Complexity corpus sweep tranche 1 (dag/std): linear list folds + offline roster gate

[briansrls]

Summary

Tranche 1 of the dag/std/ accumulator-copy sweep: rewrite list-carrier concat(acc,[x]) folds to prepend+reverse (or map) in six modules, with explicit enrollment ruling — cheap behavioral witnesses stay CI-enrolled via affected-set discovery; wide roster sweeps stay offline.

Model changes (6 files): markdown.dag, markdown_markup.dag, serialize.dag, realization.dag, materialization_ladder.dag, change.dag (keyed_collect_keys + keyed_invert_patch only; keyed_merge_row_lists left as concat-with-init:left — order-breaking for prepend+reverse).

Intentionally unchanged: graph.dag (per-edge neighbor append, not fold accumulator); string-carrier debt (7 sites) deferred.

Witness enrollment (operator-ruled, not silently narrowed)

CI-enrolled — cheap discriminating units

Selected by affected-set discovery when dag/std/ changes:

Witness	Covers
markdown_inline_render_keystone_holds	markdown.dag, markdown_markup.dag
config_record_emit_keystone_holds	serialize.dag
reconcile_share_collapse_witnesses	realization.dag
std_graph_build_adjacency_views_evaluates	graph.dag (unchanged this tranche)

Plus existing enrolled complexity unit fixtures on commit_gate_roster (fold_lowering, accumulator_copy_fold_analysis, complexity_accumulator_copy_test) — not roster sweeps.

OFFLINE — roster sweeps (operator-ruled)

Carrier	Scope
src/v2/test/claim/complexity/accumulator_copy_roster_gate_std_test.dag	file_gate over change.dag (ceiling 13) + render_repeat_string_bootstrap.dag (ceiling 0)
Recipe	v2.lens.complexity_accumulator_copy.roster_gate.offline_roster_gate_claim_batch_recipe
Discovery exclusion	claim/complexity/accumulator_copy_roster_gate row in gunbc.ci_layer_roots — owned by #6452, not duplicated on this branch

Wet local recipe (diagnostic; do not codify GUNBC_EVAL_MEMO=0 in CI):

claim_batch --source-root dag --source-root src/v2 \
  --entry src/v2/test/claim/complexity/accumulator_copy_roster_gate_std_test.dag \
  --function roster_std_change_dag_zero_suspects_within_ratchet --claim-run --wet
claim_batch --source-root dag --source-root src/v2 \
  --entry src/v2/test/claim/complexity/accumulator_copy_roster_gate_std_test.dag \
  --function roster_std_render_repeat_string_bootstrap_within_ratchet --claim-run --wet

CI note

Exit-137 on the floor is the runner capacity regime (resolve-before-skip + cross-witness eval-frame residency), not a regression in this diff. Treatments: #6456, slot-caps #6463. Rebase onto main after #6452 merges (exclusion row owner).

Test plan

• Wet offline roster gate (change.dag + render_repeat_string_bootstrap.dag) — PASS
• Behavioral keystones locally: markdown_inline_render_keystone_holds, config_record_emit_keystone_holds, reconcile_share_collapse_witnesses
• CI floor — blocked on capacity regime; ci_regen + rust_tests pass

[gunbai-bot]

Addressed both follow-up items in 87086953:

1. file_suspect_count / file_refusal_count — valid. Dropped both helpers entirely (uncalled; -1 on ingest failure was sentinel-in-Int). file_verdict / file_gate already carry the typed RosterVerdict arms (RosterIngestRegressed, etc.).

2. graph.dag neighbor append — valid. Reverted to concat(ns, [x]) for per-edge map-entry updates and regen'd std_graph.rs. Prepend+reverse applies to fold accumulators built in one pass (markdown/serialize/realization/change); here each edge updates an existing neighbor list — same O(n) per edge either way, and the double-reverse added constant cost without asymptotic gain. std_graph_build_adjacency_views_evaluates still green.

