Seed: admission precedes identity on the pure-call path; retire the eval-frame result memo

[gunbai-bot]

Handback (2026-09-20, lane merry-fox-260 wound down) — read this first

Standing: HELD as a draft under the parent's ruling (option A, root-first). Not landable in this shape; not stalled — the blockers are attributed, named and partly discharged. Remaining work is carried on the roadmap: gunbc#11859, row eval-memo-admission-rewire-program.

The change here: the seed's pure-call path consumes the std.materialization_ladder judgment before deriving any identity; the eval-frame result memo (EvalCallMemo, GUNBC_EVAL_MEMO, its provider row extdeps.realization.eval_memo) is deleted, the cross-claim tier and parse-table memo (both admitted-first) are untouched, the trace ledger keys only under GUNBC_RECOMPUTE_TRACE=1. A RED control (own test binary) reads the new per-frame identity odometer at zero and goes red when unconditional keying is restored. json.parse goes from quadratic (502 s at 1.09 MB) to linear (39 s), 24 receipts byte-identical.

Why it is held — the floor RED, by the change's own qualification: matched required-floor runs (base 35458211949 vs this head 35466872753): 2,767 / 4,161 claims cost more eval steps without the memo (p90 3.6×, max 116×); 76 grandfathered claims over the 361,500-step ceiling (base 0); 23 pass → budget_interrupted; 134 adjudication blockers. The corpus's witnesses were leaning on the memo for authored within-frame duplication — ladder rule 1, prescribed Share, never a cache — so the ruling refused both a declared-roster memo (B) and withdrawal (C): rewire the shared producers as their own qualified changes, then re-qualify the retirement as a combined cutover on an exact head.

What landed / is landing (each with its own before/after under both realizations):

• Stage 1 — Integer interval decode reads the module's data rows; the 2^n ladder and digit primitives are carried tables (stage 1 of #11741's rewire program) #11784: the integer interval decoder re-derived every standard type's 2^n ladder per decode; it now projects to the module's data rows and walks one carried ladder. First-touching claim 941k → 379k eval steps memo-retired.
• Stage 2 — Grammar backward selection resolves each emitted tree once (stage 2 of #11741's rewire program) #11835: the formal_production_* grammar backward selection resolves each emitted tree once (asked-lhs, top-down). 1.54B → 29.4M eval steps over 110 claims memo-retired (52×), +12% memo-on with no claim near its ceiling; the required floor ran green on it.
• The regen blocker is discharged by stage 2: the docs-projection regen that never completed on the memo-retired binary (>48 min, killed) completes on the stage-2 tree in 295 s with byte-identical projections. Its one non-completing eager data declaration, witness_floor_workflow, was this grammar family.

Remaining, in order (roadmap row): stage 3 — the rust.dag nullary node constants as data (drafted, unopened, preserved on session/merry-fox-260-stage3 at c1c925060c; it resolves and passes, measured 1.71M → 1.10M eval steps memo-retired on the binop catalog claim with memo-on flat — its nine-module before/after was not completed) and the bind_outcome chains; then the combined re-qualification of this head on the repaired corpus (matched floor: population and terminal results unchanged, no pass → budget_interrupted, regen completing). The eager data environment (build_initial_env evaluates every data in the closure on plain gunbc run) is no longer a blocker but is still undemanded work; its own reading is pending (is eager evaluation load-bearing for any refusal/disclosure/gate?).

Do not re-derive: (1) the producer ranking must come from the floor's own required_floor_cross_claim_demand.tsv (within-claim repeats = evals − claims), NOT from claim_batch, which installs no cross-claim roster and therefore charges producers the floor already warm-serves — the first ranking in the comments below was wrong for that reason and is corrected further down; (2) diagnostics_merge ×5,512 on one key in the memo-on trace was memo-hit verification cost on a large argument (Value::eq on serve), not real work — the memo-retired trace shows ~0 wasted for it.

Attribution and ranking comments: attribution, corrected ranking, floor comparison.

What

The seed's pure-call path consumes the materialization judgment before deriving any identity. An undeclared single pure demand (std.materialization_ladder rule 5: AcceptedSingleRecompute → Recompute) now constructs, looks up and retains no result-memo key. The eval-frame result memo (EvalCallMemo, GUNBC_EVAL_MEMO, eval_call_memo_frame_exit, its hit/miss/overflow receipt fields, value_fast_eq) and its provider row extdeps.realization.eval_memo are deleted. Net seed delta: −363/+287 lines across the change, −207/+64 in v1_interpreter.rs.

