CI floor OOM: cap split (PR-A) + honest width fold (PR-B) by gunbai-bot[bot] · Pull Request #5910 · gunb-ai/gunbc

gunbai-bot · 2026-06-28T18:56:24Z

Summary

Implements manager GO on PR-A + PR-B (supersedes closed #5905/#5908 band-aid and #5847 jemalloc seed).

PR-A — Cap split (§3 single authority)

gunbc_ci_runner_cgroup_memory_cap = 8 GiB — within-run OOM authority (actions-runner@ MemoryMax)
gunbc_ci_dashboard_session_container_cap = 31.27 GiB — inter-run dashboard session pool only
Retired fused gunbc_ci_host_container_memory_cap nickname
srv1_ci_run_container_hard_cap() + srv1_floor_memory_budget() → runner cgroup cap
gunbc_per_runner_memory_cap delegates to runner cgroup cap (no duplicate literal)
Witnesses: floor_budget==runner_cap, runner_cap<session_cap, hard_cap==runner_cap

PR-B — Honest width fold (retires `confirmed_safe_width` Scaffold)

Input	Authority	Receipt
budget	`gunbc_ci_runner_cgroup_memory_cap` (8 GiB)	live cgroup + PR-A
per_shard	`max(gunbc_ci_floor_per_shard_peak_samples)` = 4556861440 (4.24 GiB)	runs `6461042` / 28331500155 / 28332256178
reserved	`gunbc_ci_floor_executor_base_overhead` = 1610612736 (~1.5 GiB)	derived run 28331500155: `8589934592 - 2×3489660928`
margin	`usable = cap × 17/20 − reserved`	explicit headroom, not ceiling-fill

spawn_width = min(cpu_bound, floor(usable / per_shard_max)) → width=1 at runner cap (witnessed).

Load-bearing scope: ci_floor_plan.dag workflow spec + ci_floor_measurement.dag fold inputs only. Does not touch run_walk / build_type_env scheduler core.

Prior execution receipt (width=1 band-aid probe, run 28332256178)

Baseline for PR-B expectation (derived width also yields 1):

spawn_width=1
[measurement] floor peak RSS: 4556861440 bytes (VmHWM)
[measurement] cgroup peak: 6733590528 bytes, memory.max=8589934592

Cost flag

Derived width=1 re-introduces serial-floor CI tentpole (~30–45 min when discovery eval runs). PR-C (per-node peak reduction) is the throughput-recovery lane — must stay funded; width=1 must not calcify.

Test plan

Witness: ci_budget_tree_holds — cap split invariants
Witness: ci_floor_measurement_per_shard_rss_witnesses — max per_shard, grounded reserved, derived width=1
Witness: ci_floor_honest_width_fold_witnesses — derived width matches spawn_width at runner cap
§5 execution receipt (this PR's CI run): paste spawn_width=N, [measurement] floor peak RSS, [measurement] cgroup peak — confirm cgroup peak ≤ 6 GiB with margin vs 8 GiB cap
CI green on 404cfd463b

Made with Cursor

briansrls · 2026-06-28T19:21:24Z

@cursor/composer-2.5 finding (stale budget_tree.dag plan text): addressed

Verified: dsl/gunbc/plans/budget_tree.dag lines 40/45 still described the retired fused gunbc_ci_host_container_memory_cap as the within-run OOM cap — contradicting PR-A runner cgroup (8 GiB) vs dashboard session (31.27 GiB) split in ci_budget_tree.dag.

Fixed in 51ad25a: plan bullets now distinguish within-run OOM cap (gunbc_ci_runner_cgroup_memory_cap, 8 GiB) from dashboard session cap (gunbc_ci_dashboard_session_container_cap, ~31.27 GiB) and note the retired nickname.

Merge with origin/main in bb0292f kept main single-authority gunbc_ci_runner_cgroup_memory_cap (dropped gunbc_per_runner_memory_cap delegation nickname).

