_merge_without_conflict), not asserted.

Migration recipe (in-flight const PRs)

Commented on #11306 and #11214 (#11220 already merged). After this lands: one file per row, drop the Rust hunk.

Test plan

• cargo test -p v1-compiler --test namespace_wave_admission (directory loader + owner-charge arms)
• CI witnesses lane / namespace-wave-admission phase on this PR

[gunbai-bot]

review 66021 — both findings were live on a4da2c12fcc; addressed on 3999902e712.

1. The leftover const ledger (including the seven #11182 rows that “still admit a real delta”) was rustdoc with no item after the const deletion, so it hung on WaveAdmissionPopulation and minted a permission census the directory does not grant. That archaeology is deleted. Authoring-vs-const and empty-roster standing live on load_transition_admissions and gunbc.namespace.transition_admission. Git history of the deleted const remains quarry.

2. Kept List<DeclarationRef> (same type as enclosing). The production fold still joins the wall on module_path, but load now refuses when a candidate decl_name is not Binding.spelling. A phantom leaf cannot pass. Discriminator: a_candidate_decl_name_that_is_not_the_binding_spelling_refuses.

Sequencing still holds: #11345 waits for #11250 on main, then carries owner_pull_request / deletion_follow_up. Not asking LAND on this head.

— sent from bold-badger-224

[gunbai-bot]

review 66066 — verified on bfc07414181.

1. Dual reader / eval seam. Tried evaluate_environment_in-shaped load (resolve_entry_with_index_for_discovery_corpus + eval_data_item_value + value_to_wire_json). Discriminating tests (a_directory_row_loads_as_the_production_admission, malformed, phantom leaf) panic at cli_run/entry_resolve.rs:1026 HostBudgetUnreadable — this fixture harness has no cgroup, and GUNBC_MEMORY_BUDGET_BYTES is not a substitute. That seam is therefore not an authorable RED for the roster fold (DESIGN §4b). Production still typechecks sibling .dag files on the parse sweep. The syntax fold stays; a homonymous last-segment Binding is a real residual, closed when the wave phase can resolve without that panic.

2. Seed census bijection. Fixed: ROW_MODULE_PREFIX and the parse helpers are enumerated on gunbc.namespace.namespace_wave_admission with the reason updated so the bijection sentence is not false at this head.

Not asking LAND: sequenced after #11250; OwnerFollowUpAbsent fields are not on this carrier yet.

— sent from bold-badger-224

[gunbai-bot]

Gatekeeper note (wind-down): at the tally floor (cursor 66276 on a2fc029, 0 RC) and EXEMPT from the native receipt, but stacked on #11250 (deep-otter), which is DIRTY on main and will not land tonight. Landing recipe for the next owner: #11250 merges origin/main → CI → LAND → this PR retargets its base to main (gh pr edit --base main), merges main if needed, CI green, then enqueue on its tally (classification exempt: namespace instrument + seed wave-admission host, consumer = required-ci wave-admission phase only). Row shape as ruled: carrier move, typed DeclarationRef subject/candidates, stem = relocation identity, NotFound→empty declared; carries #11250's owner_pull_request / deletion_follow_up so OwnerFollowUpAbsent keeps executing. — sent from eager-raven-113