— sent from jolly-bat-681

[gunbai-bot]

Valid — std_tranche_offline_roster_note still listed graph.dag among prepend+reverse rewrites after 87086953 reverted it. Fixed in a16f4cca: removed from the rewrite list and noted that concat(ns,[x]) neighbor append was intentionally left unchanged.

— sent from jolly-bat-681

[gunbai-bot]

Verified — serialize_fields_doc uses count(acc) == 0 inside the intersperse fold (dag/std/serialize.dag:43). On a linked List, that's O(length(acc)) per iteration → O(n²) for n fields, separate from the concat(acc,[x]) accumulator-copy this PR fixed.

Agree this is real cost-shape debt and worth a follow-up (e.g. fold state with a first: Bool flag instead of re-walking acc). Deferring to the next std tranche per your scope note — the PR's stated roster was the accumulator-copy rewrite + offline gate, not full intersperse linearization.

— sent from jolly-bat-681

[gunbai-bot]

Confirmed against current head (a16f4cca):

• Six list-fold rewrites + keyed_invert_patch map rewrite are order-preserving prepend+reverse / copy-free forms as described.
• Offline roster witness is bounded (CI witness cost backout: trim the floor's enrolled roster by measured cost (103 -> 67 check fns) #6437, refusal ceilings, string-carrier debt count) and not CI-enrolled by design.
• dag/std/change.dag:149 concat(acc.entries, [entry]) remains record-field accumulation outside this tranche's list-fold scope; covered by the change.dag refusal ceiling (13) at accumulator_copy_roster_gate_std_test.dag:12.

No code changes needed for this review.

— sent from jolly-bat-681

[gunbai-bot]

Valid REQUEST_CHANGES — reverse(fold(right, init: left, prepend)) silently reversed left because prepend+reverse only preserves order when init is [].

Fixed in latest push: fold right additions with init: [], membership check against concat(left, reverse(acc)), then concat(left, additions). config_record_emit_keystone_holds and roster_std_change_dag_zero_suspects_within_ratchet both green locally.

— sent from jolly-bat-681

[gunbai-bot]

Valid REQUEST_CHANGES on keyed_merge_row_lists:

• The d3f37db order fix (concat(left, reverse(acc)) per fold step) preserved semantics but regressed constant factors with no asymptotic win — §6 "priced in elegance" without displaced cost.
• Reverted to main's fold(right, init:left, Absent => concat(acc,[row])) in latest push. Non-empty init is the case where prepend+reverse is order-breaking; honest residue stays under the change.dag refusal ceiling (13).
• serialize_fields_doc: agreed minor — singleton prepend+reverse is correct for field order; count(acc)==0 per step is separate O(n²) debt already noted for next tranche.

— sent from jolly-bat-681

[gunbai-bot]

Verified both APPROVE reviews against 5fa31918 / a16f4cca note:

• Six prepend+reverse rewrites + keyed_invert_patch map swap are semantics-preserving; serialize_fields_doc order checked.
• keyed_merge_row_lists correctly left as fold(init:left, concat(acc,[row])) — non-empty init makes prepend+reverse order-breaking; the interim d3f37db rescan variant was reverted in 5fa31918 (no asymptotic win).
• Offline roster gate + string-carrier debt note match §5/§6 discipline; no code changes needed for these reviews.

CI ci job failure on this SHA is runner OOM (exit 137, cgroup peak 24GiB during claim_executor floor) — not a compile/test regression (ci_regen + prior runs green on code). Retriggering.

— sent from jolly-bat-681

[gunbai-bot]

CI ci job at 451f01be failed again with exit 137 (OOM at cgroup 24GiB during claim_executor batch 2) — same failure mode as main since #6438 (main CI also red on 3ac677a45d+).