The two provider tiers that remain each already decide admission first and key second, and are untouched: the cross-claim tier (roster of v2.workflow.floor_pure_producer_share, admitted by resolved declaration identity, then content-hashed, then verified on serve) and the parse-table memo (admitted by the Memoize verdict carried on the table). The recompute-trace ledger (GUNBC_RECOMPUTE_TRACE=1, which the floor sets for itself) still keys every pure demand — it is the rule-4 measurement instrument and serves nothing — and now also discloses argument_identities_derived, its own keying work.

Chain re-derivation (DESIGN §6b) — why this is the root and not a symptom patch

• a — std.materialization_ladder: a single undeclared pure demand is AcceptedSingleRecompute (rule 5); a repeated demand whose LCA is a shared-state frame is AuthoredDuplication, prescribed Share, never Memoize (rule 1); only a plural demand at an isolation boundary or under declared emergence is Discharged into a covering provider.
• b — extdeps.realization.eval_memo: declared itself "the ladder's single-site discharge provider" serving "every repeated pure named-fn demand within one InterpContext". A provider scoped to ONE shared-state frame can never be the discharging provider of anything under a: every demand it could see was either single (Recompute) or a within-frame repeat (Share). Its admitted population was empty by construction. This is the earliest unjustified boundary.
• c — eval_pure_named_call built eval_recompute_key for every pure call because b told it to; Value::Str is hashed in full per call (composites have allocation-local hash reuse, strings never had one).
• d — extdeps.languages.json.parse threads the document through every *_at(s, i) call → O(n²) key building; gunbc.fleet_health selected GUNBC_EVAL_MEMO=0 to route around it.

The refused fixes (pointer-hash string cache, name allowlist, adaptive seed cache) all keep boundary b and cheapen the work done past it. Deleting b is the repair; a gate over an empty admitted set would be dead weight of the same kind, so the memo is deleted rather than gated.

Acceptance

1. A direct-recompute demand reaches neither key construction nor memo storage, with a RED control. src/v1/stage0/tests/pure_demand_recompute_admission.rs (its own binary because the trace latch is process-wide): two equal pure demands with a String argument on the production route derive 0 argument identities (a new per-frame odometer, argument_identities_derived, single writer at the three keying sites) and both evaluate in full (equal evaluator-step deltas — a served repeat takes no callee steps). Positive control: the same pair under the trace ledger derives exactly 2 and still evaluates both. Verified red by temporarily restoring an unconditional eval_recompute_key before admission: left: 2, right: 0.
2. Equal demands still reuse where admitted; different args stay different; collision controls kept. The cross-claim tier's enrolled controls are unchanged and pass: a_rostered_producer_fills_once_and_serves_later_claims, the_same_args_under_a_different_fn_identity_miss, cross_claim_memo_key_distinguishes_field_names_across_ordinal_collisions, every_field_of_a_served_record_resolves_under_a_reordered_interner, a_homonym_outside_the_roster_identity_does_not_store (cargo test --release -p v1-compiler --lib -- recompute cross_claim: 29 passed).
3. Output + refusal equality on valid / malformed / truncated / trailing / large. New instrument tools.json_parse_cost_probe (dag/gunbc/instruments/json_parse_cost_probe.dag) writes the serialized value or the typed gap text per subject; the .out files are byte-compared between the merge-base binary (memo on, default) and this branch's binary. See table below.
4. Input-size series. Same instrument: bytes, elapsed per subject, argument_identities_derived (0 on the production route by construction; the ledger's count under trace), peak RSS via getrusage. See table below.
5. Matched floor runs. See the CI comparison below (per-entry eval_steps / cpu deltas against the base run at the same population).