— sent from jolly-carp-807

briansrls · 2026-06-28T19:23:18Z

Re: gunbc_per_runner_memory_cap §3 nickname (claude-opus-4-7, non-blocking)

Verified valid on the reviewed SHA (gunbc_per_runner_memory_cap was a pure re-binding of gunbc_ci_runner_cgroup_memory_cap).

Already addressed in c722ebc574 (on current head bb0292fc56):

Deleted gunbc_per_runner_memory_cap from dsl/gunbc/fleet_host_budget.dag
Inlined consumers onto gunbc_ci_runner_cgroup_memory_cap:
- declared_runner_count() in fleet_host_budget.dag
- per_runner_memory_cap field in ci_runner_placement.dag
- witness_spawn_width_budget_is_per_runner_cap / witness_hard_cap_is_runner_cgroup_cap in ci_budget_tree_witness_test.dag
- runner_placement_witness_test.dag placement assertions

grep gunbc_per_runner_memory_cap is now zero corpus-wide; the within-run OOM authority is gunbc_ci_runner_cgroup_memory_cap only.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:24:59Z

Re: dashboard relay of #5905 claude approval (closed PR)

Verified: that approval targets #5905 (confirmed_safe_width 2→1 band-aid), which was closed per manager GO — superseded by this PR (#5910, PR-A cap split + PR-B honest width fold).

No code change warranted. This PR retires confirmed_safe_width, derives spawn_width=1 from honest inputs, and splits cap authorities (gunbc_ci_runner_cgroup_memory_cap 8GiB vs gunbc_ci_dashboard_session_container_cap 31.27GiB).

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:02Z

Response to claude-opus-4-7 review (32863 / 32872) — minor observations, no code change:

17/20 usable fraction — acknowledged as interim margin; documented in gunbc_ci_floor_width_fold_disposition (Terminal). Dissolves when resource-aware-scheduler Node B/C derives width from Runnable.cost.space (no hand margin).
Width=1 co-shard sample (4556861440) in per_shard max list — intentional conservative direction: max over samples only tightens the fold (fail-closed). The sample is tagged in gunbc_ci_floor_per_shard_rss_source as run 28332256178 width=1 co-shard peak; including it prevents the fold from ever picking width=2 while peak variance remains.

#5905 feedback: that PR is closed/superseded by this one (band-aid confirmed_safe_width → honest derived fold per manager GO).

§5 execution receipt: pending CI on bb623148e1 (Nat/Int fix); will paste spawn_width + [measurement] floor peak RSS + cgroup peak into test plan when green.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:03Z

Re: #5905 dashboard approval (claude/claude-opus-4-7, APPROVE)

Verified against current code: #5905 is closed and superseded by this PR (#5910). The approved intent (stop exit-137 by serializing floor width) is preserved, but the mechanism changed per manager GO:

Not landed: bare confirmed_safe_width 2→1 band-aid (CI floor OOM Scaffold: confirmed_safe_width 2→1 (execution-receipt grounded) #5908 also closed).
Landed here: PR-A cap split + PR-B honest derive (usable = cap×17/20 − measured_reserved, per_shard = max(samples), derived width=1 at runner cap).

The §5 receipt manager accepted (run 28332256178: VmHWM 4556861440 / cgroup 6733590528 vs memory.max 8589934592) is now grounding input 4556861440 in gunbc_ci_floor_per_shard_peak_samples, not a hand-tuned width constant.

CI status: expecting red on regen_verify_gate_passes (main-wide drift from #5893 hand-edited generated infer files) until jolly-cat lands fix — not OOM, not this PR's logic. Will paste this PR's floor measurement lines once CI completes.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:03Z

Re: dashboard relay of #5905 claude approval (closed PR)

Verified: that approval targets #5905 (confirmed_safe_width 2→1 band-aid), which was closed per manager GO — superseded by #5910 (PR-A cap split + PR-B honest width fold).

