friction: main checkout output/ holds 83.6 GB of untriaged generated artifacts

[ll7]

Summary

The main checkout's ignored output/ directory holds 83.6 GB across 124 entries and has never been triaged under the artifact-provenance policy ("output/ is temporary, worktree-local, untracked; promote durable artifacts, prune the rest"). Disk pressure is currently manageable (~115 GB free) but this is the single largest reclaimable pool on the machine.

Observed evidence (2026-09-02, main checkout)

Top entries by size (du -sh output/* | sort -rh | head):

size	entry
57G	output/slurm (SLURM job logs/scratch)
6.1G	output/issue2557-seed-variance
2.8G	output/issue4206-trace-rerun
2.8G	output/issue4206-13334-harvest
2.3G	output/issue3068-ppo-curriculum
1.9G	output/release_packet_audit_20260821
1.8G	output/issue5578-native-speed-tier
1.2G	output/issue4365-13376-harvest

Per check_worktree_capacity.py --inventory --json: category output, guidance "ignored generated output; preserve durable evidence before pruning".

Suggested change

Run a one-time artifact-provenance triage over output/: classify each top-level entry as disposable / ignored-cache / tracked-manifest / durable-required / handoff-needed; promote anything a future benchmark, report, or release needs (or record its canonical source); then prune the disposable remainder, starting with output/slurm logs older than the last audited campaign.

Boundaries

• This issue requests triage + a classification manifest, not blind deletion; durable-evidence preservation rules apply.
• Evidence grade: observed (sizes and paths above).
• No benchmark claims change; classified artifacts may later feed report provenance.

schema: goal_autopilot_preparation.v1
repository: ll7/robot_sf_ll7
issue: 8255
source_body_sha256: 08545adacc457fe2beb171ed6e287ff50d3dfb7116b62429269dfbcfc2175880
source_comments_sha256: 
audit_schema: open_issue_contract_audit.v1
audit_digest: fdd4b92c8bb28db0ca959e8680b5e4851f5ba5a323321ff4be1f2b82a96cdc19
audit_classification: state_conflict
next_action: formalize_issue
authority: lifecycle_owner
execution_mode: formalization
preferred_worker: LunaRunner
expected_pr_runner_label: runner:max
implementation_admitted: False
state_ready_change_proposed: False
readiness_gate: None
mutation_batch: goal-autopilot-01a061ac-cycle11-formalization

This packet is preparation evidence only. It never overrides live labels, exact claim state, branch state, typed dependencies, domain gates, compute authority, release authority, or scientific evidence rules.

Implementability Contract (formalization appended 2026-09-03; original sections above unchanged)

Archetype Metadata

archetype: workflow
evidence_tier: idea
linked_policy:
  - docs/context/issue_1512_issue_archetypes.md

Assumption noted: one-time triage process with a manifest deliverable (workflow). Tier idea: triage not yet executed; sizes above are supporting observations.

Scope

• In scope: read-only classification manifest over output/ top-level entries, then prune-plan and promotion follow-ups tracked as children below.
• Out of scope: blind deletion; benchmark/report/release claim changes.

Definition of Done

• Every top-level output/ entry is classified (disposable / ignored-cache / tracked-manifest / durable-required / handoff-needed).
• Durable-required items name their canonical source or promotion target.
• Disposable remainder pruned only after manifest review on this issue.

Validation / Testing

• python scripts/dev/check_worktree_capacity.py --inventory --json before/after (read-only confirmation).
• Manifest totals reconcile with du -sh output.

Affected Files

• output/slurm, output/issue2557-seed-variance, output/issue4206-trace-rerun, output/issue4206-13334-harvest, output/issue3068-ppo-curriculum, output/release_packet_audit_20260821, output/issue5578-native-speed-tier, output/issue4365-13376-harvest - Top entries by size; classification slice (workflow: output/ artifact-provenance classification manifest, no deletion (child of #8255) #8371).
• scripts/dev/check_worktree_capacity.py - Read-only inventory evidence command.

Next Implementable Children

• workflow: output/ artifact-provenance classification manifest, no deletion (child of #8255) #8371 - Classification manifest for output/ top-level entries, no deletion.
• workflow: bounded output/ prune pass on manifest-verified dispositions only (child of #8255) #8425 - Bounded prune pass on manifest-verified dispositions only (verify-then-delete).
• workflow: output/model_cache per-checkpoint triage manifest, no deletion (child of #8255) #8426 - output/model_cache per-checkpoint triage manifest, no deletion.

