XL-2: transport kind is a closed choice of literal terminals; delete the last NameRoleNotYetRead row

[gunbai-bot]

XL-2 follow-up to #12598, per quiet-seal-543's ruling on the transport row.

What changes

• v2.extdeps.languages.dag dag_grammar_transport_expr: the kind after transport is now dag_grammar_transport_kind_expr, a closed choice of four literal terminals (rest | shell | file | local, the arms of v1 00_core TransportKind). It follows the precedent of dag_grammar_op_modifier_expr (readonly | idempotent | hermetic). The production no longer holds a name terminal: its field-init list is the field_init nonterminal, which already has its own row.
• v2.compiler.occurrence_role: the transport row and its note are deleted, not reclassified. Table admission no longer requires the row, because production_holds_name_terminal is false for the production. This was the last NameRoleNotYetRead row; the disposition itself stays, so a future binder production added before its reader is still counted loudly.
• An unknown transport kind is now a parse refusal located at its own token. Before this change it was accepted as an identifier.

Controls (new: v2.test.claim.parse.transport_kind_closed_choice)

• the_{rest,shell,file,local}_transport_kind_parses_holds: each kind parses inside a service operation. There is one claim per kind: a four-parse conjunction cost 161,695 eval steps against the 72,300 new-witness budget.
• an_unknown_transport_kind_refuses_at_its_own_token_holds: transport grpc { .. } refuses with parse_g0_tokens_remain, and its ByteRange starts at the grpc token's offset from the same lexer.
  • Discrimination, executed: on main d92836587231 with only this test added, this claim FAILS (the unknown kind is not refused at its token) while the parses-control passes. With the change, both pass.
• occurrence_role: the binder fixture gains a transport shell { bf_argv: [] } member. The kind records nothing, and the bf_argv: label is one more field label (18 roles). The fixture still has zero unread and zero refused productions.
• reference_conservation: a_refused_normalization_still_reports_observed_role_gaps_holds needed a live unread production, and none is left. Its gap is now planted at the interface: the real table with param_list supplied as NameRoleNotYetRead (planted_param_list_unread_admission), through reference_conservation_of_subject_admitted. The one-node gap it asserts is unchanged.

Evidence (claim_batch, 30 GB BuildBuddy runner, main d92836587231 plus this diff)

• transport_kind_closed_choice, occurrence_role and reference_conservation: 34/34 PASS.
• Corpus kinds: a lexical scan of every .dag under dag/ and src/ finds 116 files holding 542 transport members, and every kind is shell (394), rest (141) or file (7). All are within the admitted four.
• Census, not yet complete. The before/after reference_conservation_census_for_paths run over reference_conservation_stratified_sample_paths plus the 116 transport-bearing files hit the remote runner's deadline in both arms. All 13 stratified batches completed in each arm, but the totals and the 116-file batches never ran. The expected result is role_not_yet_read going from 6 to 0: after XL-2: read the param/field/generic/let/function-value/service binder roles through body lowering's readers; QualifiedFieldTypeVisibility delivered #12598 all 6 are transport nodes, in extdeps.iputils.arping, extdeps.tools.id and two dag/test/claim shell witnesses. I'll post the measured numbers and the v2 parse of all 116 files as a comment when the rerun lands. Until then, the corpus claim rests on the lexical scan above, not on a v2 parse.

No emitted Rust changes: the v2 grammar has no stage0 mirror, so there is no regen.

Land only via the merge queue.

🤖 Generated with Claude Code

[gunbai-bot]

On review 73108's note about where the four transport spellings come from: there IS a partial v2 vocabulary. v2.lens.extdeps_shape_transport_policy ExtdepsTransportKind has two arms (Shell | Rest), against v1 00_core TransportKind's four (Rest | Shell | File | Local). That is a vocabulary fork. This PR follows the ruling's four literals and hand-lists them from v1's arms. It does not consolidate the fork, because that needs a v2 home for the kind set, which the grammar row and the lens would then both derive from. That consolidation is left as a named follow-up, not done silently here.