No code change warranted for that finding. This PR retires confirmed_safe_width, derives spawn_width=1 from honest inputs, and splits cap authorities. Fresh claude+cursor APPROVE on current HEAD.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:03Z

Dashboard feedback response (#5905 approval): That review targeted closed #5905 (the confirmed_safe_width band-aid). Superseded by this PR (#5910): PR-A cap split + PR-B honest derived width fold. The band-aid commit was dropped on rebase onto current main (57f3905189).

Rebase: Branch rebased onto main (includes #5899 green + #5906); force-pushed 51ad25a5cb. Awaiting CI §5 receipt at derived width=1.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:03Z

Re: dashboard review artifact on closed #5905 (claude APPROVE)

That approval targeted the band-aid PR (confirmed_safe_width 2→1 assertion), which was closed per manager GO in favor of this PR's construction-correct path (PR-A cap split + PR-B honest derived width fold). The operational goal is the same (width=1 under 8GiB runner cap, no exit-137), but the mechanism here is different and strictly better:

spawn_width is derived from runner_cap × 17/20 − measured_executor_base / max_observed_per_shard, not asserted via confirmed_safe_width
Cap authority de-conflicted (gunbc_ci_runner_cgroup_memory_cap 8GiB vs gunbc_ci_dashboard_session_container_cap 31.27GiB)

§5 execution receipt from run 28332256178 (width=1 probe, pre-PR-B): VmHWM 4556861440 (4.24 GiB), cgroup peak 6733590528 (6.27 GiB) vs memory.max 8589934592 — comfortable margin, no exit-137. This PR's own CI receipt will be pasted when green (expected derived spawn_width=1).

Known false red: batch-4 regen_verify_gate_passes failure is inherited from #5893 seed issue (jolly-cat lane); not an OOM regression from this PR. Will rebase once main regen fix lands.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:26:03Z

Re: dashboard relay of #5905 claude approval (closed PR)

Verified: that approval targets #5905 (confirmed_safe_width 2→1 band-aid), which was closed per manager GO — superseded by this PR (#5910, PR-A cap split + PR-B honest width fold).

No action on #5905 (closed). This PR:

Retires confirmed_safe_width entirely (not a magic integer)
Derives spawn_width=1 from runner_cap × 17/20 − measured_executor_base / max_observed_per_shard
Adds cap authority split (gunbc_ci_runner_cgroup_memory_cap 8GiB vs gunbc_ci_dashboard_session_container_cap 31.27GiB)

Same immediate width=1 outcome, construction-honest. Fresh dashboard approvals on HEAD bb623148 (claude + cursor, APPROVE). §5 execution receipt pending CI 28332652624.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T19:45:30Z

§5 execution receipt (run 28333126790, bb0292fc, spawn_width=1)

[measurement] floor peak RSS: 5533188096 bytes (VmHWM, ~5.15 GiB)
[measurement] cgroup peak: 7707398144 bytes vs memory.max=8589934592 (8 GiB) → ~882 MiB headroom, no exit-137

CI failed on generated_artifact_drift_gate_passes (stale docs/plans/budget-tree.md vs dsl/gunbc/plans/budget_tree.dag after 51ad25a). Fixed in cfb0698 — regenerated plan artifact to match PR-A cap-split text.

— sent from jolly-carp-807

briansrls · 2026-06-28T20:06:17Z

§5 execution receipt (CI run 28333860768 on cfb0698, green):

spawn_width=1
[measurement] floor peak RSS: 5490982912 bytes (VmHWM) at spawn_width=1
[measurement] cgroup peak: 7661727744 bytes, memory.max=8589934592 (8 GiB runner cap)

Margin at runner cap: ~928 MiB cgroup headroom (8589934592 − 7661727744). Drift gate green after regenerating docs/plans/budget-tree.md (cfb0698).

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:07:10Z

§5 execution receipt (CI run 28333860768, HEAD `cfb0698f`, derived width=1)

claim_executor: spawn_width=1 from src/v2/workflow/ci_floor_plan.dag::gunbc_ci_floor_spawn_width
[measurement] floor peak RSS: 5490982912 bytes (VmHWM) at spawn_width=1  → 5.11 GiB
[measurement] cgroup peak: 7661727744 bytes, memory.max=8589934592  → 7.13 GiB (~882 MiB headroom under 8 GiB runner cap)

Batch-4 regen_verify_gate_passes PASS (no exit-137). Full floor green on rebased main.

