Plan: space lens — confident memory prediction from the static .dag (Node B walking skeleton) by gunbai-bot[bot] · Pull Request #6442 · gunb-ai/gunbc

gunbai-bot · 2026-07-10T19:18:24Z

Plan doc only — awaiting operator sign-off, deliberately left draft. No code, no .dag edits.

What this is

A minimal-project plan for the space lens: a fold over the static .dag graph that returns predicted peak resident memory, fed to realization to pick shard width. The time analog (complexity lens) already exists; this is the space sibling.

Key finding — this is a resume, not a greenfield

The architecture is already operator-signed across compute-envelope-model.md, input-envelope-roadmap.md, and resource-aware-scheduler.md (the ruling "derivation is the authority; measurement is the falsifier"). The carriers exist: CostAccount.space / CostBasis::Predicted, the floor(budget ÷ per_shard_peak) width formula, and the landed InputEnvelope fail-closed admission. The gap is exactly one hollow spot: CostAccount.space is always byte_size(0), and width reads a static hardcoded row instead of a fold. This session's merged measurement PR (#6425) built the falsifier half; this plans the derivation half.

Decisions I need from you

Calibrate now vs. gate on cross-shard sharing (S2b). Each shard currently duplicates the std/spec prefix in RAM (!Send barrier), so the per-node constant re-calibrates when sharing lands. Calibrate against today's reality (useful now, re-measure later) or wait?
Homing. This lane is multi-owner and touches load-bearing substrate (std.realization*, scheduler). It should be a work item under the envelope owners, not run out of a closed measurement session. Where do you want it?

Flip to ready (gh pr ready 6442) only if you want automated review of the prose; the intended gate is your sign-off on the plan.

A temporary red probe (union_resolve_receipts_test.rs) used to characterize the width==1 typecheck-compute-counter semantics was captured by the auto-committer in 533e473 while it sat in the working tree. It asserts a deliberate RED, so it fails rust_tests. It was never intended to land. Restores the file to its state at 674381f. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…echeck counter `typecheck_compute_count()` counts typecheck cache MISSES on the current thread and is never reset in production (only tests reset it), so it equals a closure size only from a cold start. The measurement window cannot assume that: - width == 1 returns `run_discovery_rows` on the discovery thread, which has already resolved every changed file via `floor_diff_edits_from_line_ranges` plus both prefix entries. The reported count was therefore |closure(changed files) u closure(prefix) u closure(rows)| -- a function of the diff under test, not of the roster. - `floor_skip_discovery_witness` calls run_discovery_corpus three times on one thread, so the 2nd and 3rd calls read a counter warmed by the 1st. Count the union of authored module names across the graphs the shard actually resolves (prefix contexts + each roster entry) instead. That is a property of the source closure: independent of cache warmth, resolve order, and the diff. Prefix modules stay counted -- they are resident for the shard's lifetime, so they belong in the memory pairing. Note the parallel path was never the problem: run_walk dispatches DiscoveryBatch via thread::spawn, so shards do start cold. The defect was the pre-row work on the discovery thread, which shards never repeat. Adds a discriminating control: a prior same-thread resolve moves the compute counter (asserted, so the control stays discriminating) but must not move the closure count. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

gunbai-bot · 2026-07-10T19:27:23Z

The CI failure here is not a defect in this PR — it's a pre-existing gap in the CI substrate that any docs-only PR trips. Root cause, confirmed from the job log (run 29117568824, job 86444691071):

This branch's tree is exactly origin/main + one markdown file under docs/plans/. Nothing under dag/ or src/v2/ changed.
ci.yml's interim documentation-only shortcut fired correctly: claim_executor: documentation-only diff — floor skipped (CI_FLOOR_DISPOSITION=documentation_only_skipped). The floor is intentionally not run for a docs-only diff.
Because the floor didn't run, write_resolve_receipt was never called, so target/floor-resolve-receipt.txt doesn't exist.
The "Floor resolve receipt gate" step then fails closed: "floor resolve receipt missing - fail closed."

The gap: the merge-admission gate does honor the docs-only disposition (it stamps Skipped, not Success), but the resolve-receipt gate does not — it requires the receipt unconditionally. The two sibling steps disagree about a docs-only skip.

The correct fix is in the CI substrate, not this doc: the receipt gate should short-circuit (skip) when CI_FLOOR_DISPOSITION=documentation_only_skipped, mirroring the merge-admission gate. That step is generated from dag/gunbc/ci_materialization.dag (ci_floor_resolve_receipt_gate_script) and ci.yml is drift-gated, so the fix is a .dag edit + regen — a separate, small PR against load-bearing CI substrate, which I've flagged to the operator rather than bolt onto a plan-doc branch.

This PR is a plan doc awaiting sign-off and is intentionally draft; it does not need to merge before that fix lands.

