diff --git a/.github/workflows/conleche-nightly.yml b/.github/workflows/conleche-nightly.yml
index 63285936a7..767e167d96 100644
--- a/.github/workflows/conleche-nightly.yml
+++ b/.github/workflows/conleche-nightly.yml
@@ -42,7 +42,6 @@ jobs:
     # 2.6 min) -> T = max(15, 68, 65) = 68, rounded up for a shared host. The first three runs replace this basis.
     timeout-minutes: 90
     env:
-      CI_CACHE: /mnt/nvme-raid0/ci-cache/pvl-conleche
       ELAN_VERSION: v4.2.4
       ELAN_SHA256: 42b94d4244e8353142c456ec0e4ca6528fd898a6c604d4059f494e706e431f63
       LEAN_NUM_THREADS: '8'
@@ -55,6 +54,11 @@ jobs:
       - name: Pinned elan in the job's own directory (sha256-checked, never a host install)
         run: |
           set -euo pipefail
+          # The cache lives beside the runner's _work dir, as nightly.yml's cache-nightly does. The old
+          # /mnt/nvme-raid0/ci-cache path exists only on lambda; on the clean-room hosts /mnt/nvme-raid0 is
+          # root-owned, so mkdir failed and the job never ran (runs 37082537768, 37163229207).
+          CI_CACHE="$(cd "$GITHUB_WORKSPACE/../.." && pwd)/cache-conleche"
+          echo "CI_CACHE=$CI_CACHE" >> "$GITHUB_ENV"
           export ELAN_HOME="$CI_CACHE/elan"
           mkdir -p "$CI_CACHE"
           if [ ! -x "$ELAN_HOME/bin/elan" ] || ! "$ELAN_HOME/bin/elan" --version | grep -q "${ELAN_VERSION#v}"; then
diff --git a/.github/workflows/cuda-nightly.yml b/.github/workflows/cuda-nightly.yml
index 53dafd909d..c33e72b56d 100644
--- a/.github/workflows/cuda-nightly.yml
+++ b/.github/workflows/cuda-nightly.yml
@@ -707,6 +707,11 @@ jobs:
         if: ${{ !cancelled() && steps.decide.outputs.proceed == 'true' }}
         run: |
           set -uo pipefail
+          # -e OFF, as in the perf041 step: the default shell is `bash -e {0}`, so the
+          # RED control's expected exit 101 aborted the step at `; r=$?` before its
+          # assertions ran (run 37161229337: GREEN passed, step failed 101, no ::error::).
+          # Every failure path below is guarded explicitly.
+          set +e
           t=api::openai_handlers::qwen35_serve_tests::gpu_every_serve_route_prefills_through_the_batched_prefill
           model="$HOME/models/Qwen3.5-0.8B-Q4_K_M.gguf"
           [ -f "$model" ] || { echo "::error::#4250 UNMEASURABLE on ${{ matrix.name }} (provisioning): $model is absent"; exit 1; }
diff --git a/.github/workflows/mutants-nightly.yml b/.github/workflows/mutants-nightly.yml
index a5d12a8c98..2c9e251686 100644
--- a/.github/workflows/mutants-nightly.yml
+++ b/.github/workflows/mutants-nightly.yml
@@ -104,8 +104,9 @@ jobs:
                 }
               fi
               cargo mutants --version || { echo "::error::cargo-mutants not runnable after install"; exit 1; }
-              # the crates ci.yml workspace-test --excludes, derived from its nextest line (see ci.yml mutants)
-              EXCL=$(grep -m1 -E "^[[:space:]]+cargo nextest run --profile ci --workspace --lib" .github/workflows/ci.yml \
+              # the crates workspace-test --excludes, derived from its nextest line. That line moved from
+              # ci.yml into ci/sections.yml (#4655); reading the old file found nothing and failed every night.
