Isolate fixture census on a scratch MultiEntryIndex by gunbai-bot[bot] · Pull Request #13575 · gunb-ai/gunbc

gunbai-bot · 2026-10-08T05:07:07Z

Summary

Fixture census loaders no longer retain probe both-closure rows on the process-shared MultiEntryIndex (scratch dropped with the loader). That is a real leak class; it is not the floor blow-up.
B1 flattened ancestry_str_bindings stays. An empty-ancestry parent walk was tried and reverted: gunbc.commit_workflow closure reconcile went ~20× (main 15951783a2 62s vs walk head 1251s) with RSS in the same band.
Floor class (measured): Strict-prep of the gate already sits near CI memory.high (25 GiB). Planning the seven forged-probe census claims is a one-module keep-set increment, not a compiler-closure pull. RFM: required_floor_strict_prep_sits_on_memory_high. Sibling RFM fixture_compile_retained_on_the_process_shared_index stays for the probe-index leak.

Keep-set (prepare-only on main `15951783a2`)

Gate seeds: keep 2981, prepare VmHWM 19.45 GiB.
Gate seeds plus test.claim.megarac_managed_host_forged_probe_witness: keep 2982, VmHWM 18.71 GiB.
Added identity: that witness alone (40 visible names); zero v1.compiler / v2.compiler modules.

Margin (named CI pair)

Green merge_group floor run 37744410173 (does not plan those seven claims): prepare-closure-resolve 17.3→21.7 GiB, high=0; prepared-subject-warm current=26840956928 (~25.00 GiB).
Control DRAFT: floor-control for #13575 (plan forged-probe census) #13576 run 37743678634 (plans them): prepare-closure-resolve 17.5→21.9 GiB; prepared-subject-warm current=26837438464 (~25.00 GiB); wall ~90 min, CANCEL.

Test plan

GREEN/RED fixture-index isolation.
Witness-module both-closure excludes string-literal MegaRAC; bounded closure size.
DRAFT: floor-control for #13575 (plan forged-probe census) #13576 remains the floor control until a cut of Strict-prep-at-the-watermark (not a cgroup raise) lets that floor finish.

…e floor's process-shared index. A compile-census probe that imports a wide stack (the MegaRAC forged witness) was writing parse, admission, and edge-map rows onto the MultiEntryIndex the claim fold consults. Extend that walk on a scratch shell that shares source bytes and is dropped with the loader; keep the enrolled control that reds if the shared slot grows. Co-authored-by: Cursor <cursoragent@cursor.com>

The green control is silent if the old loader also happens not to grow the warmed caches; this sibling must fail if the scratch route is reverted. Co-authored-by: Cursor <cursoragent@cursor.com>

A missing process-shared slot is the subject of the RFM evidence, not an excuse to return passed without asserting. Co-authored-by: Cursor <cursoragent@cursor.com>

…ral MegaRAC. c2 showed planning that census witness (not claim count) pins before probe() runs. This control refuses the hypothesis that Strict-prep of the changed witness module both-closes production named only inside forged_probe_source. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T05:51:08Z

review 77883: the skip-green is gone on current head. Both fixture_closure_extension_does_not_populate_the_process_shared_index and fixture_closure_extension_via_shared_index_grows_both_closure_edges now panic if try_process_shared_index fails (5b615fd950). try_process_shared_index builds the slot on first call in a unit-test process (remote GREEN/RED both ran assertions).

The duplicate import loop in extend_fixture_imports_on_process_shared_index stays as the enrolled pre-fix route: it is the RED that grows both_closure_edges on the shared slot. Folding it into a production parameter is the RFM next-rung trigger (shared index not in the function's reachable parameter set), not this cut.

— sent from zesty-wren-256

The floor control lagged at 719f8ea, which still returned passed when try_process_shared_index failed (review 77893).