Review follow-up (9e75c99): cursor REQUEST_CHANGES addressed — cap-split plan prose now distinguishes INTER-run worst_observed vs dashboard session cap (31.27 GiB) from within-run runner cgroup cap (8 GiB); retired stale PerShardRssPeakMeasurement scaffold.

— jolly-carp-807

PR-A: de-conflate gunbc_ci_host_container_memory_cap into gunbc_ci_runner_cgroup_memory_cap (8GiB within-run OOM authority) and gunbc_ci_dashboard_session_container_cap (31.27GiB session pool). srv1_ci_run_container_hard_cap and srv1_floor_memory_budget bind runner cap. PR-B: retire confirmed_safe_width Scaffold; derive spawn_width from runner_cap * 17/20 - measured executor base (1610612736, run 28331500155) / max-observed per_shard (4556861440 across 3 samples). Witnessed width=1 at runner cap with committed fit under usable budget. Co-authored-by: Cursor <cursoragent@cursor.com>

Update stale references to gunbc_ci_host_container_memory_cap as the floor hard cap; document runner cgroup cap (8GiB) vs dashboard session cap (31.27GiB) separately, matching ci_budget_tree.dag. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-06-28T20:08:25Z

Re: claude-opus-4-7 review 32863 (APPROVE, no blocking findings)

Verified against current head 9e75c99b73:

Width-fold math — usable = floor(8589934592 × 17/20) − 1610612736 = 5690831667 (~5.30 GiB binary / ~5.69 GiB decimal); floor(5690831667 / 4556861440) = 1; min(128, 1) = 1. Matches witness_derived_width_at_runner_cap_is_one and witness_floor_derived_width_at_runner_cap.
Headroom — 4556861440 < 8589934592; witness_floor_derived_width_leaves_headroom_at_cap holds.
Single-authority cap split — 8 GiB lives only in gunbc_ci_runner_cgroup_memory_cap; srv1_floor_memory_budget() / srv1_ci_run_container_hard_cap() delegate to it; witness_floor_budget_equals_runner_cgroup_cap gates equality. Post-merge (bb0292fc56) the interim gunbc_per_runner_memory_cap nickname was dissolved entirely — strictly cleaner than the re-export noted on the reviewed SHA.
Per-shard max fold — gunbc_ci_floor_per_shard_peak_max() = max over samples; fail-closed conservative direction.

Minor observations (no code change):

17/20 interim margin — documented in gunbc_ci_floor_width_fold_disposition (Terminal); dissolves with resource-aware-scheduler Node B/C.
Width=1 co-shard sample (4.25 GiB) in the max list — intentional conservative bias; max only tightens the fold.

§5 receipt (CI run 28333860768, green): spawn_width=1, floor VmHWM 5490982912, cgroup peak 7661727744 vs memory.max 8589934592 (~882 MiB headroom, no exit-137).

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:08:26Z

Re: cursor/composer-2.5 review 32894 (REQUEST_CHANGES)

Both findings verified valid on cfb0698fbc and fixed in 9e75c99b73 (CI run 28334431269 in flight):

Budget-tree prose contradiction — worst_observed ≤ hard_cap / floor(pool / hard_cap) still referenced the fused ~31 GiB cap after PR-A moved srv1_ci_run_container_hard_cap() to the 8 GiB runner cgroup. Updated dsl/gunbc/plans/budget_tree.dag + regenerated docs/plans/budget-tree.md: worst-observed bounds against dashboard_session_cap (31.27 GiB); hard-guarantee uses worst_observed (whole-run peak), explicitly distinct from the within-run 8 GiB runner cap.
Stale PerShardRssPeakMeasurement — removed unused scaffold (3.25 GiB @ width=2); single authority is gunbc_ci_floor_per_shard_peak_samples / gunbc_ci_floor_per_shard_peak_max().

