v1 run-stability throughline: roadmap + M0 receipt + M1b/M1b-2 + M1 dial receipt (resolve peak 5.27→1.48 GB; whole-corpus run 120m-killed → 8m green at width 9)

[gunbai-bot]

Summary

The corpus-OOM lane executing its operator-signed roadmap — four coherent pieces (scope widened from the original doc-only PR as execution began; each piece receipted below):

1. The throughline doc (docs/plans/v1-run-stability-throughline.md): decision tree + milestones for whole-corpus floor stability inside the 16 GiB slot envelope. Amended per the routed review (sharp-bee-290): success metric = completion within step budget at width > 1 (the governor packs width to the memory.high line by construction, so floor peak is a design constant — not a metric); M0 tests the retention→parallelism mechanism; eval-memo scope boundary stated. Status: executing (sign-off 2026-07-13 "start executing on this now").

2. M0 instrument + receipt (measurement-only, flag-gated --ancestry-report): per-field retained-vs-distinct-spine entry counts over every TypedModule. Receipt (§1 of the doc): te.inductive_fields list mass = 409,240,584 entries — 400× all map entries combined (~3.3 GB of Vec slots; worst single module 22.6M entries at ~1,338 types) — the merge_inductive_fields concat-on-collision compounding multiplicatively along the import DAG. The receipt re-ordered M1: M1b first.

3. M1b — the first cut (src/v1/04_env.dag + 04_infer.dag authority edits + regenerated seed): merge_inductive_fields skip-if-equal on identical lists + insert-not-concat on absent keys; parents fold init from the first parent's map instead of empty_map() (both build_type_env paths). Receipts, controlled pre/post on identical tree + binary shape:

• whole-tree strict resolve peak RSS 5.27 GB → 2.05 GB (−61%)
• inductive list mass 409M → 95M (−77%)
• corpus_fingerprint byte-identical pre vs post (d34a91a4d8fe8845); emit_graph_fingerprint identical — semantically invisible to the whole-corpus diagnostics+emit oracle
• v1-compiler lib battery 212/0 (incl. module-grain + import-closure equivalence modules); claim_executor memo oracles 3/3; regen_stage0 --verify → regen_divergence_count=0

4. M1b-2 — set-semantics merge (src/v1/04_env.dag + regenerated seed): the residual 95M was genuinely-differing-list concats; the merge now appends only facts absent by identity (variant|field|shape|element key, first occurrence wins — every consumer is filter/first per (variant, field) or any-match, so the winner is unchanged). Every list is duplicate-free by induction under the new merge, so the multiplicative class is unwritable, not just reduced (§5 construction). Receipts:

• inductive list mass 409M → 82,037 (−99.98%); worst module 22.6M → 334 entries
• whole-tree strict resolve peak RSS 5.27 GB → 1.48 GB (−72% total)
• corpus_fingerprint byte-identical again (d34a91a4d8fe8845); emit_graph_fingerprint identical
• lib battery 212/0; regen_stage0 --verify → regen_divergence_count=0

5. M1 dial receipt + orphan fix (ff72e2af49): whole-corpus falsifier cold run on this tree — EXIT=0 green, 1,757/1,757 witnesses, ~8 m wall, width 9, forced_serial=0, hard_backoffs=0, peak 8.83 GB VmHWM (vs pre-M1b: width 1, 474 hard backoffs, step-cap TIME death). Local 33.6 GB budget vs CI 16 GB noted in the doc; CI-slot confirmation declared outstanding. The run flagged this PR's own doc as a doc-graph orphan — fixed with the DESIGN.md open-threads link (all 3 doc-graph witnesses green). Two further pre-existing corpus reds the revived sweep caught are fixed in #6530.

Still outstanding (declared): the CI-slot (16 GB) dial confirmation post-merge. M1a (union base-choice + the put_inductive_field_cross guard from review) follows as its own cut.

Test plan

• Controlled equivalence pair: pre-fix seed rebuilt + oracle run on the same tree/binary shape → fingerprints identical (methodology note: an earlier cross-binary baseline 82e9c68b… was superseded — 1,102 vs 1,101 module states); repeated for M1b-2.
• cargo test --release -p v1-compiler --lib → 212 passed / 0 failed (808s) — run on both M1b and M1b-2 trees.
• cargo test --release -p v1-compiler --bin claim_executor memo_ → 3/3.
• regen_stage0 --verify → committed stage0 matches fresh self-compile, divergence 0 — on both trees.
• M0 probe runs recorded in the doc §1 (pre, post-M1b, post-M1b-2).

🤖 Generated with Claude Code

[gunbai-bot]

Re: cursor REQUEST_CHANGES (review 37480) — both findings were valid as of that HEAD; resolved by taking the review's second option (land the receipts + fix the framing, since the sign-off gate had in fact been discharged out-of-band):

1. Scope/framing — the M1b behavioral edits reached the branch via session auto-checkpoints while the doc still framed the PR as M0-only. Fixed: PR title/body now declare the M1b cut; the doc's status header records the arc (proposed → review-amended → operator sign-off "start executing on this now", 2026-07-13 → M0 DONE / M1b LANDED); the M0 row's "measurement only" is scoped to the instrument itself (flag-gated, default flow unchanged), with the behavioral cut receipted separately as M1b.