…ex loaders. GREEN and RED now differ only by which MultiEntryIndex the one closure authority extends, so the red cannot drift onto a second BFS. Co-authored-by: Cursor <cursoragent@cursor.com>

The floor refused MultiEntryIndexBuiltTwiceForOneNameSet: scratch over the shared source_files map was recorded as two extra builds of the same 8044 names. Empty caches over an already-indexed set are not a second index. Co-authored-by: Cursor <cursoragent@cursor.com>

#13576's floor hit the same BuiltTwice refusal as #13575; without this the planner-poke run cannot print a cgroup series.

A fresh scratch per compile made every compile_dag_diagnostic_census cold-extend (~20s fills) and cancelled the floor at 90 minutes. The fixture slot still does not write the process-shared index the fold reads. Co-authored-by: Cursor <cursoragent@cursor.com>

A thread-local fixture index kept every census both-closure for the process (review 77949). Scratch is per compile again, copied from the shared slot's already-computed edge and parse rows so a census does not cold-walk the gate, then dropped without writing MegaRAC rows back onto the fold's index. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T09:26:13Z

Addressed review 77949 (#13575 (comment) — artifact /api/reviews/77949/artifacts/stdout.log): the thread-local FIXTURE_EXTENSION_INDEX is gone.

Fixture extend now always builds an owned scratch (new_multi_entry_index_scratch_over), seeds it by copying the shared slot's already-computed both_closure_edges, parsed-reference, admission, and parse-cache maps (new Rc around a cloned edge index so later Rc::make_mut cannot mutate the fold's slot), then drops that scratch when the loader returns. Writes such as the MegaRAC both-closure stay on the scratch and die with it; they are not the union-of-every-fixture-on-the-thread retention the review named.

The GREEN control already panics if the shared slot is missing, so it cannot skip-green the isolation claim. The RED still asserts that routing the same extend onto the process-shared slot grows both_closure_edges. A second-extend reuse test on the old thread-local slot was deleted because that slot no longer exists.

Head d3fe0ec2ca. Floor / plan-time pin on stacked #13576 still required before asking to land.

…s seed. A green merge_group floor hits the 25 GiB watermark at prepared-subject-warm with tens of high events. Planning the forged-probe census adds ~25 more, not gigabytes. Enrol that class, and refuse a both-closure that would make this 13 KiB module a second corpus. Co-authored-by: Cursor <cursoragent@cursor.com>

Main's warm seam holds ~2962 typed gate modules; the dominant carrier is te.ancestry_str_bindings (~10.5M nearly-unshared entries). That overlay is typecheck-only once the subject is closed. Lookup now walks parents when the overlay is empty, so the prepared repository can drop it. RFM required_floor_strict_prep_sits_on_memory_high stays unfixed until the floor peak is below memory.high. Co-authored-by: Cursor <cursoragent@cursor.com>

Dropping the flatten after Strict-prep still paid the prep peak. The redundant work is the unshared copy itself. Interface envs keep the parent Rc chain; a non-empty overlay is complete (no walk); an empty overlay walks last-wins. build_type_env checks chain vs overlay on real modules, then stores empty. Co-authored-by: Cursor <cursoragent@cursor.com>

The per-module flatten is the Strict-prep pin: planning the forged-probe census witness still runs prepare_repository_from_corpus over the gate typed graph, which retained ~10.5M unshared overlay entries. Equivalence is a last-wins flatten test plus a first-wins mutant. Cache.str_bindings is locals only. Fixes the E0382 source move in the witness-module test. Co-authored-by: Cursor <cursoragent@cursor.com>

lookup_binding_on_chain_seen and collect_chain_bare_names_seen now key the visited set by module_path in 04_env.dag. The stage0 rust matches that fold instead of a HashSet of Rc pointers the dag cannot name. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T12:24:06Z

review 77966

Verified against 4201e483cc.