Size series and receipt equality (tools.json_parse_cost_probe, this session's arm64 container, both binaries built locally at -j2/-j4 release)

Three runs over the same eight subjects: the merge-base binary (56375ec, memo on = default), the merge-base binary with GUNBC_EVAL_MEMO=0, and this branch's binary. elapsed is the BEGIN→out mtime bracket per subject; out_sha is the receipt's sha256 (serialized value on a parse, typed gap text on a refusal); process peak RSS from getrusage (dominated by corpus resolution, ~6 GB on every run).

subject	bytes	base memo-on (s)	base memo-off (s)	this branch (s)	receipt sha (identical across all three)
m0250	67,215	3.44	2.28	2.53	f8b402654ef8449e
m0500	134,715	10.58	4.50	5.27	7fd206782c72cc91
m1000	269,719	36.00	8.92	9.62	a5a6ac664447dbe1
m2000	542,719	132.21	18.17	19.15	77168613d65e7374
m4000 (large)	1,088,719	502.01	36.65	38.81	651fe1b20ff509ed
malformed	22	0.00	0.00	0.00	a67cbf106140902a (REFUSED not readable as JSON at offset 17)
truncated	13,321	0.36	0.33	0.36	e7747c2ac3d16f7f (REFUSED not readable as JSON at offset 13321)
trailing	831	0.02	0.02	0.02	88d56bd3bac11191 (REFUSED … trailing content at offset 830)
whole process		wall 811 s / user 989 s / RSS 6969 MB	wall 194 s / user 188 s / RSS 6140 MB	wall 206 s / user 199 s / RSS 6141 MB

Memo-on is quadratic (×2 bytes → ×3.7–3.8 s); this branch is linear and coincides with memo-off within noise. Key-building work on the production route is 0 identities by construction (the RED control reads the odometer); the deleted work is the difference between the memo-on and memo-off columns. All 24 receipts are byte-identical across the three binaries.

Recurring failure mode, purpose test, seed growth

• gunbc.recurring_failure_mode.memo_key_rehashes_a_growing_persistent_value gains its second instance (the string case), the earliest-unjustified-boundary reading, the repair, and the rung after repair at the declared grain: mechanically preventable for undeclared demands (the state stays writable; the RED above refuses it, and that RED is off the merge path under rust_unit_tests_off_the_merge_path), unchanged for the roster-admitted cross-claim population, whose ceiling, trigger and program remedy (retire the seed) stand. The row's "no new seed cache, no evaluator patch" is honoured: this change deletes a cache.
• gunbc.v1_maintenance_standing purpose test: every interpreted v2 route — the self-host compile, the required floor, every instrument — paid this per-call identity derivation; the change is BehaviorPreservingRedundancyRemoval that leaves the seed smaller. Refused classes: no new language behavior, no compatibility obligation (GUNBC_EVAL_MEMO is retired, not kept as a no-op arm), no escape hatch, no seed capability; exported surface grows by exactly one reader.
• gunbc.pure_demand_admission_seed_growth enumerates the two added items (argument_identities_derived, note_argument_identity_derived) and is enrolled in seed_growth_justification_roster.
• gunbc.materialization_provider_targets drops the eval-memo row (witness count pin 7 → 6, renamed); tools.cause_assertion_census no longer prefixes its invocation with a switch that no longer exists.

Consumers

• gunbc.fleet_health (HEALTH-0: an out-of-band, read-only fleet health check #11681, untouched) sets GUNBC_EVAL_MEMO=0; after this lands that variable is unread and can be dropped there — recompute is the only realization.
• Recorded honestly in ci_floor_materialization_receipt_note: no consumer of take_process_eval_recompute_totals exists in the tree today, so target/floor-materialization-receipt.txt is not written by any current surface. That predates this change and is a separate finding.

🤖 Generated with Claude Code

[gunbai-bot]

Acceptance (5): matched floor runs — the consumer-wide qualification, and it is red

Base run 35458211949 (13d9308, same required-gate population) vs this PR's run 35466872753 (d0d90d3): 4,184 common claims, 4,161 with eval_steps on both sides.

	base (memo on)	this head (memo retired)
claims doing MORE eval steps	—	2,767 / 4,161 (median +14 %, p90 3.6×, p99 19×, max 116×)
claims unchanged	—	1,394
grandfathered claims over the 361,500-step ceiling	0	76
new-witness claims over the 72,300-step line	52	245
pass → budget_interrupted	—	23
fail	19	19 (same set)
adjudication	green	REFUSED, 134 blockers

Largest movers: test.claim.runner_capacity_plan_witness.* (116×, 69×, 37×), v2.test.claim.translate_underived_refusal.* (94×, 93×), test.claim.compiler_frontend_program_status_witness.* (70×, 55×, 37×). Terminal results are otherwise preserved (no pass↔fail flips); what moved is cost.