— sent from merry-owl-649

… knobs) The host OOM is systematic, not intermittent: a width==1 shard is killed by host-level oversubscription under its own slot cap, so narrowing width cannot help. The per-shard prediction therefore feeds TWO knobs — within-job width and cross-job packing/admission — and the systematic case is served by admission- refusal, not width. Records RSS-at-kill as a lower-bound calibration point. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

gunbai-bot · 2026-07-10T20:25:17Z

Closing as parked, not abandoned — the content is fully preserved and this reopens in one click.

This is a complete plan doc (docs/plans/space-lens-minimal-project.md) awaiting operator sign-off on two things: (a) calibrate-now vs gate on the interface/body split, and (b) homing under the compute-envelope owners as the fused measurement+predictor work item that loyal-wren-398's affected-set lane is ready to bind receipts to. It is not implementation waiting to merge, so "mark ready" doesn't apply — and it can't cleanly go non-draft anyway until the separate docs-only-skip resolve-receipt gate bug is fixed (that's why ci is red here; root-caused in the comment above, it's a CI-substrate gap, not a defect in this doc).

Rather than sit as a nagging red draft, I'm parking it: the doc lives on session/merry-owl-649 and in history, nothing is lost, and reopening (or re-PRing from a fresh branch under the fused work item) resumes it immediately. The design conversation and the loyal-wren coordination that shaped it are captured; the first calibration datum (16.3 MiB/node censored-at-cap, run 29119564221) is already flowing to it.

Reopen when the homing decision is made.