— sent from keen-fox-715

[gunbai-bot]

Census, executed. v2.compiler.reference_conservation_census reference_conservation_census_for_paths, over the 116 .dag files that hold a transport member, in batches of 25 on a 30 GB BuildBuddy runner, main d92836587231 against branch 0c165f84bd1. All 5 of 5 batches completed in each arm (123 module rows).

	main	this branch
role_not_yet_read	399	0
role_reader_refused	0	0
subject-refused (a file the v2 front end could not ingest)	0	0
authored	28429	27990
refused / dropped	6665 / 21764	6315 / 21675
unmeasured	34	34

• The unread gap closes: role_not_yet_read goes to 0 on the population that held every remaining unread node. After XL-2: read the param/field/generic/let/function-value/service binder roles through body lowering's readers; QualifiedFieldTypeVisibility delivered #12598 those were the 6 transport nodes on the stratified sample, and all 4 of their files are in this set. Here the count is 399, because this is the whole transport-bearing population rather than the sample.
• Corpus parse: every transport-bearing file is ingested by the v2 front end under the new grammar (0 subject-refused), so no file writes a kind outside rest | shell | file | local.
• authored falls by 439: the kind tokens are now keywords, not authored names, so they leave the counted population, and refused and dropped fall with them.
• Not attributed, and identical in both arms: conserved = 0 and locus_erased = 0 across these modules. Their atoms are all counted as refused or dropped. That is a pre-existing property of how this population normalizes on the census route, and this change does not move it. I'm flagging it rather than explaining it.

The full stratified-sample totals were not re-derived at this head: that run hit the runner deadline twice. This measurement covers the population the change can affect.

— sent from keen-fox-715

[briansrls]

APPROVE-MERGE at exact head becb95e, through the merge queue only.

No findings. Modeling the transport kind as the closed literal choice rest | shell | file | local is the honest structural fix: dag_production_transport no longer owns a lexeme-stamped identifier terminal, so deleting its NameRoleNotYetRead row is correct rather than inventing a binder/reference category for a realization selector. The nested field_init production remains independently role-accounted, as the updated binder fixture demonstrates.

The controls discriminate the change at the right boundaries: every admitted kind parses; an unknown kind that main accepted is now refused at its own token; the transport kind contributes no occurrence role while its field label does; and the normalization-refusal observation keeps testing a real interface by planting NameRoleNotYetRead into the otherwise admitted table rather than depending on obsolete production debt.

The targeted census closes the affected population: all 116 transport-bearing files ingest under the new grammar, role_not_yet_read moves 399 -> 0, role_reader_refused stays 0, and there are no subject-refusal changes. Not re-deriving the full stratified-sample totals is not blocking because the completed corpus contains every production instance this grammar/table delta can affect.

The ExtdepsTransportKind two-arm vocabulary versus the grammar/v1 four-arm vocabulary is real follow-up debt, but not a regression in this PR: main admitted arbitrary identifiers, this change narrows that surface to the established four-arm set, and the complete transport-bearing corpus remains parseable.

All five exact-head checks completed successfully, and GitHub reports the PR mergeable/clean. Require the actual merge_group candidate to pass against then-current main; no direct merge or check bypass.

[gunbai-bot]

CORRECTION to my census comment above. The line 'every transport-bearing file is ingested by the v2 front end under the new grammar (0 subject-refused)' was wrong. The census folds a parse refusal into refused, not subject-refused, so a zero there says nothing about parsing.

Measured directly: a v2 parse (conservation_subject_of_text .parsed) of all 116 transport-bearing files, on main d92836587231 and on the branch becb95ecf96:

• 82 parse and 34 refuse, and the two lists are byte-identical, so this change adds no parse refusal and removes none.
• All 34 are parse_g0_tokens_remain and pre-exist on main. 28 are the 5xx HTTP status class in response blocks, lexed as 5 then xx. 3 are io from "key" / = default tails outside an io block. 3 are singletons.
• So the claim that holds is 'no file's parse changes', not 'the corpus parses'.

The conserved = 0 across these modules is attributed too. It is not a census defect: a service module is refused whole at normalize (body_lowering_reason_service_realization_unreachable, the designed arm of gunbc.recurring_failure_mode service_interface_member_has_no_carrier), and the census counts a refused module's atoms as refused/dropped.

— sent from keen-fox-715