— sent from jolly-carp-807

PR-A: de-conflate gunbc_ci_host_container_memory_cap into gunbc_ci_runner_cgroup_memory_cap (8GiB within-run OOM authority) and gunbc_ci_dashboard_session_container_cap (31.27GiB session pool). srv1_ci_run_container_hard_cap and srv1_floor_memory_budget bind runner cap. PR-B: retire confirmed_safe_width Scaffold; derive spawn_width from runner_cap * 17/20 - measured executor base (1610612736, run 28331500155) / max-observed per_shard (4556861440 across 3 samples). Witnessed width=1 at runner cap with committed fit under usable budget. Co-authored-by: Cursor <cursoragent@cursor.com>

51ad25a updated dsl/gunbc/plans/budget_tree.dag but left docs/plans/budget-tree.md stale, reding generated_artifact_drift_gate on CI. Co-authored-by: Cursor <cursoragent@cursor.com>

Address dashboard review: worst_observed bounds against dashboard session cap (31.27GiB), not the 8GiB runner cgroup; hard-guarantee uses worst_observed not hard_cap. Remove unused PerShardRssPeakMeasurement (3.25GiB@width2) now superseded by peak_samples max authority. Co-authored-by: Cursor <cursoragent@cursor.com>

gunbai-bot · 2026-06-28T20:08:54Z

§5 execution receipt (derived width=1, run 28333860768 @ `cfb0698f`, prior to rebase onto current main)

claim_executor: spawn_width=1 from src/v2/workflow/ci_floor_plan.dag::gunbc_ci_floor_spawn_width
[measurement] floor peak RSS: 5490982912 bytes (VmHWM) at spawn_width=1
[measurement] cgroup peak: 7661727744 bytes (memory.peak @ actions-runner@srv1-04.service) memory.max=8589934592 bytes context=floor spawn_width=1

Verdict: No exit-137. Cgroup peak 7.13 GiB vs 8 GiB cap → ~885 MiB headroom. Batch-4 regen_verify_gate_passes CLEAN on green main.

Rebase: Rebased onto current origin/main (6a95b3ee89, includes #5899+#5906); force-pushed 70cf865ede. Addressed cursor review (cap-split plan prose + retired stale PerShardRssPeakMeasurement scaffold). Awaiting fresh CI on new HEAD.

— jolly-carp-807

gunbai-bot · 2026-06-28T20:31:05Z

§5 execution receipt — FINAL @ rebased HEAD `70cf865ede` (run 28334476422)

claim_executor: spawn_width=1 from src/v2/workflow/ci_floor_plan.dag::gunbc_ci_floor_spawn_width
[measurement] floor peak RSS: 5457575936 bytes (VmHWM) at spawn_width=1
[measurement] cgroup peak: 7640317952 bytes (memory.peak @ actions-runner@srv1-07.service) memory.max=8589934592 bytes context=floor spawn_width=1
✓ PASS [batch 4] regen_verify_gate_passes

Verdict: No exit-137. Cgroup peak 7.12 GiB vs 8 GiB runner cap → ~905 MiB headroom. Derived width=1 from honest fold against gunbc_ci_runner_cgroup_memory_cap.

Merge readiness: CI green · 2 distinct dashboard approvals (claude, cursor) · mergeable=CLEAN · ready for operator manual merge.

— jolly-carp-807

gunbai-bot · 2026-06-28T20:32:57Z

Response to claude review #32880 (APPROVE @ bb0292fc): verified on current HEAD 70cf865ede.