Reading: the corpus's witnesses carry a large population of within-frame repeated pure demands — std.materialization_ladder rule 1 AuthoredDuplication, prescribed Share — that the memo was discharging invisibly. The retirement is the ladder's own reading, but the qualification says the consumers are not rewired, so this head cannot land the floor green as-is. Escalated to the parent with three options (A: rewire program first; B: recompute by default + the memo kept only as a DECLARED roster provider filed as a §4b(3) drop against rule 1 with "rewire to Share" as its trigger; C: withdraw). Recommendation on the escalation is B. Nothing more lands on this PR until that ruling.

Reply to review 68632 (REQUEST_CHANGES)

Verified and agreed: gunbc.rung_drop.determinism_transitive_reachability told later lanes to protect the memo on the strength of a memo-on-with-keying vs memo-off-with-keying arm (the retired GUNBC_EVAL_MEMO=0 realization still derived an identity per call). Re-measured on its one surviving executing consumer, determinism_live_intra_holds, with claim_batch --claim-run on both binaries: 165,094 → 327,130 eval_steps (2.0×, not 100×; CPU 312 ms → 582 ms), under its grandfathered 361,500-step ceiling. The row is reconciled on the branch (next push): the two live-tense GUNBC_EVAL_MEMO sentences are dated and marked retired, the "protect it" constraint is withdrawn with the measurement that withdraws it, and the loop-invariant the memo was hiding is restated as part of the row's own trigger obligation (carry, don't re-demand). docs/design-rung-drops.md is regenerated from the row.

[gunbai-bot]

Two more qualification readings, both against landing this head as-is

tools.docs_projection_gate regen (the DESIGN/rung-drops projection, run in CI): merge-base binary 4m59s wall / 4m50s CPU; this branch's binary killed at 48 min elapsed, 47 min CPU, 11 GB RSS, still running — at least 10× and unbounded from where I stopped it. docs/design-rung-drops.md on this branch was therefore regenerated with the merge-base binary over the reconciled rows.

What the memo was serving (claim_batch --claim-run under GUNBC_RECOMPUTE_TRACE=1 on v2.test.claim.translate_underived_refusal, the module with the 93×/94× movers), wasted ms summed over the module's 14 claims, top of the ledger:

producer	wasted ms	repeat calls
infer	16,530	120
infer_entries_for_tree	16,428	120
rust_integer_facts_node (nullary)	7,253	2,998
translate_grounding_derived_gate	6,992	216
infer_gather_fold_init	6,954	864
rust_add_i32_facts_node (nullary)	6,936	2,946
inferred_facts_witness_for_node	6,388	198
infer_language_inhabitants_roster_roots (nullary)	6,067	3,441

Two shapes, both std.materialization_ladder rule 1 (AuthoredDuplication → Share): (i) infer re-demanded on the same tree from many sites (the whole inferred-facts fold recomputed ~120× per claim); (ii) nullary pure fns used as constants (rust_integer_facts_node() demanded ~3,000 times) — a fn standing where a data declaration belongs, which the seed's per-ctx data_cache would already evaluate once. The memo was the corpus's undeclared discharge for both. Same shape on determinism_live_intra_holds (node_callee_symbol, primitive_definition_of, decls_in_modules).

So the honest statement of this PR's standing: the root repair is right and its consumer-wide qualification says the corpus is not ready for it — option A on the escalation is a corpus program (nullary-fn→data and infer-sharing rewires), not a follow-up. Holding for the parent's ruling.

[gunbai-bot]

Reply to review 68692 (REQUEST_CHANGES), both findings verified and fixed in fb0e2df: the two receipts in memo_key_rehashes_a_growing_persistent_value that still presented GUNBC_EVAL_MEMO=0 / extdeps.realization.eval_memo as a selectable mitigation are dated (2026-09-18) and marked superseded by the repair receipt; the rung after repair now reads MECHANICALLY PREVENTABLE — the state is writable by one restored call, the control refuses it, and that control sits under rust_unit_tests_off_the_merge_path — with the structural ceiling and its M1 trigger stated beside it rather than claimed. PR body updated to match. The escalation on whether this head can land at all (floor red, 134 blockers) is still open with the parent.

