Carry the grammar digest and production index on the prepared grammar; share the parse-witness fixture trees warm

[gunbai-bot]

Symptom (so a symptom search lands here)

Main's required floor is red: ten identities refused INTERRUPTED-BEFORE-VERDICT at the 500ms per-claim CPU ceiling (cpu_at_least=506-610ms, right-censored) on every main head since #10946 / #10961 (runs 34556215902, 34556223550, 34560949186):

• v2.test.parse.d1_declaration_grammar_parse.* (7)
• v2.test.parse.where_refinement_clause_parse.where_refinement_clause_parses_and_is_carried_holds, …survives_normalize_graft_holds
• v2.test.claim.body_lowering.declaration_structure_preserved.where_refinement_clause_interior_is_not_retained_holds
• v2.test.parse.expression_bodied_fn_decl_parse.braced_fn_decl_survives_normalize_holds

claims_failed=0: no floor run has ever reached their verdict. (Run 34560218821 at the same sha is the fleet-desired workflow, not the floor — there is no flipping head.)

Cost, measured (the interrupted figure is a censored lower bound)

where	method	the ten, per claim
uncontended amd64 EPYC (BuildBuddy runner, load ~1), main bf56299d	claim_batch --entry <file> --functions … --eval-budget-ms 60000	318–523ms
the lane's GH runner, same run 34556223550	the 8 file-mate rows that DID complete: 353–482ms there vs 187–292ms on the EPYC → 1.4–2.1× (median ~1.6×)	extrapolated ~510–840ms
arm64 session host, load 60–78 (contended by other sessions' builds)	same claim_batch	266–705ms

Structurally over, not marginal.

What they spend it on — required_floor_cross_claim_demand.tsv of run 34556223550 names it:

• parse_grammar_digest keyed, claims=19 evals=19 total_ms=2424 → ~128ms per claim, plus grammar_to_node ×2 per claim (~68ms each) — a content hash over the whole grammar node tree and an O(P) production index, rebuilt on every parse_module_prepared call although the grammar was already the warm-shared dag_prepared_grammar. This is exactly the finding Share the native_decl_selection fixture at the resolved seam so its four witnesses fit the 500ms floor #10992's write-up flagged for a separate lane.
• Under it, one module-constant source tokenized+parsed by 2–4 claims of one file, once per claim frame (d1_tokenize_parse ~480ms inclusive per call on the runner; lex_try_rules ~97ms per claim).

Cost shape answer for root: the same class as #10992 — a shared precondition re-derived per isolated claim frame — third instance (native four → these ten). #11015 filed that failure-mode row and was closed unmerged by the operator at 02:52Z; not re-minted here, flagged to sunny-tern-813.

Repair (both homes; the line is untouched)

1. v2.compiler.parse — GrammarFirstAnalysis now carries grammar_digest and productions_by_name, computed once in compute_grammar_first_analysis_given (so once per prepare_grammar) and read by parse_table_for_production_analyzed. The grammar's own locus, read only by the rejection arm that carries no diagnostic of its own, is derived on that arm alone. Same digest value, paid once; every claim corpus-wide that parses .dag text inherits it. Not a parser shortcut — no production, no table, no acceptance changes.
2. The three parse-witness files share each multi-consumer source at the parse tree through nullary producers (d1_where_reproducer_parsed ×4 consumers, d1_where_bare_parsed ×3, d1_where_control_parsed ×2, d1_cache_identity_where_parsed ×2, where_refinement_bare_parsed ×2, expression_bodied_fn_parsed / expression_bodied_literal_fn_parsed / braced_fn_control_parsed ×2 each), enrolled warm in v2.workflow.floor_pure_producer_share with the rationale beside the row. Normalize stays per claim so every normalize claim still executes the stage it names. Single-consumer sources (the refusal probes) are not shared — a single-consumer warm row is a budget relocated, not a recurrence removed — and the one three-source conjunction where_refinement_clause_parses_and_is_carried_holds is split into three rows (…brand_clause_parses_holds, …range_clause_parses_holds added; nothing removed).

declaration_structure_preserved is not edited: its row parses one source in one claim and is repaired by (1) alone (363 → 239ms on the EPYC).

Evidence

• After (1) alone, same EPYC host: the ten at 101–460ms, all under the line (d1: 343/279/207/242/385/339/352; wr: 460/264; eb braced-normalize 183; ds where 239). (2) then removes the remaining ~130–170ms tokenize+parse from each multi-consumer row.
• All 37 claims across the four files + pure_producer_share_refusal_test (the live-roster collision witnesses) PASS under claim_batch on this branch.
• Positive control, one per family, by execution: with the where_refinement_clause optional removed from type_variant and the = expr arm removed from fn_decl (grammar then restored byte-identical to HEAD): d1 cause_a_where_clause_parse_tree_carries_holds FAIL (control …where_control_normalizes_holds PASS, …empty_where_clause_refuses_holds PASS); wr where_refinement_clause_parses_and_is_carried_holds FAIL (…empty_clause_is_refused_holds PASS); eb expression_bodied_fn_decl_parses_holds FAIL (braced_fn_decl_still_parses_holds PASS); ds where_refinement_clause_interior_is_not_retained_holds FAIL (its three siblings PASS).
• The instrument that decides it is this PR's own required-witnesses-floor run: [floor-shared-fill] showing the eight fills landing at preparation with disposition=Stored, joined against required_floor_claim_cost.tsv showing the ten identities at verdict_reached=true with an observed cost under 500ms.

Foreclosed / not done

No ceiling change, no floor_cost_debt row, no expected-red or quarantine, no narrowing of any assertion, no D1 / where-grammar extension.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YApoxsmnC96ad2nWpRJdja