[ll7]

Artifact-provenance classification manifest for output/ (closes #8371)

Method (2026-09-03, main checkout, read-only): du -sh output/* | sort -rh
(full 131-entry inventory), stat mtimes, top-level content probes, and
check_worktree_capacity.py category guidance ("ignored generated output;
preserve durable evidence before pruning"). Total 78G (parent reported
83.6G on 2026-09-02; drift is expected — e.g. model_cache touched 2026-09-02,
issue_audit_plan.json rewritten today by active autopilot runs).
Top-8 sum = 75.9G; remainder ≈ 2.1G across 123 small entries. Totals reconcile
with du -sh output/ within rounding. Nothing was deleted or moved.
Confidence on campaign-dir dispositions is ~70%: prune candidates below are
conditional on per-issue evidence confirmation, not deletion orders.

Large entries (all >1G)

size	mtime	entry	classification	rationale / condition
56G	Aug 18	output/slurm/10-issue5346-s30-trustworthy-rerun-job-13529	durable-required (pending)	Trustworthy-rerun evidence for the #5346 S30 campaign; keep until that campaign's evidence is promoted or ruled superseded
~1G	Jun–Aug	output/slurm/ (remaining ~40 job dirs, largest 782M oracle-trace, 293M S30-canary)	prune-candidate (conditional)	Job logs/scratch older than the last audited campaign; prune only after confirming no open campaign cites them
6.1G	Jun 27	output/issue2557-seed-variance	prune-candidate (conditional)	Stale campaign scratch; prune after #2557 evidence check
2.8G	Jul 8	output/issue4206-trace-rerun	prune-candidate (conditional)	Rerun scratch superseded by issue4206-13334-harvest; prune after #4206 evidence check
2.8G	Jul 4	output/issue4206-13334-harvest	handoff-needed	Harvest bundle; confirm promotion target in #4206 before any prune
2.3G	Jul 14	output/issue3068-ppo-curriculum	prune-candidate (conditional)	Curriculum scratch; prune after #3068 evidence check
1.9G	Aug 30	output/release_packet_audit_20260821	durable-required	Recent release-packet audit input; keep until v0.0.6 release disposition
1.8G	Jul 30	output/issue5578-native-speed-tier	handoff-needed	Native speed-tier evidence; confirm promotion target in #5578
1.2G	Jul 12	output/issue4365-13376-harvest	handoff-needed	Harvest bundle; confirm promotion target in #4365

Medium entries (10M–1G): ignored-cache / regenerable unless noted

model_cache 262M (ignored-cache; checkpoints need a per-model triage slice before any prune),
external_data/sdd 425M (durable-required pending license/provenance check — staged dataset),
coverage 188M (ignored-cache, regenerable),
issue3216-* CPU-requeue 594M + harvest 206M + hashrefresh 88M (prune-candidates conditional on #3216),
issue1554-* post-finalizer-run 190M + split-mem180-run 179M + l40s-mem180 177M + mean-matched-harness 178M +
13375-harvest 17M + sipp-compare 9.2M + best-ppo-compare-run 5.9M (campaign scratch; conditional prune
candidates except harvests, which are handoff-needed),
issue3810-* ppo-only-rerun 36M + snqi 32M + extroster-run 31M + longhorizon-run 23M + hybrid-roster 23M
(conditional prune candidates),
issue4830-safety-wrapper-factorial-v1 27M (handoff-needed — active #7684 campaign),
issue4205-codesign-loop-run 3.8M + autogen-trainings 4.4M + tmp_criticality 3.9M (disposable),
issue5326-objective-comparison-20260802 2.7M + issue6095-* 2.6M/12K (conditional prune candidates).

Small entries (<10M, ~90 items): disposable / handoff-needed

• disposable (regenerable scratch, no evidence value): tmp/, quickstart_smoke/,
  issue_3216_local_dedup_dry_run, all *-preflight dirs with a matching -run counterpart
  (~15 dirs, ~68–192K each), payload/, results/, shortlist/, validation/, benchmarks/,
  fidelity_*, licca-*, repos/ (stub README), release_verification/, trace_viewer/,
  threejs_viewer/, compute_feasibility_matrix_v1/, issue_refinement/, issue-drafts/,
  pr_bodies/, per-issue *-acceptance/*-inventory-20260630* stubs (16–52K each).
• handoff-needed (active review/audit context, Sep 2–3): pr-7468/7468/7983/8192/8293/8298-review-*.md,
  pr_4608_followup_body.md, issue-7649-release-race-comment.md,
  issue_audit_plan*.json (rewritten today by active runs — do not touch),
  project_priority_score/, issue7684-safety-wrapper pointer, issue5164-sources 392K
  (evidence-source staging; confirm promotion).
• durable-required: paper_experiment_matrix_7planners_v1_release_v0_0_2_*_publication_bundle.tar.gz
  532K (publication bundle — keep until Zenodo deposit), release/ (release-note drafts),
  issue_4904_* figures/csv (published-figure artifacts; confirm report linkage).
• tracked-manifest: none (no tracked manifests live in ignored output/ by policy).