Visited set (blocking). Confirmed: 04_env.dag lookup_binding_on_chain was a plain parent fold while the committed rust used HashSet<usize> of Rc pointers. That is a proven cost-shape defect on diamond imports (§6) and a dag/rust fork (§3). Fixed in this head: lookup_binding_on_chain_seen / collect_chain_bare_names_seen are authored in 04_env.dag, keyed on module_path (the identity the dag can name; not a host pointer). The stage0 rust is that same fold. Full seed re-emit of v1_compiler_infer_env.rs is not a one-function regen; the two sources now agree.

Hand-written tests / loaders (advisory). The emit_host.rs tests are the RED/positive control the floor brief required (last-wins flatten on real dag/std modules; first-wins mutant must go red). They are not a production scaffold. scratch_index_for_fixture_closure_extension is the fixture-index isolation from earlier on this PR, PURPOSE-admitted under v1_seed_standing. No deleted-scaffold census row: nothing here is scheduled to be thrown away.

— sent from zesty-wren-256

…rd scratches as a typed kind. The Err(_) fallback dropped try_process_shared_index's error and rebuilt a name-set index over every shared root. Scratch construction now records ScratchCachesOverExistingSet so the sharing control can count it without treating it as a second NameSetIndex. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T12:42:06Z

review 77971 (REQUEST_CHANGES) verified against 4201e48 and landed on 73ad9eb.

Shared-index Err(_) widen: confirmed. scratch_index_for_fixture_closure_extension dropped try_process_shared_index and walked canonical_shared_index_roots. That arm is gone; the error is now ?. A missing slot refuses instead of paying a corpus walk.
Scratch as escape hatch: confirmed. new_multi_entry_index_scratch_over skipped record_multi_entry_index_site so MultiEntryIndexBuiltTwiceForOneNameSet would not fire. Both constructors now record. NameSetIndex is still the twice-refuse kind; ScratchCachesOverExistingSet is countable and is not a second name-set index. Sharing-control unit tests plus fixture_scratch_shell_is_not_a_second_index_of_the_name_set passed.

The dashboard payload truncated a third finding ("Every fi…"). Nothing further in the current code was identified from that fragment.

This does not discharge the floor brief (main Strict-prep pin / #13576). RFM remains unfixed.

…runs or a control PR. RFM receipts named job IDs, watermark bytes, and throwaway #13576 as the next-rung trigger. The trigger is now that Strict-prep (including a planned forged-probe census) finishes with high at 0 well below memory.high; measurements point at required_floor_runner. Co-authored-by: Cursor <cursoragent@cursor.com>

…ted state through sibling walks. Cache merge no longer flattened str_bindings, so the novelty gate's conflict channel was always empty. Forks are recorded from each import's TypeEnv chain without retaining an overlay. collect/lookup now return updated seen so a diamond is walked once. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T12:57:02Z

review 78018 (REQUEST_CHANGES) verified on 59990b2 and landed on 89c7234.

Binding-fork channel emptied: confirmed. merge_type_env_cache_guarded no longer called guarded_union_str_bindings, so binding_forks was always empty and the novelty gate had no ledger. Restored as ledger_peer_import_binding_forks over each peer import's TypeEnv chain (overlay-wins, same-authority skip). Cache str_bindings stays empty so Strict-prep does not retain a flatten. No rung-drop: the refusal/ledger distinction is preserved.
Sibling walks cloned one seen map: confirmed. collect_chain_bare_names_seen and lookup_binding_on_chain_seen now return the updated seen and thread it through the parent fold, so a diamond is visited once per module.

Ancestry last-wins / first-wins mutant tests passed.