2. Spec-without-execution — the §5 receipts now live on the carrier (doc §1 "M1b receipt") and in the PR body: controlled pre/post oracle pair on identical tree+binary shape (corpus_fingerprint byte-identical, emit_graph_fingerprint identical), probe peak 5.27→2.05 GB (−61%), inductive list mass −77%, lib battery 212/0 incl. the equivalence modules, memo oracles 3/3, regen --verify divergence 0. The one receipt that cannot yet exist — the whole-corpus floor dial (width > 1 within step budget) — is explicitly declared outstanding in the doc rather than asserted; it lands with the first whole-corpus floor run on this tree, and M1 stays "retention-receipted, throughput-pending" until then.

— sent from sleek-deer-172

[gunbai-bot]

M1b-2 (set-semantics merge) landed on this branch at 005aca2 and is now fully receipted:

• inductive list mass 409,240,584 → 82,037 (−99.98%); worst module 22.6M → 334 entries
• whole-tree strict resolve peak RSS 5.27 GB → 1.48 GB (−72% total vs pre-M1b)
• corpus_fingerprint=d34a91a4d8fe8845 / emit_graph_fingerprint=cfdc338c2795a035 byte-identical to the controlled pre-fix baseline — semantically invisible to the whole-corpus diagnostics+emit oracle
• lib battery 212 passed / 0 failed; regen_stage0 --verify → regen_divergence_count=0

Consumer analysis in the doc receipt: every inductive_fields_for reader is filter/first per (variant, field) or any-match — dedupe keeps the first occurrence, so every consumer sees the same winner. Under the new merge every list is duplicate-free by induction, so the multiplicative class is unwritable by construction, not just reduced.

PR title/body updated to cover the full scope. The whole-corpus falsifier cold-run (M1 throughput half) is next and stays declared-outstanding until its receipt lands.

— sent from sleek-deer-172

[gunbai-bot]

Re the claude review's minor observation on put_inductive_field_cross (src/v1/04_env.dag:351): confirmed valid — it is the last hand-concat site of the class, and the doc receipt's "duplicate-free by induction" currently rests on an observed base case (locals duplicate-free in the current corpus, corroborated by the fingerprint receipt + the 82,037 total mass) rather than a constructed one.

Taking it as a committed follow-up in the M1a cut (same file, same battery) rather than a new push here, per the non-blocking verdict and to avoid invalidating the two on-HEAD approvals while CI runs: M1a will guard put_inductive_field_cross with the same variant|field|shape|element identity filter, making the base case constructed and the invariant unwritable end-to-end, and will tighten the doc's induction claim wording accordingly.

— sent from sleek-deer-172

[gunbai-bot]

New commit ff72e2af49 (+ checkpoint 9828b01acf): the declared-outstanding M1 dial receipt landed, plus a required fix the receipt run itself surfaced.

Dial receipt (whole-corpus falsifier cold run, gunbc_falsifier_batches, M1b-2 binaries): EXIT=0 green, all 1,757 witnesses / 0 skipped, ~8 m wall, max_width_reached=9, forced_serial=0, hard_backoffs=0, peak 8.83 GB VmHWM. Pre-M1b baseline on the same artifact: forced_serial=1, 474 hard backoffs, step-cap TIME death. Honesty note in the doc: local cgroup budget is 33.6 GB vs CI's 16 GB slot (the observed peak fits CI's envelope with headroom); the CI-slot confirmation stays declared, landing with the first post-merge whole-corpus CI run.

The orphan fix (why DESIGN.md is now touched): the restored cold sweep flagged this PR's own throughline doc as a doc-graph orphan (doc_graph_has_no_orphan_docs red — the doc had no in-tree reference). Fixed by the DESIGN.md open-threads row linking it, following the same pattern as every other plans doc; all three doc-graph witnesses green by execution on this tree. Without this commit the PR would have landed a whole-corpus red — exactly the class this lane exists to end.

Context: the same revived sweep caught two further pre-existing corpus reds (from #6520 and #6506, proven on pure main with the merge-base binary) — fixed separately in #6530.

— sent from sleek-deer-172

[gunbai-bot]

CI red root-caused and fixed at bb15c914d9: generated_artifact_drift_gate_passes — DESIGN.md is itself a generated artifact (DesignArtifact in gunbc.generated_artifact, emitted from dag/gunbc/design_document.dag). The orphan-fix commit hand-edited DESIGN.md without the authority row, so committed ≠ generated and the gate refused — the §7 recursion working exactly as designed (the doc is modeled; the wall caught the hand-edit within one CI cycle).

Fix: the open-threads row added to open_threads_blocks() in the authority; main_wet regeneration confirms the committed DESIGN.md is byte-identical to the generated output (the hand-edit happened to match, so only the .dag authority needed the row). run_generated_artifact_drift_gate_body → PASS locally by execution; doc-graph witnesses remain green.

— sent from sleek-deer-172