— sent from merry-owl-649

…ew catch) The two cli_run.rs comments cited docs/plans/space-lens-minimal-project.md, which does not exist on main — the predictor design is in flight on PR #6442 (merry-owl-649's lane) and was never landed under that path. Rebound: the shared closure-definition authority is stated as this function itself, with the in-flight design cited by PR number and the landed parent-lane authorities cited by real paths (compute-envelope-model.md fleet envelope; input-envelope-roadmap.md admission). No behavior change; cargo check clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…hrunk by affected set) + falsifier host-OOM receipt (#6438) * WIP: Re-land affected-set CI enrollment flip once shard resolve footprint is * WIP: Re-land affected-set CI enrollment flip once shard resolve footprint is * Fix CI OOM: pin discovery corpus spawn_width_cap to 1. Run 28999086030 OOM-killed at width=2 on the 24GiB live slot during the discovery-flip corpus batch. Gate workloads still fit at width=2; only the tree-wide discovery batch serializes via spawn_width_cap=1, with a receipt witness and dissolve-on note. Co-authored-by: Cursor <cursoragent@cursor.com> * Fix type error in corpus discovery spawn width cap helper. Both if-branches must return Nat (hardware_thread_count_value), not a bare Int literal. Co-authored-by: Cursor <cursoragent@cursor.com> * Merge affected-set enrollment flip (takeover of PR #6403 from session/gentle-stag-677-flip) Conflict resolutions onto post-#6422/#6431/#6432/#6435 main: - ci_witness_optin_inversion: main's typed Scaffold Disposition -> Terminal (dissolve fired at the flip; roster stays as explicit-entry home, exclusivity dissolved) - floor step timeout: 45 (main) vs 60 (flip) -> 60 with provenance note - ci_spec notes: kept flip notes, stale fixed-8 phrasing dated - falsifier width note: + RESIDUE paragraph (5x exit-137 on srv2-class slots = converge lane, not plan width) Verified before commit: discovery batches charge corpus_resolve_nanos (own key), resolve_nanos=0 -> resolves_total declared 3 is NOT moved by the flip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * WIP: affected set processing * WIP: affected set processing * WIP: affected set processing * Correct the false flip receipt + kill-surviving calibration receipts (space-lens loop) CORRECTED RECEIPT: run 29000557166's floor never completed - the executor was host-OOM-killed mid-discovery-corpus at ~8min; the log's ExitSuccess belongs to the merge-admission stamp tool stamping CI_FLOOR_EXIT=137. No flipped corpus has completed in CI (0/1 ci + 0/5 falsifier). Both carrier notes corrected. Calibration (coordinated with merry-owl-649's space-lens lane): - roster_import_closure_nodes_pre_resolve: shared closure-count authority (pure import walk, closure grain not entry grain), emitted BEFORE the heavy resolve so killed runs still yield the (nodes, peak) lower-bound pair - width-1 definition-drift oracle: pre-resolve walk must equal post-resolve resolved union on completed runs; refuses on divergence. Proven by execution: pre == post == 190/213 nodes across Off/Applied modes - kill-surviving cgroup memory.peak pre/post steps (post is always()) in the ci job and falsifier.yml; scope semantics labeled on the emission lines (reset=ok floor-scoped; span compares post>pre; never silently conflated) - job backstop timeouts extended by the two aux steps in both jobs Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Calibration pairs carry censored-vs-exact labels (methodology: killed runs are censored observations) Floor steps get id=floor; the always() post-peak step emits floor_outcome so each (closure_nodes, peak) pair is explicitly labeled: success = exact point, anything else = censored lower bound (true demand strictly greater than read). Prevents the fit from treating cap-kill reads as point estimates, which would drag the slope down and make the predictor underestimate - the dangerous direction (merry-owl methodology catch, 2026-07-10). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Review fixes: explicit witness import + complete step-budget ledger - ci_floor_plan_witness_test.dag: import witness_ci_corpus_discovery_serializes_at_width_one explicitly (cursor catch on #6438). The call resolved pre-fix via the seed resolver's flat namespace, so the witness ran green by execution; the explicit import restores the file's per-symbol import convention. - ci_workflow.dag: the resolve-receipt gate step had NO step cap — a hang there could only die by job-cancel (the #6323 starvation-kill class). It now carries the aux cap (script is a sub-second receipt read) and the ci job backstop counts four aux terms (peak pre/post, resolve-receipt gate, merge-admission gate): every step budgeted, backstop = step-sum + prelude (claude review catch on #6438, sharpened). - falsifier_workflow_witness_test.dag: falsifier_backstop_is_step_sum_plus_prelude asserted the pre-calibration formula — latent red proven by execution (FAIL receipt), fixed to the live two-aux sum, re-run PASS. - ci.yml regenerated byte-stable from the carrier (backstops 125->130, gate step timeout 5m). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Rebind calibration provenance comments to real artifacts (cursor review catch) The two cli_run.rs comments cited docs/plans/space-lens-minimal-project.md, which does not exist on main — the predictor design is in flight on PR #6442 (merry-owl-649's lane) and was never landed under that path. Rebound: the shared closure-definition authority is stated as this function itself, with the in-flight design cited by PR number and the landed parent-lane authorities cited by real paths (compute-envelope-model.md fleet envelope; input-envelope-roadmap.md admission). No behavior change; cargo check clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Remove stray empty file (shell-redirect artifact the auto-committer flushed) An internal-messaging command's backtick content was command-substituted by bash; a '-> fail-closed' fragment became a stdout redirect and created an empty file at the repo root, which the auto-committer then committed as 38f0a46. No tree content beyond the empty file; removing it restores the branch to e99c757's content. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Scaffold-mark the cgroup peak calibration shell (cursor review catch) The three concat-built calibration runners (ci_cgroup_peak_locate_shell, ci_floor_peak_pre_script, ci_floor_peak_post_script) landed without the on-carrier scaffold markers repo convention requires for hand-shell. Each now carries a Disposition = Scaffold { dissolves_to: RealizationDispatch } row binding the decl (the ci_materialization pattern), and both scripts embed the shared dissolve-on note as a shell comment (the ci_spec pattern): dissolution = bash-emit (#5828 / gap-B emit(intent, Bash)) realizing the observation as an emitted ShellProgram intent or a typed host Observe effect. ci.yml + falsifier.yml regenerated; falsifier_workflow_witness_holds and the flip witnesses re-run PASS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Merge origin/main into session/loyal-wren-398 (resolve #6441 step-list conflict) Conflict resolution, all consciously declared: - ci_job steps: union — the calibration peak pre/post steps wrap the floor step (this branch) and #6441's materialization receipt gate slots between the resolve-receipt gate and the merge-admission gate (main). - The incoming materialization receipt gate step landed with timeout none — the same uncapped-step starvation-kill class this branch's review fix eliminated — so it now carries the aux cap, and the ci backstop counts FIVE aux terms (peak pre, peak post, resolve-receipt gate, materialization receipt gate, merge-admission gate); the budget disposition note records the merge provenance. - ci_run_step_natures_are_claims_counted_not_silent: RunStep count pin bumped 17 -> 19, acknowledging the two peak calibration steps #6441's count predates (conscious-count discipline; proven red at 17 then green at 19 by execution). - ci.yml regenerated from the merged carriers (backstops 130 -> 135). Verified on the merged tree: release bins rebuilt on merged Rust; falsifier workflow witness, flip witnesses, floor-plan/optin/width witnesses, and all 8 ci_materialization witnesses PASS; generated-artifact regen ExitSuccess. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