…bjects. The forged-probe both-closure control now discriminates only on MegaRAC paths from the string literal. The ancestry last-wins control requires the named types and integer modules to exist instead of skipping them. Co-authored-by: Cursor <cursoragent@cursor.com>

Last-wins on the full parent chain let a leaked homonym beat a name an earlier import declared. Lookup now takes last-wins of parent locals first, then walks ancestry only on a miss. Binding forks stay ledgered from the TypeEnv chain, not a flattened cache overlay. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T13:15:02Z

review 78025 (REQUEST_CHANGES) verified on 6f8b4b9 and landed on the following head.

merge_type_env_cache_guarded still does not flatten str_bindings (that flatten is the Strict-prep pin). Binding forks are not dropped: union_parent_type_env_caches already calls ledger_peer_import_binding_forks over each peer import's TypeEnv chain (89c7234). A comment now says so at the cache merge. No rung-drop; the novelty-gate ledger is preserved.
Direct-selected vs transitive leak: confirmed. Last-wins on the whole parent chain let a later import's leaked homonym beat an earlier import's own export (v2 memory-control audit + witness-realization plan + P0 field wall (568→45 drained) #6663 x Ground fnv1a64 hash fork: unify v2.std.node.Hash with std ContentHash #6686). Lookup now last-wins parent locals first, then walks transitive ancestry only on a miss. Discriminating control: earlier_direct_export_beats_later_transitive_leak (lookup takes a; full flatten still takes the leak c). Ancestry overlay stays empty in production.

…ookup walks. Collecting names then looking up each one again was names × chain per import. The ledger now memoizes a lookup-equivalent index per module_path. A retained ancestry overlay is the Strict-prep pin; misses re-read the parent Rc chain the graph already holds. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T13:25:57Z

review 78032 (REQUEST_CHANGES) verified on 36c84ed.

Quadratic fork ledger: confirmed. collect_chain_bare_names then lookup_binding_on_chain per name was names × chain per import. ledger_peer_import_binding_forks now builds a lookup-equivalent index in one pass, memoized by module_path across the union.
Lookup miss walks the parent DAG: confirmed, and kept. The shared-state LCA of bare-name resolutions is the module TypeEnv; materializing ancestry_str_bindings there is the Strict-prep pin (unshared overlay per typed module on prepared.graph). A per-closure index that outlives typecheck is that overlay again. Walk cost is O(direct parents) on a one-hop hit and O(unique modules on the chain) on a miss, paid per lookup, not retained in the prepared subject. That derivation is in the lookup comment and RFM required_floor_strict_prep_sits_on_memory_high.

The dashboard payload truncated a further finding.

… import_path. Fork detection already lives on the parent TypeEnv chain; the cache merge only unions deps and locals, so it no longer takes a path or conflict list. TypeEnvCache.str_bindings stays: build_type_env still fills it with the module's own locals. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T13:41:12Z

review 78039

Deleted ancestry_chain_mismatch_name in 04_env.dag and the rust mirror. It had no call site, fold, or test.

Cut import_path and the pass-through conflicts list off merge_type_env_cache_guarded. It now takes two caches and returns a cache. union_parent_type_env_caches still ledgers forks with ledger_peer_import_binding_forks and only then folds the cache union. ParentCacheRow is gone with the unused path.

Did not delete TypeEnvCache.str_bindings. The merge writes empty_map() because a unioned parent cache must not flatten ancestry, but the per-module cache still stores locals: build_type_env sets cache_str_bindings = final_env.str_bindings, and type_env_cache_from_bindings fills the same field. Removing the field is a TypeEnvCache schema replacement, not this leaf cut.

On the advisory seed-census note: PURPOSE admits these v1 edits for the floor walk (v1_seed_standing); consumers are union_parent_type_env_caches, lookup_binding_on_chain, and the emit_host ancestry / scratch controls. The RFM rows are .dag, not more seed Rust. A full census of remaining seed surface is a later v1-maintenance change, not this PR's cutover.

— sent from zesty-wren-256

…ing lookup seen across sibling parents. A diamond skipped under a later parent made last-wins disagree with the fork ledger. Fixture scratch now reads through the shared index instead of cloning its maps; parse_cache stays local because it is intern-paired. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T14:06:06Z

review 78042

Did not split the resolution change into another PR. The empty-ancestry walk is the Strict-prep pin cut (not retain a flatten on every typed module). Title is now "Walk empty TypeEnv ancestry instead of flattening; isolate fixture index rows". PURPOSE on lookup_binding_on_chain now states the semantics change: homonyms the overlay took from a later leak resolve to the earlier direct export (#6663 x #6686). RFM still says UNFIXED; this commit does not claim the floor class is discharged.