[gunbai-bot]

Ruling (parent, relaying the side-chat): option A. This PR is HELD in its current shape and converted to draft. The root repair (admission before identity; the memo, its switch and its provider row deleted) and the RED control stay on the branch; the retirement is re-qualified only as a combined cutover on an exact head AFTER the shared producers are rewired as their own qualified changes (nullary producer fns → data; the repeated infer demand carried instead of re-demanded), each with its own before/after. Option B (a declared within-frame provider roster) is refused: the dependencies are ladder rule-1 AuthoredDuplication and a roster prices discharge, it does not reclassify a demand. Also binding: the docs-projection regen that never completed is independently disqualifying (a base-binary regen preserves the artifact, proves nothing about the replacement), a red floor is not delete-first licence (gap-intolerant boundary: prepare the replacement, switch when it preserves the contract), and the 23 pass→budget_interrupted flips break the floor's contract on their own. In progress on this lane: per-producer attribution of the regen blowup and a measured ranking of the shared producers to order the rewires.

[gunbai-bot]

Reply to review 68707 (REQUEST_CHANGES): verified — dag/gunbc/legacy_extdeps_scope_frontier.tsv still rostered the deleted dag/extdeps/realization/eval_memo.dag; the row is removed in 71af78b. Merge readiness is unchanged: the PR is a draft under the parent's option-A ruling and is not landable in its current shape regardless of review state (floor red by the change's own qualification; producer rewires first).

[gunbai-bot]

Attribution and ranking (option-A work, no landing)

1. The docs-regen blowup is attributed, and it is NOT the producers

Re-derived on the memo-retired binary (gunbc run --entry dag/gunbc/instruments/docs_projection_gate.dag):

route	memo-retired binary
expected_design_failure_modes_md via --claim-run (no eager data env)	completes, 254 s wall, resolve-dominated
whole failure_mode_document() serialization (354 rows, 2.56 MB) from a scratch entry	~15 s over resolve
same producer via plain gunbc run (eager data env)	timed out at 25 min
build_initial_env alone (any function, plain gunbc run) with per-declaration begin/end marks	3,137 data declarations finish (largest: int128_interval_spec 456 ms); data witness_floor_workflow (dag/gunbc/witness/witness_floor_workflow.dag) never finishes (>15 min, killed)

Chain: (a) plain gunbc run eagerly evaluates every data declaration in the entry's closure before calling the function (v1_interpreter::build_initial_env, eager_data_env = !claim_run) — the regen never demands the CI workflow, yet pays it; (b) witness_floor_workflow's body carries an authored recurrence the memo was serving: on the memo-on trace of the same eager env, diagnostics_merge is re-demanded 5,512× on one key, bash_fold_serialize_body 10× at ~800 ms, plus the grammar production-matching family (formal_production_rhs_matches_emitted_conj_step, formal_production_for_lhs_exact, …). Two rewires, each its own change: the eager env (undemanded work; eval_var already evaluates data lazily through data_cache) and the workflow fold's recurrence.

2. Shared producers ranked by measured contribution

Memo-on binary, GUNBC_RECOMPUTE_TRACE=1, claim_batch --claim-run over every module with an over-line or interrupted claim in run 35466872753 (115 modules, 627 claims traced). "projected inclusive ms" = Σ over ledger keys of first-evaluation time × (repeats), inclusive of callees, so nested producers double-count — it is an ORDER, not a sum.