… gated) (#6455) * Inferred materialization increment 1: floor demand ledger + receipt gate (#6441) * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * WIP: Duplicate Computation * Analysis spine: spine_receipt composes dependency_view + materialize (increment 1) The Rc-safe half of run ≜ realize ∘ materialize ∘ dependency_view: compose the two catamorphisms that need no interpreter thread-safety. - src/v2/std/spine.dag: level-profile fold over the DependencyView DAG → critical_path_depth (longest dependency chain / the reduce spine) + independence_width (max nodes at one level / what parallelizes to hw width). spine_receipt emits the §9 numbers by COMPUTATION, not assertion, joining materialize's content-hash Share/dedup counts. - Discriminating RED (§9): single chain A→B→C → depth 3 / width 1 (serial); root over N independent leaves → depth 2 / width N. - Witnesses 9/9 (spine) green by execution; materialize 6/6 unchanged. The RUNNER (wall-clock parallel execution, §8 steps 3-4) stays increment 2+, gated on interpreter Rc→Arc (v1 interp is !Send). This increment measures the achievable width/critical-path floor the runner targets; it does not fabricate wall-clock N×. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Unbundle recompute-trace interpreter extension from the analysis-spine PR CI root cause: the recompute-trace extension edited v1_interpreter.rs, which hosts the atom_identity_hash/hash_combine fnv1a64 intrinsics that dag_collect_fingerprint_witness executes. That correctly invalidated the re-verify 'assumed-green node-frontier' SKIP of that witness (cone-touch, fail-closed) — but the witness is a pre-existing 806s (13min) fnv1a64 whole-corpus walk, so un-skipping it blew the 30-min floor budget. Main stays green only because its interpreter cone is untouched (witness stays skipped). The interpreter extension is a separable, env-gated (GUNBC_RECOMPUTE_TRACE=1), READ-mode diagnostic — not a dependency of the spine/materialize analysis substrate (pure .dag, content_hash-based). Reverting v1_interpreter.rs to main restores byte-identical interpreter content → the fingerprint witness returns to assumed-green SKIP → floor budget restored. The recompute-trace extension lands as its own follow-up PR where the fingerprint-witness re-run is expected. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * WIP: Duplicate Computation * Materialization ladder: the state x decision law, as executable witnesses The operator's 4-rule business logic generalized on one axis: the decision is a function of WHEN the redundancy is knowable and WHETHER what was knowable was prepared for. Errors fire only on knowable-but-unprepared; genuine emergence and declared triviality are typed acceptances, never silence. std.materialization_ladder (dag/std): - Frame/FrameKind: SharedStateFrame | IsolatedChildrenFrame | ReplayedFrame{attempts} | UnboundedSiblingsFrame — the nested-scope axis; a 'run' at any layer is a frame, never a different kind of thing. - DemandNature: Pure | IdempotentEffect | WorldRead{envelope_declared} | FreshEffect — the D3 effect gate; TTL = unmodeled dependency XOR declared staleness, never an eviction knob. - CacheProvider: keying (ContentKeyed|ExistenceKeyed) + EvictionPolicy (ScopeExit|SpacePacked) REQUIRED by construction — rule 3 unwritable-bad. - LadderVerdict (total): AuthoredDuplication | Discharged | RefusedNoProvider | RefusedScopeTooNarrow | RefusedExistenceKeyed (the build-if-absent #6352 wall) | RefusedUnmodeledWorldRead | ExemptFreshEffect | AcceptedBelowCostFloor | AcceptedSingleRecompute. - Declared-emergent frames obligate UP FRONT: retry frames derive checkpointing, unbounded-sibling frames (server loops, CI-runs-over-time) derive persistent caches — prepare-before-demand. 12/12 witnesses green by execution; each test is one cell of the table, fixtures mirror the live worked example (workflow/jobs/steps, retry, fleet). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * WIP: Duplicate Computation * Docs crystallization: derived materialization placement into roadmap ④ + the state x decision table - roadmap_authority ④ (ROADMAP.md regenerated via main_wet, drift-consistent): the nested-frame law, plurality cells, declared-emergent prepare-up-front, keying/staleness walls, the caching-completion un-shelve, the ①-vs-§2 cross-run-caching contradiction subsumed into per-node derived verdicts, and the v1.compiler.ownership §3-convergence row. - duplicate-work design doc: the state x decision table as the qualifier's final verdict logic + the ownership consolidation subsection. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Consolidation plan: realization + materialization + ownership are one law Operator synthesis (2026-07-09): Materialization = the verdict vocabulary; ownership = the verdict computation at the eval frame; realization = provider selection discharging the verdict at each frame's carrier. Share's handler is layer-aware — the cheapest cache is a REFERENCE (same program + ref-capable carrier + readonly => point-at-it, no copy, no destroyed memory), demoting through HAMT structural share / process memo / artifact / CAS, and demotion must be priced, never silent (#6249 clone-fallback is the receipt). Census of every hand-rolled instance found in-tree (16 rows): each named with its frame, ladder cell, action, and dissolution trigger — ownership (consolidated), Rc/HAMT (reference tier), clone-fallback (demotion wall), ParseTable + cached_stage (C5 grounding), M1 memo / intern / assumed-green / sccache / cargo-cache (C2 provider rows), PROCESS_RESOLVE_STORE (C3 = the 9GB scope fix), build-if-absent (landed as the keying wall), resolve-cache #5789 (rule-4 requirements), recompute-trace (state-4 finding source). Peers kept distinct: affected-set, Independence/Placement, mutable state. Sequence C1-C5 with C1 = this PR (ladder + live CI gate). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * WIP: Duplicate Computation * Ownership refactored onto materialization: the value tier, tested as providers First consolidation increment (operator-directed): v1.compiler.ownership's decision IS provider selection at the value grain, now expressed in the ladder and proven by mirror witnesses. - ProviderTier: Reference | Copy | Memo{keying} | Artifact{keying} | Cas{keying} — keying folded INTO store tiers, so an existence-keyed reference is unwritable BY SHAPE (§5 construction); frame obligations at isolation LCAs are dischargeable only by store tiers (a reference cannot cross an isolation boundary — witnessed). - Value tier: AccessMode (Read|Consume|Carry|Project), ValueUsage, value_materialization -> ValueDead | ValueMoved | ValueSharedByReference | ValueDemotedToCopy (PRICED, requires a declared CopyTier provider — the #6249 silent clone-fallback made refusable) | ValueRefusedAffine | ValueRefusedNoCarrierProvider. - Faithful to v1's THREE distinct plurality readings, kept separate: take_count = semantic_consumer_count (Consumed only, affine axis); value_access_plurality = binding_fan_out (Carry excluded, reference axis); borrow_count = whole_value_borrow_count (Read+Carry not Project, movability). - Fold-accumulator case mirrored at conservative pre-FoldAccUnwrap semantics: carries excluded from plurality but blocking the move -> demoted-to-copy. 23/23 witnesses green (14 frame cells + 9 value cells). Corpus-level equivalence against live v1 ownership folds = the C4 receipt. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * WIP: Duplicate Computation * WIP: Duplicate Computation/Materialization * G: exhaustive cache/share census (52 mechanisms, 6 groups) + C0 catalog-grounding of provider rows + forward-wiring walls The sweep found the tree already models caches: extdeps/cache cache_catalog (CacheInterfaceCatalogFacts, 9 rows) + extdeps/realization carrier facts + the warm==cold purity oracle. C0 grounds CacheProvider on that catalog (provider_from_catalog projection; CacheProvider.id -> CacheInterfaceId) so the ladder and the catalog never restate each other's half. Three forward walls (demands-from-DependencyView, hand-cache shape lens, live-provider-or-red) make new caches born as provider rows only. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * C0: provider rows derive from the cited cache catalog (extdeps/cache/materialization.dag projection) Operator ruling 2026-07-09: extdeps stays the ground-truth/interface layer; every USE of a cache goes through materialization so memo never gets re-invented. provider_from_catalog is the only door from a catalog row into the ladder: keying/tier/eviction derived from cited facts (mechanism->class: InProcess=>ScopeExit; Ttl/Lru/SizeBounded=>SpacePacked; Never/Manual outside a process => typed ProjectionRefused, counted by the new enrolled witness; HandAuthored/prefix-fallback keys => ExistenceKeyed until key-completeness upgrades them). CI sccache row now derived, hand-typed tier facts deleted, 6/6 + 23/23 witnesses green. 10b re-scoped: v1-internal caches = inventory only (retire with seed); ParseTable/cached_stage promoted to next. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * Fail-closed hardening: RefusedNatureConflict arm, weakest-claim nature default, UsesStep silence counted, ladder cells enrolled in CI The complexity-lens false-clean lesson applied to this gate before anyone relies on it: (1) heterogeneous natures on one identity were first-wins — now a typed RefusedNatureConflict (24th ladder cell, RED witness); (2) an unrostered run-step claimed PureComputation — now IdempotentEffect, the weakest claim (verdict-identical, no purity overclaim; a misdeclared WorldRead can no longer be legally memoized by default); (3) UsesStep foreign actions produced silence — now a pinned denominator (==9; a new foreign action must consciously bump it); (4) the ladder's 24 cells enroll on the CI surface so the law's own witnesses gate every run. 24/24 + 7/7. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * M1 ratchet bump: declared resolve count 1 -> 3 (gate's first live catch — enrollment pays counted resolves) Run 29058798771: enrolling the ladder witness file moved resolves_total 1 -> 3 (each enrolled entry file pays one closure resolve against the shared index) and the receipt gate redded that exact run — the designed semantics, receipted in the note. Bump acknowledges the two witness-entry resolves as declared debt; M2 (one closure resolve per roots set) ratchets back toward 1. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Retire the M1/M2 labels from carriers (no-codename rule; collided with DESIGN's floor-memoization M1/M2) Step renamed to 'Floor resolve receipt gate (declared cold-resolve count)'; notes and the gate echo now say what the things are: the counted cold-resolve receipt, and the shared-resolve rewire (one closure resolve per source-roots set). Naming-retirement note left on the carrier. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * Run-step natures pinned as counted claims, not affordances Operator ruling 2026-07-10: hardcoded shell scripts get no affordances — run-steps are supposed to be modeled as typed intents and emitted (shell-emission-model slice 4 covers ci_workflow RunSteps). Until that slice lands, every nature in job_run_demands is a claim (rostered FreshEffect / weakest-default IdempotentEffect), never derived from a modeled effect row. Count the whole claimed-nature surface (16 by execution) and pin it in the enrolled witness: a new raw-script step must consciously bump the pin, and slice-4 migration becomes a countable ratchet ending at zero, where the pin becomes a wall. Rejected alternative recorded in the carrier note: a per-step declared-natures roster (a second parallel ledger over scripts the shell-emission plan already governs). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Regen DESIGN.md after main merge (duplicate-work row + #6373 body-lowering row coexist) Conflict was both sides editing the open-threads list in design_document.dag: ours added the duplicate-work thread, main's #6373 rewrote body-lowering to Stages 1-3 LANDED. Resolution keeps both rows in the .dag authority; DESIGN.md regenerated from it, never hand-edited. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * Inferred materialization increment 1: floor demand ledger + receipt-or-red gate Running IS enrolling: the interpreter ledgers every keyed pure call and every InterpContext absorbs its totals into a process accumulator on Drop — by construction, no eval path escapes the receipt. claim_executor writes target/floor-materialization-receipt.txt at walk end; trace defaults ON in the executor, and an explicit =0 zeroes keyed_calls which the gate refuses. Gate arms this push (all verified under dash from the emitted ci.yml): receipt missing/malformed/keyed_calls=0 fail closed. Exact pins for unkeyed_calls and duplicated_keys land from the FIRST CI receipt (the resolve gate's measure-then-pin path) — unkeyed is known nonzero on the floor (count_matching takes a predicate closure; closures are the one disclosed identity-less class, dissolve-on captured-env content identity). Unit witness green: resolve -> eval -> ctx Drop -> absorb -> drain-once. Step addition bumped the claimed-natures pin 16 -> 17 consciously. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Review fix: drop the racy drain-once assertion from the receipt unit test opus-4-7 finding on #6441: under plain cargo test (still the documented runner) tests share a process, the env latch is OnceLock-sticky, and sibling ctx drops could absorb between the two takes — making the drain-empty assertion racy. Absorb-on-Drop stays asserted (monotone under concurrent absorbs: siblings only ADD); drain-once is Option::take by construction, not asserted through the shared global. Comment states the sharing semantics explicitly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization * WIP: Duplicate Computation/Materialization --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt (#6438) * WIP: Re-land affected-set CI enrollment flip once shard resolve footprint is * WIP: Re-land affected-set CI enrollment flip once shard resolve footprint is * Fix CI OOM: pin discovery corpus spawn_width_cap to 1. Run 28999086030 OOM-killed at width=2 on the 24GiB live slot during the discovery-flip corpus batch. Gate workloads still fit at width=2; only the tree-wide discovery batch serializes via spawn_width_cap=1, with a receipt witness and dissolve-on note. Co-authored-by: Cursor <cursoragent@cursor.com> * Fix type error in corpus discovery spawn width cap helper. Both if-branches must return Nat (hardware_thread_count_value), not a bare Int literal. Co-authored-by: Cursor <cursoragent@cursor.com> * Merge affected-set enrollment flip (takeover of PR #6403 from session/gentle-stag-677-flip) Conflict resolutions onto post-#6422/#6431/#6432/#6435 main: - ci_witness_optin_inversion: main's typed Scaffold Disposition -> Terminal (dissolve fired at the flip; roster stays as explicit-entry home, exclusivity dissolved) - floor step timeout: 45 (main) vs 60 (flip) -> 60 with provenance note - ci_spec notes: kept flip notes, stale fixed-8 phrasing dated - falsifier width note: + RESIDUE paragraph (5x exit-137 on srv2-class slots = converge lane, not plan width) Verified before commit: discovery batches charge corpus_resolve_nanos (own key), resolve_nanos=0 -> resolves_total declared 3 is NOT moved by the flip. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * WIP: affected set processing * WIP: affected set processing * WIP: affected set processing * Correct the false flip receipt + kill-surviving calibration receipts (space-lens loop) CORRECTED RECEIPT: run 29000557166's floor never completed - the executor was host-OOM-killed mid-discovery-corpus at ~8min; the log's ExitSuccess belongs to the merge-admission stamp tool stamping CI_FLOOR_EXIT=137. No flipped corpus has completed in CI (0/1 ci + 0/5 falsifier). Both carrier notes corrected. Calibration (coordinated with merry-owl-649's space-lens lane): - roster_import_closure_nodes_pre_resolve: shared closure-count authority (pure import walk, closure grain not entry grain), emitted BEFORE the heavy resolve so killed runs still yield the (nodes, peak) lower-bound pair - width-1 definition-drift oracle: pre-resolve walk must equal post-resolve resolved union on completed runs; refuses on divergence. Proven by execution: pre == post == 190/213 nodes across Off/Applied modes - kill-surviving cgroup memory.peak pre/post steps (post is always()) in the ci job and falsifier.yml; scope semantics labeled on the emission lines (reset=ok floor-scoped; span compares post>pre; never silently conflated) - job backstop timeouts extended by the two aux steps in both jobs Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Calibration pairs carry censored-vs-exact labels (methodology: killed runs are censored observations) Floor steps get id=floor; the always() post-peak step emits floor_outcome so each (closure_nodes, peak) pair is explicitly labeled: success = exact point, anything else = censored lower bound (true demand strictly greater than read). Prevents the fit from treating cap-kill reads as point estimates, which would drag the slope down and make the predictor underestimate - the dangerous direction (merry-owl methodology catch, 2026-07-10). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Review fixes: explicit witness import + complete step-budget ledger - ci_floor_plan_witness_test.dag: import witness_ci_corpus_discovery_serializes_at_width_one explicitly (cursor catch on #6438). The call resolved pre-fix via the seed resolver's flat namespace, so the witness ran green by execution; the explicit import restores the file's per-symbol import convention. - ci_workflow.dag: the resolve-receipt gate step had NO step cap — a hang there could only die by job-cancel (the #6323 starvation-kill class). It now carries the aux cap (script is a sub-second receipt read) and the ci job backstop counts four aux terms (peak pre/post, resolve-receipt gate, merge-admission gate): every step budgeted, backstop = step-sum + prelude (claude review catch on #6438, sharpened). - falsifier_workflow_witness_test.dag: falsifier_backstop_is_step_sum_plus_prelude asserted the pre-calibration formula — latent red proven by execution (FAIL receipt), fixed to the live two-aux sum, re-run PASS. - ci.yml regenerated byte-stable from the carrier (backstops 125->130, gate step timeout 5m). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Rebind calibration provenance comments to real artifacts (cursor review catch) The two cli_run.rs comments cited docs/plans/space-lens-minimal-project.md, which does not exist on main — the predictor design is in flight on PR #6442 (merry-owl-649's lane) and was never landed under that path. Rebound: the shared closure-definition authority is stated as this function itself, with the in-flight design cited by PR number and the landed parent-lane authorities cited by real paths (compute-envelope-model.md fleet envelope; input-envelope-roadmap.md admission). No behavior change; cargo check clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Remove stray empty file (shell-redirect artifact the auto-committer flushed) An internal-messaging command's backtick content was command-substituted by bash; a '-> fail-closed' fragment became a stdout redirect and created an empty file at the repo root, which the auto-committer then committed as 38f0a46. No tree content beyond the empty file; removing it restores the branch to e99c757's content. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * WIP: affected set processing * Scaffold-mark the cgroup peak calibration shell (cursor review catch) The three concat-built calibration runners (ci_cgroup_peak_locate_shell, ci_floor_peak_pre_script, ci_floor_peak_post_script) landed without the on-carrier scaffold markers repo convention requires for hand-shell. Each now carries a Disposition = Scaffold { dissolves_to: RealizationDispatch } row binding the decl (the ci_materialization pattern), and both scripts embed the shared dissolve-on note as a shell comment (the ci_spec pattern): dissolution = bash-emit (#5828 / gap-B emit(intent, Bash)) realizing the observation as an emitted ShellProgram intent or a typed host Observe effect. ci.yml + falsifier.yml regenerated; falsifier_workflow_witness_holds and the flip witnesses re-run PASS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Merge origin/main into session/loyal-wren-398 (resolve #6441 step-list conflict) Conflict resolution, all consciously declared: - ci_job steps: union — the calibration peak pre/post steps wrap the floor step (this branch) and #6441's materialization receipt gate slots between the resolve-receipt gate and the merge-admission gate (main). - The incoming materialization receipt gate step landed with timeout none — the same uncapped-step starvation-kill class this branch's review fix eliminated — so it now carries the aux cap, and the ci backstop counts FIVE aux terms (peak pre, peak post, resolve-receipt gate, materialization receipt gate, merge-admission gate); the budget disposition note records the merge provenance. - ci_run_step_natures_are_claims_counted_not_silent: RunStep count pin bumped 17 -> 19, acknowledging the two peak calibration steps #6441's count predates (conscious-count discipline; proven red at 17 then green at 19 by execution). - ci.yml regenerated from the merged carriers (backstops 130 -> 135). Verified on the merged tree: release bins rebuilt on merged Rust; falsifier workflow witness, flip witnesses, floor-plan/optin/width witnesses, and all 8 ci_materialization witnesses PASS; generated-artifact regen ExitSuccess. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * ShellProgram -> DAG: transport intent collapse + live importer ratchet (phase 0 of the sidecar dissolution) (#6449) * WIP: ShellProgram -> DAG * WIP: ShellProgram -> DAG * WIP: ShellProgram -> DAG * WIP: ShellProgram -> DAG * WIP: ShellProgram -> DAG * bash fold: native Concat/CmdSubst/WithRedir coverage + measured cost wall on the whole-tree emit path Fold-family productions extended so the fold no longer refuses word Concat/CmdSubst or stmt WithRedir (all four Redir variants): new concat_parts/word-compound/with_redir production families, lex tokens, kind-tag emit transforms, and recursive bundle arms. Byte-identity vs serialize_bash proven by execution: five depth-2 oracle tests plus the depth-4 assign_root_stmt manual probe (ROOT=$('git' 'rev-parse' '--show-toplevel' 2>/dev/null || 'pwd') byte-exact). The delegated fail-closed RED control repoints from WithRedir (now native) to Heredoc (still delegated) so the boundary guard stays discriminating. Measured cost wall, declared on-carrier (bash_program_emit_cost_wall_note): whole-tree backward row-selection is ~alternatives^depth (1s flat stmt, 64s for the single depth-4 stmt, DNF >8min for the full witness_bin program) because formal_production_unique_lhs_exact_match deep-validates every candidate per level and the descent re-validates each level again. Real-program oracles therefore stay MANUAL probes, not test fns (a discovery-run test would hang the local floor); the probe note on the test carrier names the dissolution triggers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * B1: NameResolutionPolicy row + position-tracked resolve (NamespaceOnlyY gated). Add v2.std.resolution_policy (ImportScoped default | NamespaceOnlyY). Thread ResolveContext { position, expected, policy } through resolve walk; namespace-only skips import module bindings and uses symbol_index at position. Policy pilot witness: green under NamespaceOnlyY at type position, RED under ImportScoped at module position. Import-scoped global default unchanged — zero corpus churn. Co-authored-by: Cursor <cursoragent@cursor.com> * Wave 1A - namespace-only name resolution: make the syntactic containment tree the single naming authority (qualified name = nesting position, reference = lexical lookup up ancestors, . = projection one level down). Path: confirm loyal-heron SymbolIndex scaling receipt FIRST, then SymbolIndex = conta (#6451) * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * v2.std.symbol_index: materialize containment tree as single naming authority Add SymbolIndex fill from nesting (qualified path → Node), qualified-name path algebra bridges, and discriminating witnesses. Retarget #6436 variant- visibility scaffold to dissolve into symbol_index_lexical_lookup; harvest_unique stays interim until module-scoped index scan lands. Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * Fix rust fmt on qualified-name bridge host functions (CI rust_tests gate). Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * Remove symbol_index_has_path; tests match symbol_index_lookup directly Dissolve Optional→Bool predicate in new std/ surface per review. Lexical lookup root termination already uses qualified_name_is_empty (not dotted string projection). Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * Address review: is_empty authority, host scaffold binds, layer split - Remove qualified_name_is_empty; lexical lookup uses is_empty(xs: position) - Host dispositions bind to from/to_dotted_string bridges; add P5 receipt tests - Move extdeps-coupled fill to v2.compiler.symbol_index_fill (layer DAG fix) Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * Wire resolver through SymbolIndex; dissolve harvest_unique_disj interim. Build SymbolIndex at admission and thread it through Namespace. resolve_atom falls back to symbol_index_lexical_lookup for unbound atoms after import-scoped lookup_chain. Fill-time unique-variant aliases (suffix-scan equivalent) replace #6436 harvest_unique_disj. Equivalence witnesses prove SymbolIndex-alone covers variant visibility (green + without-alias RED control); end-to-end wire green. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Brian Searls <briansearls1@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com> * WIP: Wave 1A - namespace-only name resolution: make the syntactic containment * Drop unused ResolveContext.expected until expected-type lane lands. Removes the dead field and uncalled resolve_ctx_with_expected helper flagged in #6454 review 36717 — B1 uses position-only disambiguation; expected-type filtering returns when that slice is scoped. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: gunbai-bot[bot] <289086189+gunbai-bot[bot]@users.noreply.github.com> Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Brian Searls <briansearls1@gmail.com> Co-authored-by: Brian Searls <11205878+briansrls@users.noreply.github.com>