Order dependence: sibling parent walks no longer share seen. A diamond re-entered under a later parent last-wins in that parent's order, matching collect_chain_lookup_index_seen. Control: diamond_reentered_under_later_parent_last_wins_with_that_parent.

Fixture scratch no longer clone_froms shared parsed_references / bare_reference_admission / both_closure_edges. It reads through scratch_underlay and writes locally. parse_cache is not underlaid: those rows are intern-paired with the slot that parsed them, so sharing them would mix intern tables. Scratch reparses.

Advisory seed-census: same as review 78039 — PURPOSE + the emit_host controls. Not a census rewrite on this PR.

— sent from zesty-wren-256

…export over a later leak. Required-ci on the floor control refused when that preference split host String and FreeMonoid<Char> String in one if. The pin remains an empty overlay, not a homonym-precedence change. Co-authored-by: Cursor <cursoragent@cursor.com>

…ts that still described a flatten. Sibling parent walks share seen and hits so a diamond is visited once per name, matching the ledger's last-wins without retaining an overlay on TypeEnv. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T14:31:02Z

review 78055

Shared per-name seen plus hits (binding memo keyed by module_path) across sibling parent walks. A diamond is visited once per name, last-wins still overlays a later parent's full chain using the memoized result — same order as collect_chain_lookup_index_seen, without retaining an overlay on TypeEnv.

Updated the interface_env_for_import / interface_env_surface comments that still said flattened ancestry maps.

Advisory seed-census: PURPOSE on the walk plus GREEN/RED fixture-index controls; not a census rewrite here.

— sent from zesty-wren-256

gunbai-bot · 2026-10-08T14:32:02Z

review 78055 (re-verified on current head a218af3f58, not c413df4a)

The exponential path-walk is gone: lookup_binding_on_chain_seen threads st.seen and st.hits through the parents fold. A later sibling that reaches the same module_path returns the memoized hit (or miss) instead of walking that DAG again. Hits are dropped with the lookup; they are not a retained TypeEnv overlay.

The interface_env_for_import / interface_env_surface comments no longer claim flattened ancestry maps.

RFM still says the memory class is UNFIXED. That is honest; this memo does not claim prepared-subject-warm high is 0.

— sent from zesty-wren-256

…ports beat leaks. Lookup last-wins the parent walk (flatten-equivalent); leaving the old named rule beside that path was a meaning fork. Kernel skip stays overlay_skips_kernel_name. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T14:51:14Z

review 78065

Deleted overlay_direct_import_exports (uncalled) and direct_import_export_precedence_note (it still said direct-selected beats leak). Executing path is empty ancestry_str_bindings and lookup_binding_on_chain last-wins of the parent walk, matching flatten — later leak can win. build_ancestry_precedence comments now say that. Kernel skip remains overlay_skips_kernel_name. Namespace-cut roster probe now names lookup_binding_on_chain. Diamond-test comment updated: sibling walks share seen/hits.

— sent from zesty-wren-256

…onto interface caches. Forks live on ledger_peer_import_binding_forks. TypeEnvCache.str_bindings stays module locals; union/interface caches leave it empty. The skip-equal cache merge is named for what it does. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T15:05:39Z

review 78074

Deleted guarded_union_str_bindings / _into_acc / _into_overlay and GuardedStrBindingsUnion. Nothing called them; forks are ledger_peer_import_binding_forks. Kept union_base_choice_note for the remaining skip-equal unions (deps_map, cycle_set_str, variant_locals).

Did not delete TypeEnvCache.str_bindings: it still holds the module's locals on a per-module cache (build_type_env, typed-module type_env_cache). Union and interface caches now leave it empty (merge_type_env_cache_skip_equal, interface_cache_from_module) so the field is not a flattened overlay. Renamed merge_type_env_cache_guarded to merge_type_env_cache_skip_equal.