+              EXCL=$(grep -m1 -E "^[[:space:]]+cargo nextest run --profile ci --workspace --lib" ci/sections.yml \
                 | grep -oE -- "--exclude [a-z0-9_-]+" | sed "s/^--exclude /--exclude-crate /" | tr "\n" " ")
               [ -n "$EXCL" ] || { echo "::error::no --exclude list found on the workspace-test nextest line"; exit 1; }
               # shellcheck disable=SC2086
diff --git a/.github/workflows/nightly.yml b/.github/workflows/nightly.yml
index 1f34b0e156..5cfabf0079 100644
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@@ -80,7 +80,9 @@ jobs:
           ref: main
 
       - name: Case table
-        run: python3 scripts/nightly_manifest.py --self-test
+        run: |
+          python3 scripts/nightly_manifest.py --self-test
+          bash scripts/nightly_prune_release.sh --self-test
 
       - name: Decide (reused | build | red-ci | ci-pending)
         id: gate
@@ -259,6 +261,11 @@ jobs:
       - name: Build apr (host-native, pinned toolchain)
         run: |
           set -euo pipefail
+          # Every later step reads ./target. The mini's runner env redirects cargo's
+          # target dir (CARGO_TARGET_DIR=~/data/cargo-target), so cargo built
+          # elsewhere and the --version line found no apr (run 37171118510, exit
+          # 127). Same fix as binary-release.yml's mini lane (#4660).
+          export CARGO_TARGET_DIR="$GITHUB_WORKSPACE/target"
           rustup show active-toolchain
           cargo build --release --locked -p apr-cli --target "$T"
           "target/$T/release/apr" --version
@@ -321,6 +328,13 @@ jobs:
           cat nightly-manifest.json
           echo "decision=$(python3 -c 'import json;print(json.load(open("nightly-manifest.json"))["decision"])')" >> "$GITHUB_OUTPUT"
 
+      # publish lists one page of assets (100); the release overflowed it, so staged
+      # leftovers and old copies past page one made every upload 422 (run 37171118510).
+      - name: Prune the release to what publish can see
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: bash scripts/nightly_prune_release.sh dist nightly-manifest.json "$NIGHTLY_SHA"
+
       # REST, not `gh`: the fleet boxes carry no gh (binary-release.yml, run 34448908554).
       - name: Publish green arches, then nightly-manifest.json
         env:
diff --git a/ci/sections.yml b/ci/sections.yml
index 410f4f350e..c305724594 100644
--- a/ci/sections.yml
+++ b/ci/sections.yml
@@ -1394,6 +1394,7 @@ jobs:
           bash scripts/ci_resolve_dirty.sh --selftest
           python3 scripts/lib/roadmap_merge.py --selftest
           python3 scripts/nightly_manifest.py --self-test   # #4189 nightly gate/manifest/publish plan
+          bash scripts/nightly_prune_release.sh --self-test   # nightly publish overflowed page one (422)
       - name: "target-watch: before-cargo-install-bashrs"
         # aprender#2822 boundary probe (2 stat, 1 readdir, 1 df, 1 pgrep).
         # Instrumentation, never a gate: the script always exits 0.
diff --git a/docs/roadmaps/entries/GH-4718.yaml b/docs/roadmaps/entries/GH-4718.yaml
new file mode 100644
index 0000000000..715d659fbd
--- /dev/null
+++ b/docs/roadmaps/entries/GH-4718.yaml
@@ -0,0 +1,17 @@
+- id: GH-4718
+  github_issue: 4718
+  item_type: task
+  title: '0702-P1: nightly producers for the deep lanes (deep-doctests, deep-nodefault, deep-bins-smoke)'
+  status: inprogress
+  priority: critical
+  assigned_to: null
+  created: 2026-10-04T09:17:34.904751550+00:00
+  updated: 2026-10-04T09:17:34.904751550+00:00
+  spec: null
+  acceptance_criteria: []
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - P0
+  notes: null
diff --git a/docs/roadmaps/entries/PMAT-4715.yaml b/docs/roadmaps/entries/PMAT-4715.yaml
new file mode 100644
index 0000000000..398c99140b
--- /dev/null
+++ b/docs/roadmaps/entries/PMAT-4715.yaml
@@ -0,0 +1,19 @@
+- id: PMAT-4715
+  github_issue: 4715
+  item_type: task
+  title: 'B2: root-cause fixes for red nightlies on main'
+  status: inprogress
+  priority: critical
+  assigned_to: aprender-wmainred
+  created: 2026-10-04T00:00:00Z
+  updated: 2026-10-04T00:00:00Z
+  acceptance_criteria:
+  - Each red or cancelled scheduled workflow on main has a root cause named in its fix PR, one PR per defect, listed on issue #4715
+  - Each fix PR changes no gate's acceptance; a threshold or acceptance change goes to the train lead's sign-off batch instead
+  - After each fix merges, the next scheduled or dispatched run of that workflow on main gets past the step that failed
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - kind:ci
+  notes: 'Parent #4678 (0.70.2), bar B2: no check red on main for more than 24 h. Proposed by aprender-a7, minted by cop aprender-14.'