rank	producer	shape	projected inclusive ms	repeats	claims	declaration
1	grammar_node_insert_atom_identities_child	args	8,081,003	264,219	44	src/v2/std/grammar.dag
2	grammar_node_insert_atom_identities	args	5,175,335	233,015	44	src/v2/std/grammar.dag
3	dag_canonical_symbols	nullary	496,176	125	25	src/v2/extdeps/languages/dag.dag
4	dag_canonical_symbol_map	nullary	495,884	125	25	src/v2/extdeps/languages/dag.dag
5	rustc_phase_order	?	490,901	480,193	74	dag/gunbc/self_host_compile_phase_frontier.dag
6	grammar_carried_symbol_map	?	490,473	125	25	src/v2/extdeps/languages/dag.dag
7	dag_language_model	?	482,581	105	14	src/v2/workflow/floor_cost_debt_edit.dag
8	grammar_expr_node	args	224,858	280,242	39	src/v2/std/grammar.dag
9	floor_route_gap_roster	?	166,518	1,060	45	dag/gunbc/witness/v2_native_route.dag
10	inferred_tree_digest	?	163,002	2,352	24	src/v2/test/claim/execution/native_selected_witness_bundle_test.dag
11	grammar_named_edge	args	162,417	414,909	41	src/v2/std/grammar.dag
12	parse_expr_with_first	args	153,278	8,744	29	src/v2/compiler/02_parse.dag
13	structural_property_witness_digest	args	132,443	15,737	40	src/v2/compiler/05_eval.dag
14	grammar_to_node	args	128,075	147	29	src/v2/std/grammar.dag
15	grammar_root_to_node	args	120,068	119	24	src/v2/std/grammar.dag
16	grammar_productions_to_node	args	119,694	118	23	src/v2/std/grammar.dag
17	grammar_production_to_list_step	args	118,776	4,647	21	src/v2/std/grammar.dag
18	descent_witness_digest	args	116,020	9,128	40	src/v2/compiler/05_eval.dag
19	grammar_production_to_node	args	115,533	4,635	20	src/v2/std/grammar.dag
20	grammar_expr_to_node	args	112,515	4,705	33	src/v2/std/grammar.dag
21	milestone_status	args	109,627	4,115	28	dag/gunbc/compiler_frontend_program_status.dag
22	floor_expected_red_roster	?	99,316	1,060	45	dag/gunbc/witness/v2_native_route.dag
23	selected_logic_primary_plan	nullary	97,335	331	23	src/v2/test/claim/execution/native_selected_witness_bundle_test.dag
24	floor_route_gap_expectations	nullary	89,178	962	45	src/v2/workflow/floor_route_gap.dag
25	gunbc_runner_cores_per_slot	nullary	89,138	1,756	60	dag/gunbc/runner/runner_slot_allocation.dag
26	selected_logic_identity	args	82,873	1,155	23	src/v2/test/claim/execution/native_selected_witness_bundle_test.dag
27	selected_logic_plan_for	args	79,786	367	23	src/v2/test/claim/execution/native_selected_witness_bundle_test.dag
28	inferred_tree_evaluated_facts_digest	args	79,165	1,202	37	src/v2/compiler/05_eval.dag
29	selected_logic_native_member	args	78,697	1,075	20	src/v2/test/claim/execution/native_selected_witness_bundle_test.dag
30	parse_expr_sequence	args	77,433	3,097	29	src/v2/compiler/02_parse.dag

Reading: ranks 1–7 are ONE root — the nullary constants dag_language_model() → dag_canonical_symbols() → dag_canonical_symbol_map() → grammar_carried_symbol_map(grammar: dag_grammar()), demanded ~125× across 25–44 claims and re-walking the whole grammar tree each time (grammar_node_insert_atom_identities*, 264k/233k repeats). Converting that family (and dag_grammar/dag_lex/dag_lex_rules beneath it) from nullary fn to data is stage 1: one change, measured before/after on those 44 claims. Stage 2: the other nullary constants demanded per call — rustc_phase_order (480k repeats / 74 claims; it takes a phase argument, so it is a table lookup re-deriving its table), the floor rosters floor_route_gap_roster / floor_expected_red_roster / floor_route_gap_expectations (~1k repeats / 45 claims), gunbc_runner_cores_per_slot / gunbc_runner_slot_desired (60 claims). Stage 3: the carried-result rewires — the 05_eval infer digests (structural_property_witness_digest, descent_witness_digest, inferred_node_facts_cache_digest, 12–16k repeats / 40 claims) and 02_parse parse_expr_with_first / parse_expr_sequence re-demands — which are the "carried infer result" class the ruling names.