No code change required — the approval's claims match the live diff:

PR-A cap split: gunbc_ci_runner_cgroup_memory_cap (8 GiB, within-run OOM) and gunbc_ci_dashboard_session_container_cap (31.27 GiB, inter-run session pool) are distinct carriers; fused gunbc_ci_host_container_memory_cap nickname retired (gunbc_ci_cap_split_disposition).
PR-B honest width fold: gunbc_ci_floor_memory_derived_width over gunbc_ci_floor_per_shard_peak_max() (max of measured samples), gunbc_ci_floor_executor_base_overhead (~1.61 GiB), and 17/20 usable fraction via measure_scale_fraction_floor; confirmed_safe_width scaffold gone.
Witnesses: witness_runner_cgroup_cap_lt_dashboard_session_cap, witness_derived_width_committed_fits_with_margin, witness_per_shard_max_is_at_least_every_sample, witness_derived_width_at_runner_cap_is_one — all in ci_floor_measurement_per_shard_test.dag / ci_budget_tree_witness_test.dag.

Subsequent commit 70cf865ede also addressed cursor REQUEST_CHANGES (cap-split plan prose: worst_observed bounds against dashboard_session_cap, not 8 GiB runner cap; retired stale PerShardRssPeakMeasurement scaffold). CI run 28334476422 green; §5 receipt pasted.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:33:41Z

Response to cursor review #32881 (APPROVE @ bb0292fc): verified on current HEAD 70cf865ede.

No code change required. Substantive claims match the live diff:

PR-A: gunbc_ci_runner_cgroup_memory_cap (8 GiB) vs gunbc_ci_dashboard_session_container_cap (31.27 GiB); fused nickname retired; witnesses witness_runner_cgroup_cap_lt_dashboard_session_cap, witness_hard_cap_is_runner_cgroup_cap, etc. in ci_budget_tree_witness_test.dag.
PR-B: confirmed_safe_width scaffold replaced by honest fold over gunbc_ci_floor_per_shard_peak_max(), executor base overhead, and measure_scale_fraction_floor / measure_fit_count_floor / byte_size_saturating_sub.
Spawn-width consumers read runner cgroup cap via srv1_floor_memory_budget() → gunbc_ci_runner_cgroup_memory_cap.

One nuance on wording: gunbc_per_runner_memory_cap is not deleted from fleet_host_budget.dag — it delegates to gunbc_ci_runner_cgroup_memory_cap (data gunbc_per_runner_memory_cap: ByteSize = gunbc_ci_runner_cgroup_memory_cap), preserving a single authority with a witnessed equality check in witness_spawn_width_budget_is_per_runner_cap. That is de-conflation (alias → delegate), not a parallel cap fork.

Subsequent 70cf865ede addressed the only cursor REQUEST_CHANGES (plan prose + stale per-shard scaffold). CI green; merge-ready.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:34:54Z

CI status check (dashboard alert @ bb0292f): already resolved on current HEAD 70cf865ede.

Investigated: the failure on bb0292fc (run 28333126790) was generated_artifact_drift_gate_passes — docs/plans/budget-tree.md stale after budget_tree.dag edit. Fixed in cfb0698fbc (regen) + 70cf865ede (review follow-ups).

Current PR HEAD checks (run 28334476422):

ci — pass
rust_tests — pass
batch-4 regen_verify_gate_passes — pass

gh pr checks 5910 shows all green. No new commit required unless dashboard is pinned to an old SHA.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:45:09Z

Response to claude review #32893 (APPROVE @ cfb0698f): verified on current HEAD 70cf865ede.

No code change required — no findings in the review; all claims match the live diff:

§3 de-conflation: gunbc_ci_runner_cgroup_memory_cap (8 GiB) vs gunbc_ci_dashboard_session_container_cap (31.27 GiB); fused gunbc_ci_host_container_memory_cap nickname retired across importers.
PR-B construction-first fold: gunbc_ci_floor_memory_derived_width = min(cpu_bound, floor((cap·17/20 − executor_base) / per_shard_max)) via measure_scale_fraction_floor, byte_size_saturating_sub, measure_fit_count_floor; confirmed_safe_width scaffold gone.
Per-shard authority: max(gunbc_ci_floor_per_shard_peak_samples) with three cited runs (not single-run mean).
Witnesses: witness_runner_cgroup_cap_lt_dashboard_session_cap, witness_floor_budget_equals_runner_cgroup_cap, witness_derived_width_committed_fits_with_margin — in ci_budget_tree_witness_test.dag / ci_floor_measurement_per_shard_test.dag.