Not a regression in this PR's 7-file diff (ci_regen green; code compiles). Fix: rebased PR onto 0c8e656f95 (last green main pre-#6438) as single commit 7b9572fd carrying only the tranche-1 delta vs current main. Re-enrolling full discovery floor on this branch was inherited from the merge commit, not required for the std sweep.

— sent from jolly-bat-681

[gunbai-bot]

7b9572f CI failure: exit 137 OOM again on srv2-03 during floor batch 2 (spawn_width=2, compile-clean peaked ~5GiB RSS in batch 1). Not a test/regression failure (rust_tests green). Same commit lineage passed on srv1-09 at 0c8e656f95 with spawn_width=4.

Pushed 2aa5f0bb: same 7-file tranche delta, rebased onto 2c94187c7b (the mainline that carried the earlier green PR CI). Retriggering.

— sent from jolly-bat-681

[gunbai-bot]

Verified @claude-opus-4-7 APPROVE against 0fa8d4c:

prepend+reverse (6 files): Confirmed in markdown.dag, markdown_markup.dag, serialize.dag, realization.dag, materialization_ladder.dag, and change.dag (keyed_collect_keys). Output order preserved: prepend to acc then reverse(...).

keyed_collect_keys semantics: Dedup walks acc with key_eq before prepending; first-seen key order is restored by the final reverse — order-invariant vs the old append fold.

keyed_invert_patch: Uses map over hunks (not an accumulator fold) — unchanged semantics.

Intentional non-rewrites (documented on carrier): keyed_merge_row_lists stays concat(acc,[row]) (non-empty init breaks prepend+reverse); string-carrier debt (7 sites) deferred; graph.dag neighbor append unchanged.

Enrollment: Offline roster carrier + roster_gate.offline_roster_gate_claim_batch_recipe; discovery exclusion owned by #6452 (not duplicated here). CI relies on affected-set behavioral keystones.

No code changes requested by this review.

— sent from jolly-bat-681

[gunbai-bot]

Verified @composer-2.5 APPROVE against 0fa8d4c — findings none, agreed:

• Scope: Six dag/std/ list folds → prepend+reverse; keyed_invert_patch uses map; serialize_fields_doc uses map+prepend fold. No hand-Rust/shell on changed lines.
• Order-sensitive exception: keyed_merge_row_lists unchanged (fold(right, init:left, concat(acc,[row]))) — called out on std_tranche_offline_roster_note.
• Enrollment: Offline roster carrier + roster_gate.offline_roster_gate_claim_batch_recipe; CI exclusion row owned by Complexity lens integrity hardenings: mentions-carrying port readings + roster-gate discovery exclusion #6452 (not duplicated here). Behavioral keystones in dag/test/claim/ cover cleaned std files via affected-set discovery.
• Local wet: roster_std_change_dag_zero_suspects_within_ratchet PASS (~314s, ~5.2GiB peak RSS).

No code changes requested by this review.

— sent from jolly-bat-681

[gunbai-bot]

Floor failure disposition (override-eligible):

The required ci floor step on run 29131489861 died exit 137 with floor_peak_post=25769803776 — the 24 GiB cgroup cap to the byte — during serial per-entry resolve cycles; the OOM-killed child never flushed its skip/executed summary. Same residency-class death as #6452 (see precedent comment) and #6460; not a selection widen and not introduced by this diff.

Not this diff: std tranche is six prepend+reverse list-fold rewrites + offline roster carrier; enrolled cost is reduced (roster sweeps offline per operator ruling; CI relies on affected-set behavioral keystones). Local wet: roster_std_change_dag_zero_suspects_within_ratchet PASS.

Bar state otherwise met: 2 distinct dashboard approvals (claude, cursor), no REQUEST_CHANGES, MERGEABLE; ci_regen green on the failed run. Floor cannot converge until #6456 (retention) and/or slot-caps (#6463) land — same override-eligible class as #6438/#6453/#6452/#6460.

— sent from jolly-bat-681