diff --git a/docs/roadmaps/roadmap.yaml b/docs/roadmaps/roadmap.yaml
index 211749d65b..8dd540de02 100644
--- a/docs/roadmaps/roadmap.yaml
+++ b/docs/roadmaps/roadmap.yaml
@@ -18483,6 +18483,23 @@ roadmap:
   estimated_effort: null
   labels: []
   notes: 'Before: v0.70.1 tag run completed 25 min after start with no ci / coverage job; the gate refused only at preflight, after the tag and the GitHub release were public (up to 90 tries x 60 s, 2 gh calls/try). After: refused at cut_tag, 0 gh calls, 0 s. NOTE: ci.yml today declares no ci / coverage job (coverage runs inside x86-main), so every release refuses at cut_tag until one is declared or the gate is retargeted.'
+- id: GH-4718
+  github_issue: 4718
+  item_type: task
+  title: '0702-P1: nightly producers for the deep lanes (deep-doctests, deep-nodefault, deep-bins-smoke)'
+  status: inprogress
+  priority: critical
+  assigned_to: null
+  created: 2026-10-04T09:17:34.904751550+00:00
+  updated: 2026-10-04T09:17:34.904751550+00:00
+  spec: null
+  acceptance_criteria: []
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - P0
+  notes: null
 - id: PMAT-3351
   github_issue: 3347
   item_type: task
@@ -24081,6 +24098,25 @@ Not done, pending the ruling: moving the 31 fixture files to an obviously synthe
   labels:
   - kind:docs
   notes: 'Operator C197, 2026-10-01 12:40Z: re-date via the spec''s own procedure; check unchanged; not a waiver.'
+- id: PMAT-4715
+  github_issue: 4715
+  item_type: task
+  title: 'B2: root-cause fixes for red nightlies on main'
+  status: inprogress
+  priority: critical
+  assigned_to: aprender-wmainred
+  created: 2026-10-04T00:00:00Z
+  updated: 2026-10-04T00:00:00Z
+  acceptance_criteria:
+  - Each red or cancelled scheduled workflow on main has a root cause named in its fix PR, one PR per defect, listed on issue #4715
+  - Each fix PR changes no gate's acceptance; a threshold or acceptance change goes to the train lead's sign-off batch instead
+  - After each fix merges, the next scheduled or dispatched run of that workflow on main gets past the step that failed
+  phases: []
+  subtasks: []
+  estimated_effort: null
+  labels:
+  - kind:ci
+  notes: 'Parent #4678 (0.70.2), bar B2: no check red on main for more than 24 h. Proposed by aprender-a7, minted by cop aprender-14.'
 - id: PMAT-4719
   github_issue: 4719
   item_type: task
diff --git a/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/findings.sarif b/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/findings.sarif
new file mode 100644
index 0000000000..772d58e67b
--- /dev/null
+++ b/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/findings.sarif
@@ -0,0 +1,232 @@
+{
+  "$schema": "https://docs.oasis-open.org/sarif/sarif/v2.1.0/errata01/os/schemas/sarif-schema-2.1.0.json",
+  "version": "2.1.0",
+  "runs": [
+    {
+      "tool": {
+        "driver": {
+          "name": "pmat"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": [
+        {
+          "ruleId": "no_existing_paginated_prune",
+          "level": "note",
+          "message": {
+            "text": "pmat query at head for paginated deletion of stale release assets returns 5 results; the nearest is nightly_manifest.py publish, which lists one page only, and none paginates or prunes. The duplication scan searched 12 needles over 1380/1380 sibling branches with 0 hits."