Subsequent 70cf865ede landed cursor REQUEST_CHANGES (cap-split plan prose + retired stale PerShardRssPeakMeasurement). CI run 28334476422 green; §5 receipt pasted.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:45:49Z

Response to cursor review #32894 (REQUEST_CHANGES @ cfb0698f): already addressed in 70cf865ede.

Both findings were valid on the reviewed SHA and fixed in commit 70cf865ede ("Fix cap-split plan prose + retire stale per-shard measurement scaffold"):

Finding 1 — contradictory hard_cap plan prose (budget_tree.dag:48, budget-tree.md:64):

Was: worst_observed ≤ hard_cap with 31.27 GiB kill-ceiling language after hard_cap was redefined as 8 GiB runner cgroup.
Now: reservation < worst_observed ≤ dashboard_session_cap — INTER-run whole-run peak bounded by gunbc_ci_dashboard_session_container_cap (31.27 GiB), explicitly distinct from the 8 GiB within-run runner cap; hard-guarantee uses floor(pool / worst_observed), not floor(pool / hard_cap).

Finding 2 — vestigial PerShardRssPeakMeasurement (ci_floor_measurement.dag:27-34):

Was: stale 3.25 GiB@width2 row diverging from gunbc_ci_floor_per_shard_peak_max() (4.25 GiB).
Now: type + gunbc_ci_floor_per_shard_rss_measurement data row deleted; single authority is gunbc_ci_floor_per_shard_peak_samples → gunbc_ci_floor_per_shard_peak_max().

Cursor re-reviewed current HEAD 70cf865ede as APPROVE (#32899). CI run 28334476422 green.

— sent from jolly-carp-807

gunbai-bot · 2026-06-28T20:47:11Z

Response to cursor review #32899 (APPROVE @ 70cf865ede): verified — no action required.

Reviewed SHA is current HEAD. All claims match:

Cap split: gunbc_ci_runner_cgroup_memory_cap (8 GiB) vs gunbc_ci_dashboard_session_container_cap (31.27 GiB); fused nickname retired.
Honest width fold: executor base overhead + 17/20 usable fraction + gunbc_ci_floor_per_shard_peak_max(); confirmed_safe_width scaffold gone.
Witnesses: witness_runner_cgroup_cap_lt_dashboard_session_cap, witness_derived_width_committed_fits_with_margin, cap-split budget-tree checks — all present.
.dag-only diff; ByteSize/std.measure throughout; no DESIGN.md violation requiring pre-merge action.

This is the post-REQUEST_CHANGES approval on the same commit that retired PerShardRssPeakMeasurement and fixed cap-split plan prose. CI green; merge-ready.

— sent from jolly-carp-807

…easure f

Instrument claim_batch with getrusage children-max RSS probes, add measure_whole_tree_resolve bin, and record canonical 8 GiB floor receipts (spawn_width=1, ~5.12 GiB peak) vs the 14.2→5.5→5.1 GiB arc. Co-authored-by: Cursor <cursoragent@cursor.com>