Prune-candidate list, largest first (conditional — reclaimable bytes)

1. output/slurm/ remainder excl. trustworthy-rerun ≈ 1.0G
2. output/issue2557-seed-variance 6.1G
3. output/issue4206-trace-rerun 2.8G
4. output/issue3068-ppo-curriculum 2.3G
5. output/issue3216-headline-ci-rank-stability-cpu-requeue-20260630h 594M
6. output/issue3216-13274-harvest 206M + ...-hashrefresh-20260630e 88M
7. output/coverage 188M + output/tmp* ~4.2M + preflight-only dirs ~2M

Conditional prune pool ≈ 13.2G (excl. the 56G trustworthy-rerun dir, which stays until
#5346 campaign disposition). No durable-required or handoff-needed entry is listed above.

Follow-ups (out of scope for this slice)

• Per-checkpoint triage slice for output/model_cache (262M).
• License/provenance check for output/external_data/sdd (425M).
• Deletion/promotion execution slice under friction: main checkout output/ holds 83.6 GB of untriaged generated artifacts #8255 after review of this manifest.

[ll7]

Artifact-provenance classification manifest for output/ (closes #8371)

Method (2026-09-03, main checkout, read-only): du -sh output/* | sort -rh
(full 131-entry inventory), stat mtimes, top-level content probes, and
check_worktree_capacity.py category guidance ("ignored generated output;
preserve durable evidence before pruning"). Total 78G (parent reported
83.6G on 2026-09-02; drift is expected — e.g. model_cache touched 2026-09-02,
issue_audit_plan.json rewritten today by active autopilot runs).
Top-8 sum = 75.9G; remainder ≈ 2.1G across 123 small entries. Totals reconcile
with du -sh output/ within rounding. Nothing was deleted or moved.
Confidence on campaign-dir dispositions is ~70%: prune candidates below are
conditional on per-issue evidence confirmation, not deletion orders.

Large entries (all >1G)

size	mtime	entry	classification	rationale / condition
56G	Aug 18	output/slurm/10-issue5346-s30-trustworthy-rerun-job-13529	durable-required (pending)	Trustworthy-rerun evidence for the #5346 S30 campaign; keep until that campaign's evidence is promoted or ruled superseded
~1G	Jun–Aug	output/slurm/ (remaining ~40 job dirs, largest 782M oracle-trace, 293M S30-canary)	prune-candidate (conditional)	Job logs/scratch older than the last audited campaign; prune only after confirming no open campaign cites them
6.1G	Jun 27	output/issue2557-seed-variance	prune-candidate (conditional)	Stale campaign scratch; prune after #2557 evidence check
2.8G	Jul 8	output/issue4206-trace-rerun	prune-candidate (conditional)	Rerun scratch superseded by issue4206-13334-harvest; prune after #4206 evidence check
2.8G	Jul 4	output/issue4206-13334-harvest	handoff-needed	Harvest bundle; confirm promotion target in #4206 before any prune
2.3G	Jul 14	output/issue3068-ppo-curriculum	prune-candidate (conditional)	Curriculum scratch; prune after #3068 evidence check
1.9G	Aug 30	output/release_packet_audit_20260821	durable-required	Recent release-packet audit input; keep until v0.0.6 release disposition
1.8G	Jul 30	output/issue5578-native-speed-tier	handoff-needed	Native speed-tier evidence; confirm promotion target in #5578
1.2G	Jul 12	output/issue4365-13376-harvest	handoff-needed	Harvest bundle; confirm promotion target in #4365

Medium entries (10M–1G): ignored-cache / regenerable unless noted

model_cache 262M (ignored-cache; checkpoints need a per-model triage slice before any prune),
external_data/sdd 425M (durable-required pending license/provenance check — staged dataset),
coverage 188M (ignored-cache, regenerable),
issue3216-* CPU-requeue 594M + harvest 206M + hashrefresh 88M (prune-candidates conditional on #3216),
issue1554-* post-finalizer-run 190M + split-mem180-run 179M + l40s-mem180 177M + mean-matched-harness 178M +
13375-harvest 17M + sipp-compare 9.2M + best-ppo-compare-run 5.9M (campaign scratch; conditional prune
candidates except harvests, which are handoff-needed),
issue3810-* ppo-only-rerun 36M + snqi 32M + extroster-run 31M + longhorizon-run 23M + hybrid-roster 23M
(conditional prune candidates),
issue4830-safety-wrapper-factorial-v1 27M (handoff-needed — active #7684 campaign),
issue4205-codesign-loop-run 3.8M + autogen-trainings 4.4M + tmp_criticality 3.9M (disposable),
issue5326-objective-comparison-20260802 2.7M + issue6095-* 2.6M/12K (conditional prune candidates).