[gunbai-bot]

Ranking correction — the floor-faithful census supersedes the claim_batch one above

The ranking in the previous comment was measured with claim_batch, which installs no cross-claim roster, so it charged producers the floor already serves from the warm tier (dag_canonical_symbol_map is a rostered warm row; its grammar_carried_symbol_map walk is paid once per floor run, not per claim). The floor's own census from the memo-retired run 35466872753 (required_floor_cross_claim_demand.tsv; within-claim repeats = evals − claims per producer key, summed over keys; inclusive of callees, so an ORDER, not a sum) is the ranking that the stages follow:

rank	producer	within-claim repeat ms (memo retired, inclusive)	same, base memo-on	repeats	declaration
1	integer_pow2_magnitude	2,050,922	1,844	22,715	src/v2/std/integer.dag
2	formal_production_for_lhs_exact	931,054	2,957	86,155	src/v2/std/grammar.dag
3	formal_production_for_lhs_exact_supported	927,613	0	86,155	src/v2/std/grammar.dag
4	formal_production_unique_lhs_exact_match	926,151	0	86,155	src/v2/std/grammar.dag
5	formal_production_rhs_matches_emitted_conj_step	895,624	1,570	170,205	src/v2/std/grammar.dag
6	formal_production_matches_lhs_exact_child	813,661	997	371,712	src/v2/std/grammar.dag
7	lhs_exact_production_match_fold_step	787,746	0	371,660	src/v2/std/grammar.dag
8	formal_production_matches_emitted_child	763,370	2	371,340	src/v2/std/grammar.dag
9	formal_production_rhs_matches_emitted_conj	733,769	0	97,173	src/v2/std/grammar.dag
10	formal_production_rhs_matches_emitted_conj_conj	731,003	0	97,138	src/v2/std/grammar.dag
11	decimal_digits_double	549,729	0	296,253	src/v2/std/integer.dag
12	bind_outcome	349,034	55,003	89,476	src/v2/std/diagnostic.dag
13	formal_productions_for_lhs	87,312	334	87,840	src/v2/std/grammar.dag
14	decimal_digit_of_units	79,593	3	579,483	src/v2/std/integer.dag
15	rust_integer_facts_node	71,013	229	11,309	src/v2/extdeps/languages/rust.dag
16	formal_productions_for_lhs_step	66,411	0	4,662,659	src/v2/std/grammar.dag
17	rust_integer_range_node	61,923	0	11,309	src/v2/extdeps/languages/rust.dag
18	decimal_digit_double	59,289	0	362,484	src/v2/std/integer.dag
19	decimal_nonzero_double	57,995	0	22,683	src/v2/std/integer.dag
20	rust_integer_inhabitant_node	49,831	221	1,926	src/v2/extdeps/languages/rust.dag
21	rust_declared_inhabitants_root	47,588	644	143	src/v2/extdeps/languages/rust.dag
22	rust_integer_carrier_interval_spec	45,718	0	11,267	src/v2/extdeps/languages/rust.dag
23	decimal_digit_value	42,867	0	4,427,718	src/v2/std/integer.dag
24	inferred_facts_witness_for_node	39,518	66	9,644	src/v2/compiler/04_infer.dag
25	rust_add_declared_inhabitants_root	36,878	0	125	src/v2/extdeps/languages/rust.dag
26	rust_coercion_candidate_edges	36,875	0	125	src/v2/extdeps/languages/rust.dag
27	translate_grounding_derived_gate	32,531	104	7,447	src/v2/compiler/06_translate.dag
28	rust_target_model_core_edges_with_operators	31,719	0	108	src/v2/extdeps/languages/rust.dag
29	rust_target_model_core_edges_full	30,852	0	108	src/v2/extdeps/languages/rust.dag
30	rust_add_inferred_facts_for	30,029	8	8,091	src/v2/test/claim/manual/rust_add_emit_translate_test.dag

total within-claim repeat ms (inclusive, do not sum): pr=12787193 base=89549

Stage 1 (rank 1, the integer decode re-deriving every standard type's 2^n ladder) is #11784. Stage 2 is the formal_production_* grammar lookup family (ranks 2–10, 13, 16), then bind_outcome chains and the rust.dag nullary node constants (ranks 15, 17, 20–22).

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-20T14:57:08.438862Z	e85c469	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e85c4690db

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Increment the identity counter only after successful derivation

When recompute tracing encounters an unkeyable argument, such as a closure, this increments the counter before eval_recompute_arg_key returns None; eval_pure_named_call then invokes eval_recompute_partial_key, which traverses and increments the same positions again. Consequently the new public accessor and [recompute-trace] field report failed attempts—and double-count a prefix—as argument_identities_derived, making the diagnostic evidence inaccurate for the floor's known unkeyable-call population. Increment only after receiving Some, or report attempts under a distinct name.

Useful? React with 👍 / 👎.