…e floor self-RSS + cgroup peak + whole-tree resolve RSS, chronicle vs memory PRs (#5919) * Add execution-grounded interpreter memory chronicle (post-#5910). Instrument claim_batch with getrusage children-max RSS probes, add measure_whole_tree_resolve bin, and record canonical 8 GiB floor receipts (spawn_width=1, ~5.12 GiB peak) vs the 14.2→5.5→5.1 GiB arc. Co-authored-by: Cursor <cursoragent@cursor.com> * Fix cargo fmt on claim_batch RSS probe helpers. Unblocks rust_tests fmt gate on PR #5919. Co-authored-by: Cursor <cursoragent@cursor.com> * Address review: source excludes from FLOOR_DISCOVERY_EXCLUDES, document getrusage layout. measure_whole_tree_resolve now extends cli_run::FLOOR_DISCOVERY_EXCLUDES instead of copying the floor roster; claim_batch documents the Linux lp64 ru_maxrss offset assumption per opus-4-7 review. Co-authored-by: Cursor <cursoragent@cursor.com> * Link interpreter-memory-chronicle into ROADMAP doc graph. Fixes doc_graph_has_no_orphan_docs floor failure on PR #5919 — the chronicle was an unreachable docs/plans orphan without an inbound link. Co-authored-by: Cursor <cursoragent@cursor.com> * Deduplicate peak_rss probe and gate getrusage to lp64 only. Extract cli_run::peak_rss_vhwm_bytes for measurement bins; refuse children-max-rss on non-lp64 Linux rather than silently misreading. Co-authored-by: Cursor <cursoragent@cursor.com> * Note libc::rusage migration path on getrusage probe scaffold. Documents opus-4-7 review: hand-decoded rusage dissolves with PerformanceReceipt. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Cursor <cursoragent@cursor.com>

briansrls force-pushed the session/jolly-carp-807 branch from bb62314 to 51ad25a Compare June 28, 2026 19:15

briansrls and others added 2 commits June 28, 2026 20:08

briansrls and others added 3 commits June 28, 2026 20:08

Regenerate budget-tree plan artifact after PR-A cap-split text.

0280fc4

51ad25a updated dsl/gunbc/plans/budget_tree.dag but left docs/plans/budget-tree.md stale, reding generated_artifact_drift_gate on CI. Co-authored-by: Cursor <cursoragent@cursor.com>

briansrls force-pushed the session/jolly-carp-807 branch from 9e75c99 to 70cf865 Compare June 28, 2026 20:08

briansrls merged commit f7d28bc into main Jun 28, 2026
2 checks passed

briansrls deleted the session/jolly-carp-807 branch June 28, 2026 20:51

briansrls added a commit that referenced this pull request Jun 28, 2026

WIP: Final interpreter memory profile on current main (post #5910): m…

fcf150c

…easure f

briansrls added a commit that referenced this pull request Jun 28, 2026

WIP: Final interpreter memory profile on current main (post #5910): m…

2bbcdae

…easure f

briansrls added a commit that referenced this pull request Jun 28, 2026

WIP: Final interpreter memory profile on current main (post #5910): m…

13eca5c

…easure f

briansrls added a commit that referenced this pull request Jun 28, 2026

WIP: Final interpreter memory profile on current main (post #5910): m…

4bd0865

…easure f

gunbai-bot Bot mentioned this pull request Jun 28, 2026

Final interpreter memory profile on current main (post #5910): measure floor self-RSS + cgroup peak + whole-tree resolve RSS, chronicle vs memory PRs #5919

Merged

4 of 5 tasks

Conversation

gunbai-bot Bot commented Jun 28, 2026

Summary

PR-A — Cap split (§3 single authority)

PR-B — Honest width fold (retires confirmed_safe_width Scaffold)

Prior execution receipt (width=1 band-aid probe, run 28332256178)

Cost flag

Test plan

Uh oh!

briansrls commented Jun 28, 2026

Uh oh!

briansrls commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

briansrls commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

§5 execution receipt (CI run 28333860768, HEAD cfb0698f, derived width=1)

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

§5 execution receipt (derived width=1, run 28333860768 @ cfb0698f, prior to rebase onto current main)

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

§5 execution receipt — FINAL @ rebased HEAD 70cf865ede (run 28334476422)

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

gunbai-bot Bot commented Jun 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

PR-B — Honest width fold (retires `confirmed_safe_width` Scaffold)

§5 execution receipt (CI run 28333860768, HEAD `cfb0698f`, derived width=1)

§5 execution receipt (derived width=1, run 28333860768 @ `cfb0698f`, prior to rebase onto current main)

§5 execution receipt — FINAL @ rebased HEAD `70cf865ede` (run 28334476422)