Small entries (<10M, ~90 items): disposable / handoff-needed

• disposable (regenerable scratch, no evidence value): tmp/, quickstart_smoke/,
  issue_3216_local_dedup_dry_run, all *-preflight dirs with a matching -run counterpart
  (~15 dirs, ~68–192K each), payload/, results/, shortlist/, validation/, benchmarks/,
  fidelity_*, licca-*, repos/ (stub README), release_verification/, trace_viewer/,
  threejs_viewer/, compute_feasibility_matrix_v1/, issue_refinement/, issue-drafts/,
  pr_bodies/, per-issue *-acceptance/*-inventory-20260630* stubs (16–52K each).
• handoff-needed (active review/audit context, Sep 2–3): pr-7468/7468/7983/8192/8293/8298-review-*.md,
  pr_4608_followup_body.md, issue-7649-release-race-comment.md,
  issue_audit_plan*.json (rewritten today by active runs — do not touch),
  project_priority_score/, issue7684-safety-wrapper pointer, issue5164-sources 392K
  (evidence-source staging; confirm promotion).
• durable-required: paper_experiment_matrix_7planners_v1_release_v0_0_2_*_publication_bundle.tar.gz
  532K (publication bundle — keep until Zenodo deposit), release/ (release-note drafts),
  issue_4904_* figures/csv (published-figure artifacts; confirm report linkage).
• tracked-manifest: none (no tracked manifests live in ignored output/ by policy).

Prune-candidate list, largest first (conditional — reclaimable bytes)

1. output/slurm/ remainder excl. trustworthy-rerun ≈ 1.0G
2. output/issue2557-seed-variance 6.1G
3. output/issue4206-trace-rerun 2.8G
4. output/issue3068-ppo-curriculum 2.3G
5. output/issue3216-headline-ci-rank-stability-cpu-requeue-20260630h 594M
6. output/issue3216-13274-harvest 206M + ...-hashrefresh-20260630e 88M
7. output/coverage 188M + output/tmp* ~4.2M + preflight-only dirs ~2M

Conditional prune pool ≈ 13.2G (excl. the 56G trustworthy-rerun dir, which stays until
#5346 campaign disposition). No durable-required or handoff-needed entry is listed above.

Follow-ups (out of scope for this slice)

• Per-checkpoint triage slice for output/model_cache (262M).
• License/provenance check for output/external_data/sdd (425M).
• Deletion/promotion execution slice under friction: main checkout output/ holds 83.6 GB of untriaged generated artifacts #8255 after review of this manifest.

[ll7]

Triage manifest: output/model_cache (issue #8426, read-only slice)

Inventory taken 2026-09-04 on the maintaining machine. Total reconciles: 90M + 4.9M + 4.8M + 4.0K dir ≈ 99M (du -sh output/model_cache = 99M; the 262M figure in the issue predates earlier prune passes — current output/ total is 3.1G).

entry	size	mtime (UTC+2)	identity / source	classification	rationale
ppo_expert_issue_791_reward_curriculum_eval_aligned_large_capacity_20260417/ppo_expert_issue_791_reward_curriculum_eval_aligned_large_capacity_20260417-model.zip	90M	2026-08-17	sha256 2b30df81… matches model/registry.yaml release asset; registry local_path points here; referenced by configs/baselines/ppo_issue_791_eval_aligned_* and eval configs	keep	canonical local copy, release-backed, actively referenced
predictive_proxy_selected_v1/predictive_proxy_selected_v1-predictive_model.pt	4.9M	2026-08-17	sha256 0e7e775f… matches registry release asset; referenced by model_id in configs/algos/prediction_planner_*.yaml	keep	verified working copy, release-backed
predictive_proxy_selected_v2_full/predictive_proxy_selected_v2_full-predictive_model.pt	4.8M	2026-06-21	sha256 a28aed6d… matches registry release asset (size 4958329)	keep	verified working copy, release-backed
local_model/	4.0K (empty dir)	2026-08-17	no content, no registry entry, no config reference	prune-candidate	0 reclaimable bytes; remove with a later prune pass

Reclaimable bytes: 0 (nothing with content is a prune candidate).

Promotion-check item for a later slice (not executed here): the registry local_path for predictive_proxy_selected_v1 and predictive_proxy_selected_v2_full still points at long-gone output/tmp/predictive_planner/... paths while the verified copies live in output/model_cache/. Recommend a registry-only follow-up updating those two local_path pointers (21 registry entries already point at output/model_cache, so the convention exists).

Confirmation: check_worktree_capacity.py --inventory --json re-run: capacity pass, output = 3264032768 bytes, status review. Nothing deleted or moved in this slice.