briansrls and others added 12 commits July 9, 2026 22:17

WIP: ci floor measurement

42bfc38

WIP: ci floor measurement

da3f0ac

Merge remote-tracking branch 'origin/main' into session/merry-owl-649

dba467d

Merge remote-tracking branch 'origin/main' into session/merry-owl-649

674381f

WIP: ci floor measurement

533e473

WIP: ci floor measurement

ee01f00

WIP: ci floor measurement

ed8f9c6

Merge remote-tracking branch 'origin/main' into session/merry-owl-649

745a283

WIP: ci floor measurement

f3beae2

Merge remote-tracking branch 'origin/main' into session/merry-owl-649

4291580

gunbai-bot Bot changed the title ~~ci floor measurement~~ Plan: space lens — confident memory prediction from the static .dag (Node B walking skeleton) Jul 10, 2026

gunbai-bot Bot closed this Jul 10, 2026

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

Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438

Merged

3 of 4 tasks

Navigation Menu

Repository navigation

Plan: space lens — confident memory prediction from the static .dag (Node B walking skeleton) - #6442

Plan: space lens — confident memory prediction from the static .dag (Node B walking skeleton)#6442
gunbai-bot[bot] wants to merge 13 commits into
mainfrom
session/merry-owl-649

gunbai-bot Bot commented Jul 10, 2026 •

edited

Loading

Uh oh!

gunbai-bot Bot commented Jul 10, 2026

Uh oh!

gunbai-bot Bot commented Jul 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

gunbai-bot Bot commented Jul 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this is

Key finding — this is a resume, not a greenfield

Decisions I need from you

Uh oh!

gunbai-bot Bot commented Jul 10, 2026

Uh oh!

gunbai-bot Bot commented Jul 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

gunbai-bot Bot commented Jul 10, 2026 •

edited

Loading