+          },
+          "properties": {
+            "grounding": "measured",
+            "command": [
+              "pmat",
+              "query",
+              "paginate release assets delete stale",
+              "--limit",
+              "5"
+            ],
+            "exit_code": 0,
+            "stdout_sha256": "8b607c9c54be06b5796a61cbc2edc00cd53724230579ef1c29344a9e30dae4d5",
+            "rationale": "No helper exists to reuse; the script fills the gap publish leaves past page one.",
+            "failure_scenario": "If a paginated prune existed elsewhere, a second one would drift from it; none was found.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "plan_mirrors_publish_plan",
+          "level": "note",
+          "message": {
+            "text": "The plan mirror selects green targets at SHA built by this run, as publish_plan does."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 66",
+            "excerpt": "      | select(.status == \"green\" and .green_sha == $sha and .built_run_id == $r)",
+            "excerpt_sha256": "eb797e09f2aa3f3f39583a574fb88f67d7bd1881b67daac7e518139c8360115b",
+            "rationale": "The jq select applies the three filters publish_plan applies at scripts/nightly_manifest.py:320, so only names publish re-uploads lose their old copy.",
+            "failure_scenario": "If a filter were dropped, an asset outside the plan (red, stale sha, reused run) would be deleted and not replaced; mutants m4/m5/m6 each turn 2 self-test rows red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "publish_plan_filter_source",
+          "level": "note",
+          "message": {
+            "text": "publish_plan's own selection, cited as the reference the mirror is checked against."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_manifest.py at head, line 320",
+            "excerpt": "        if e[\"status\"] != \"green\" or e[\"green_sha\"] != sha or e.get(\"built_run_id\") != manifest[\"run_id\"]:",
+            "excerpt_sha256": "a17d4cc48e1b62552ed9d0a0de37f08b40e430f45c6463e5396254b26cbc8ab2",
+            "rationale": "This is the line the script mirrors; the review compared them field by field.",
+            "failure_scenario": "A divergence here would let the prune delete what publish skips; both lanes found no divergence on a well-formed manifest.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "non_array_page_fails_closed",
+          "level": "note",
+          "message": {
+            "text": "A 200 asset page that is not a JSON array exits nonzero."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 88",
+            "excerpt": "    jq -e 'type == \"array\"' <<< \"$batch\" >/dev/null 2>&1 || die \"assets page $page is not a JSON array\"",
+            "excerpt_sha256": "7608f5eb023d469aabda327d4d11a5c06b58e87c74bd56e7c2efdb83eb1bbd59",
+            "rationale": "Round 1 lane 1 (claude-sonnet-5-5) found a non-array 200 page read as the empty last page; this guard closes it and the self-test row answers {} so only the guard can catch it.",
+            "failure_scenario": "Without the guard a {} page ends the listing early and past-page-one copies survive to a rename 422; mutant m1 turns 1 row red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "exact_membership",
+          "level": "note",
+          "message": {
+            "text": "Membership in the plan is an exact whole-line match."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 110",
+            "excerpt": "    if [[ \"$up\" == *$'\\n'\"$name\"$'\\n'* ]]; then",
+            "excerpt_sha256": "4530bc84d68b9c99990c3a3b3de473a7cf6c6577f438f1aaa4ce7b95c192e08d",
+            "rationale": "Membership is newline-delimited so a planned name that contains another as a substring does not match it.",
+            "failure_scenario": "A substring match deletes a234 because a234.old is planned; mutant m10 turns 5 rows red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "jq_rewrite_equivalent",
+          "level": "note",
+          "message": {
+            "text": "58b6456d07 rewrites the self-test fixture's $all[] as $all | .[]; the same jq program, measured equal on a sample (jq -n: [1,2,3] as $all | [$all[]|select(.>1)] == [$all|.[]|select(.>1)] -> true)."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 153",