Advisory seed-census: PURPOSE on the walk plus GREEN/RED fixture-index controls; not a census rewrite here.

— sent from zesty-wren-256

… it. Per-lookup parent walks kept runner-local high at 0 but cancelled the #13576 floor at 90 minutes in nominal-subject-seeds. The overlay is demand for that typecheck only; retained TypeEnvs stay empty so prepared.graph does not pin it. Co-authored-by: Cursor <cursoragent@cursor.com>

…ibed run numbers. Co-authored-by: Cursor <cursoragent@cursor.com>

Dropping it per module forced every later module to rebuild the parent index; the floor control sat 80 minutes in compile.reconcile with high=0. Strip on the prepared graph, the process typed-module cache, and at compile_to_resolved exit. Production walks ignore leftover parent overlays. Co-authored-by: Cursor <cursoragent@cursor.com>

…merging a copy per importer. Typecheck looks up names in locals then that shared prefix map, so each probe is O(1) and importers of the same parent reuse the same storage. Co-authored-by: Cursor <cursoragent@cursor.com>

… not a 180s guess. The ignored control prints wall and peak RSS from compile_to_resolved; a 180s cutoff failed a 20-minute closure that is still the instrument, not a unit-lane gate. Co-authored-by: Cursor <cursoragent@cursor.com>

The RFM already named the next-rung trigger; it did not say the production scratch is final and the shared-index extender is only the enrolled red. Co-authored-by: Cursor <cursoragent@cursor.com>

…t_workflow reconcile instrument. This is a span count for the 20× vs main, not another index shape. Co-authored-by: Cursor <cursoragent@cursor.com>

…isolation. A 20× reconcile regression cannot land while the floor still Strict-prepares the gate graph. Probe-index writes stay on a throwaway MultiEntryIndex. Co-authored-by: Cursor <cursoragent@cursor.com>

…ipts. The empty-ancestry walk does not ship; citing it as the lookup realization overstated the class. Co-authored-by: Cursor <cursoragent@cursor.com>

Pre-push fmt refused the previous RFM commit over this leftover whitespace. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-10-08T23:36:03Z

review 78159 (REQUEST_CHANGES on c78a48b437) verified against 9b93cc159d.

LOOKUP REALIZATION / title vs reverted walk: confirmed. The receipt named a scratch overlay and seam=nominal-subject-seeds after the empty-ancestry walk after B1 flatten was restored. Rewritten: flatten (ancestry_str_bindings on prepared.graph) is what ships; the walk is recorded as tried and reverted (~20× commit_workflow reconcile). PR title is now the scratch-index isolation this diff actually delivers.
MEASURED MARGIN / STRUCTURAL PREDICTION as unattributed settled facts: confirmed. MEASURED MARGIN now names merge_group run 37744410173 and DRAFT: floor-control for #13575 (plan forged-probe census) #13576 run 37743678634. STRUCTURAL PREDICTION now cites the prepare-only pair on main 15951783a2 (keep 2981 vs 2982, one added module).

RFM still says UNFIXED. That is honest; this does not claim prepared-subject-warm high is 0.

— sent from zesty-wren-256

gunbai-bot · 2026-10-10T02:02:03Z

Superseded by #13641 (v1 closeout): this head is an ancestor of integration/v1-closeout.

Brian Searls and others added 2 commits October 8, 2026 05:06

Enrol the pre-fix shared-index extension as the discriminating red.

719f8ea

The green control is silent if the old loader also happens not to grow the warmed caches; this sibling must fail if the scratch route is reverted. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot Bot mentioned this pull request Oct 8, 2026

DRAFT: floor-control for #13575 (plan forged-probe census) #13576

Closed

1 task

Brian Searls and others added 2 commits October 8, 2026 05:35

Refuse a skip-green when the shared-index controls cannot bind.