+            "excerpt": "              tools: ([$all | .[] | select(. as $n | $skip | index($n) | not)] | map({key: ., value: {asset: .}}) | from_entries)}}",
+            "excerpt_sha256": "c21fd10a4f40fec7eb572c5dc90e0b1f6fed8dcbb6a782564e4f026bbdc3af18",
+            "rationale": "A syntax-only change in the self-test fixture so CI's bashrs 7.4.1 stops reporting SC1087; the production path is unchanged and the self-test still passes with 21 rows.",
+            "failure_scenario": "If the rewrite changed the fixture's target set, the self-test rows built from it would test a different manifest than intended.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "planted_question_rejected",
+          "level": "note",
+          "message": {
+            "text": "Quorum round 4: unanimous PASS, degraded: same-family; the planted question was rejected by both lanes."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "git show --stat --format= 58b6456d07 (first line)",
+            "excerpt": " scripts/nightly_prune_release.sh | 2 +-",
+            "excerpt_sha256": "c001cb6a0e49c11a718987ec225f31ac8bc6356f4769ffbe68a0c2262de5cc8d",
+            "rationale": "The planted contrary question Q3 claimed 58b6456d07 also changed the production delete loop near line 110; both round-4 lanes (claude-sonnet-5-5, claude-haiku-4-5-20251001) rejected it: the commit is one line in, one out, at self-test line 153.",
+            "failure_scenario": "If 58b6456d07 had touched the delete loop, the plan-membership guarantee reviewed in round 3 would no longer be the code at head.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "quorum_r4_lane1_sonnet",
+          "level": "note",
+          "message": {
+            "text": "Round 4 lane 1, model claude-sonnet-5-5: PASS, planted Q3 rejected, mutation of its own choosing RED."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "/mnt/nvme-raid0/cop-inbox/handoff/signoff-q-4777-prune.md, quorum receipt table, round 4 row",
+            "excerpt": "| 4 | 4cc22e57b4 | 1 | claude-sonnet-5-5 | PASS | rejected (58b6456d07 = 1+/1-, line 153 only; loop :102-112 untouched) | yes | :110 membership test -> `if true`, RED 10 |",
+            "excerpt_sha256": "6d3961b47255553f2c091182a8661d4fe580288d61cb1dd8c80af266b285b332",
+            "rationale": "Lane 1 verdict recorded on its own, with exact model id.",
+            "failure_scenario": "Under C312-9 a same-family answer stands only if unanimous; a FAIL or a lane that accepted the planted premise in either row would void the review.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "quorum_r4_lane2_haiku",
+          "level": "note",
+          "message": {
+            "text": "Round 4 lane 2, model claude-haiku-4-5-20251001: PASS, planted Q3 rejected, mutation of its own choosing RED."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "/mnt/nvme-raid0/cop-inbox/handoff/signoff-q-4777-prune.md, quorum receipt table, round 4 row",
+            "excerpt": "| 4 | 4cc22e57b4 | 2 | claude-haiku-4-5-20251001 | PASS | rejected (only line 153 changed; zero delete-loop lines) | yes | :110 leading newline dropped, RED 7 |",
+            "excerpt_sha256": "74709a1bba35c570fead2330a17707a1e6e35faa028bdf3ebd6e7b08eb323ba4",
+            "rationale": "Lane 2 verdict recorded on its own, with exact model id; its first reply was verdict-only and was asked once for cited answers. With lane 1 it makes the vote unanimous (2/2 PASS).",
+            "failure_scenario": "Under C312-9 a same-family answer stands only if unanimous; a FAIL or a lane that accepted the planted premise in either row would void the review.",
+            "precision_class": "advisory"
+          }
+        }
+      ]
+    },
+    {
+      "tool": {
+        "driver": {
+          "name": "cargo-mutants"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": [
+        {
+          "ruleId": "guard_mutants_killed",
+          "level": "note",
+          "message": {
+            "text": "Self-test at head: 21 ok rows, PASS."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "author mutant run at head 4cc22e57b4 (scripts/nightly_prune_release.sh --self-test)",
+            "excerpt": "head: 21 ok, nightly_prune_release self-test: PASS",
+            "excerpt_sha256": "33217b4d9b0db767ffea077dff8b6bfd2816b902b75428fd50e3dfdbb4437df6",
+            "rationale": "The unmutated script passes all 21 self-test rows.",
+            "failure_scenario": "If the self-test were vacuous, mutants would pass; see the next result.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "guard_mutants_all_red",
+          "level": "note",
+          "message": {
+            "text": "10 of 10 guard mutants killed."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "author mutant run at head 4cc22e57b4, 10 mutants",
+            "excerpt": "m1-no-array-guard: nightly_prune_release self-test: FAIL (1)\nm2-dist-only: nightly_prune_release self-test: FAIL (6)\nm3-plan-read-open: nightly_prune_release self-test: FAIL (1)\nm4-no-status: nightly_prune_release self-test: FAIL (2)\nm5-no-sha: nightly_prune_release self-test: FAIL (2)\nm6-no-run: nightly_prune_release self-test: FAIL (2)\nm7-page-one-only: nightly_prune_release self-test: FAIL (10)\nm8-no-staged-purge: nightly_prune_release self-test: FAIL (2)\nm9-no-manifest-name: nightly_prune_release self-test: FAIL (1)\nm10-substring-match: nightly_prune_release self-test: FAIL (5)",
+            "excerpt_sha256": "5dd9f7cabbb2ffc1c35f72e6d5cf22a63f76c5771ffebd2e3abc7d0e65695194",
+            "rationale": "Each of 10 mutants (array guard, dist-only membership, plan read open, status/sha/run filters, page-one-only, staged purge, manifest name, substring match) turns the self-test red; each mutant copy parses.",
+            "failure_scenario": "A surviving mutant would mark a row that cannot fail; none survived.",
+            "precision_class": "advisory"
+          }
+        }
+      ]
+    },
+    {
+      "tool": {
+        "driver": {
+          "name": "antigravity"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": []
+    }
+  ]
+}
diff --git a/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/receipt.intoto.jsonl b/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/receipt.intoto.jsonl
new file mode 100644
index 0000000000..fe27da242c
--- /dev/null
+++ b/evidence/pr-review/4777/4cc22e57b485bf1cd9aae8a2122cbafeb2205929/receipt.intoto.jsonl
@@ -0,0 +1 @@
+{"_type":"https://in-toto.io/Statement/v1","subject":[{"name":"git+https://github.com/paiml/aprender","digest":{"sha1":"4cc22e57b485bf1cd9aae8a2122cbafeb2205929"}}],"predicateType":"https://paiml.dev/attestations/pr-review/v2","predicate":{"skill_version":"2.1.0","attestation_level":"L1-self","pr":4777,"base_sha":"6192aa9b886a31341a9cc7465662a0598964ce8c","head_sha":"4cc22e57b485bf1cd9aae8a2122cbafeb2205929","author_actor":{"kind":"agent","id":"agent:claude-opus-5-5/aprender-wtix-odd"},"reviewer_actor":{"kind":"agent","id":"agent:claude-sonnet-5-5/session-independent-review-4cc22e57b4"},"affected_crates":[],"verdict":"DEGRADED","degraded_reason":"degraded: same-family — measured out: gemini (empty); Claude Code lanes: claude-sonnet-5-5,claude-haiku-4-5-20251001","consultations":{"pmat":{"status":"consulted","transport":"cli","transport_unavailable":["mcp: not attempted this run (cli answered)"],"index_commit":"4cc22e57b485bf1cd9aae8a2122cbafeb2205929","index_is_ancestor":true,"index_worktree_dirty":false,"complexity_delta":[],"tdg_delta":[],"satd_introduced":[],"cache_hits":0,"duplication_hits":[],"duplication_coverage":{"rust":"lexical","shell":"lexical","python":"lexical","config":"lexical","docs":"lexical","other":"lexical","sibling_branches":"lexical","merge_base_to_main":"lexical"},"duplication_horizon":["head=4cc22e57b485bf1cd9aae8a2122cbafeb2205929","siblings=refs/remotes/origin/* unmerged into origin/main","merge_base_to_main=6192aa9b886a31341a9cc7465662a0598964ce8c..refs/remotes/origin/main"],"horizon_branches_total":1380,"horizon_branches_scanned":1380,"merge_base_to_main_files":0,"symbols_searched":12,"queries":[{"q":"paginate release assets delete stale","command":["pmat","query","paginate release assets delete stale","--limit","5"],"cwd":"the review worktree at head","results":5,"note":"index built at head; the nearest result is nightly_manifest.py publish, which lists one page; no paginated