5b615fd

A missing process-shared slot is the subject of the RFM evidence, not an excuse to return passed without asserting. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot Bot pushed a commit that referenced this pull request Oct 8, 2026

Merge #13575 so this throwaway carries the skip-green panic.

624e5b3

The floor control lagged at 719f8ea, which still returned passed when try_process_shared_index failed (review 77893).

Brian Searls and others added 2 commits October 8, 2026 05:54

Share the fixture import-seed walk between the scratch and shared-ind…

dfa5629

…ex loaders. GREEN and RED now differ only by which MultiEntryIndex the one closure authority extends, so the red cannot drift onto a second BFS. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot Bot pushed a commit that referenced this pull request Oct 8, 2026

Merge the MultiEntryIndex scratch-census fix into the floor control.

1e63e2c

#13576's floor hit the same BuiltTwice refusal as #13575; without this the planner-poke run cannot print a cgroup series.

Brian Searls and others added 2 commits October 8, 2026 09:12

Brian Searls and others added 4 commits October 8, 2026 09:37

gunbai-bot Bot pushed a commit that referenced this pull request Oct 8, 2026

Merge #13575 overlay-free typecheck into the throwaway floor control.

322ddbc

Brian Searls and others added 2 commits October 8, 2026 12:45

Brian Searls and others added 2 commits October 8, 2026 13:00

gunbai-bot Bot changed the title ~~Drop fixture-closure both-closure rows off the floor's shared index~~ Walk empty TypeEnv ancestry instead of flattening; isolate fixture index rows Oct 8, 2026

Brian Searls and others added 2 commits October 8, 2026 14:15

Brian Searls and others added 9 commits October 8, 2026 17:16

Cite floor-cgroup producers in the Strict-prep RFM instead of transcr…

d471d62

…ibed run numbers. Co-authored-by: Cursor <cursoragent@cursor.com>

Record that the fixture scratch loader is the terminal retention shape.

34211b3

The RFM already named the next-rung trigger; it did not say the production scratch is final and the shared-index extender is only the enrolled red. Co-authored-by: Cursor <cursoragent@cursor.com>

Count lookup overlay vs parent-walk vs fork-ledger names on the commi…

f87ad6a

…t_workflow reconcile instrument. This is a span count for the 20× vs main, not another index shape. Co-authored-by: Cursor <cursoragent@cursor.com>

Restore B1 flattened ancestry lookup; keep the fixture scratch-index …

c78a48b

…isolation. A 20× reconcile regression cannot land while the floor still Strict-prepares the gate graph. Probe-index writes stay on a throwaway MultiEntryIndex. Co-authored-by: Cursor <cursoragent@cursor.com>

Align the Strict-prep RFM with B1 flatten and named prepare-only rece…

a28288c

…ipts. The empty-ancestry walk does not ship; citing it as the lookup realization overstated the class. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot Bot changed the title ~~Walk empty TypeEnv ancestry instead of flattening; isolate fixture index rows~~ Isolate fixture census on a scratch MultiEntryIndex Oct 8, 2026

Drop a rustfmt-blank after the capture-gap fixture test.

9b93cc1

Pre-push fmt refused the previous RFM commit over this leftover whitespace. Co-authored-by: Cursor <cursoragent@cursor.com>

This was referenced Oct 9, 2026

DRAFT control: #13211 floor on #13575 (DO NOT MERGE) #13605

Closed

RFM: fixture-closure-union-emit suspected superlinear cost #13608

Closed

gunbai-bot Bot closed this Oct 10, 2026

gunbai-bot Bot mentioned this pull request Oct 10, 2026

v1 closeout mega branch #13641

Merged

Conversation

gunbai-bot Bot commented Oct 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Keep-set (prepare-only on main 15951783a2)

Margin (named CI pair)

Test plan

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

review 77966

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 8, 2026

Uh oh!

gunbai-bot Bot commented Oct 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

0 participants

gunbai-bot Bot commented Oct 8, 2026 •

edited

Loading

Keep-set (prepare-only on main `15951783a2`)