prune exists to reuse"}]},"cuda":{"status":"consulted","trigger_reason":"The S3.B path trigger fires on .github/workflows/cuda-nightly.yml. Its one hunk turns bash -e off before an existing serve-route GPU test step so the RED control exit 101 reaches its assertions; it changes no kernel, launch, stream or device code. The only claim adjacent to device behaviour is that exit 101 comes from the Rust test harness, not from the CUDA runtime, so that was asked.","queries":[{"q":"process exit code when a CUDA host program test harness panics: does the CUDA runtime define or alter the host process exit status (e.g. 101) on context teardown?","result":"no-authority-found","source_uri":"mcp__nvidia-cuda-docs__search_cuda_docs","claim_checked":"Exit 101 in the step is the Rust test harness panic code. The corpus returned only sample programs that return 0 or 1 themselves, nothing saying the CUDA runtime sets or alters a host exit status, so the claim stands on the Rust side and no finding is raised."}]},"crux":{"status":"not-triggered","trigger_reason":"the diff touches no CLI subcommand or flag, HTTP route, MCP tool, config key or output format; it adds a script run by one nightly workflow step"},"mutation":{"status":"consulted","scope":"guard","attempted":10,"killed":10,"survivors":[]},"antigravity":{"status":"unreachable","attempted":0,"reason":"gemini bucket state empty (budget 0); the budget probe was run and no agy call was made"}},"findings_ref":{"path":"findings.sarif","sha256":"55eabb47fd58047b1e9f63954fded43cdbd1b57a80254a2471390cd18698132f"},"cost":{"input_tokens":0,"output_tokens":0,"wall_seconds":0},"diff_patch_id":"02d78de2f774ff5f864c17e4386daf75614a72a2","diff_patch_id_algo":"git-patch-id-verbatim/pinned-diff-v1"}}
diff --git a/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/findings.sarif b/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/findings.sarif
new file mode 100644
index 0000000000..8651770fc3
--- /dev/null
+++ b/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/findings.sarif
@@ -0,0 +1,216 @@
+{
+  "$schema": "https://docs.oasis-open.org/sarif/sarif/v2.1.0/errata01/os/schemas/sarif-schema-2.1.0.json",
+  "version": "2.1.0",
+  "runs": [
+    {
+      "tool": {
+        "driver": {
+          "name": "pmat"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": [
+        {
+          "ruleId": "no_existing_paginated_prune",
+          "level": "note",
+          "message": {
+            "text": "pmat query at head for paginated deletion of stale release assets returns 5 results; the nearest is nightly_manifest.py publish, which lists one page only, and none paginates or prunes. The duplication scan searched 8 needles over 1376/1376 sibling branches with 0 hits."
+          },
+          "properties": {
+            "grounding": "measured",
+            "command": [
+              "pmat",
+              "query",
+              "paginate release assets delete stale",
+              "--limit",
+              "5"
+            ],
+            "exit_code": 0,
+            "stdout_sha256": "8b607c9c54be06b5796a61cbc2edc00cd53724230579ef1c29344a9e30dae4d5",
+            "rationale": "No helper exists to reuse; the script fills the gap publish leaves past page one.",
+            "failure_scenario": "If a paginated prune existed elsewhere, a second one would drift from it; none was found.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "plan_mirrors_publish_plan",
+          "level": "note",
+          "message": {
+            "text": "The plan mirror selects green targets at SHA built by this run, as publish_plan does."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 66",
+            "excerpt": "      | select(.status == \"green\" and .green_sha == $sha and .built_run_id == $r)",
+            "excerpt_sha256": "eb797e09f2aa3f3f39583a574fb88f67d7bd1881b67daac7e518139c8360115b",
+            "rationale": "The jq select applies the three filters publish_plan applies at scripts/nightly_manifest.py:320, so only names publish re-uploads lose their old copy.",
+            "failure_scenario": "If a filter were dropped, an asset outside the plan (red, stale sha, reused run) would be deleted and not replaced; mutants m4/m5/m6 each turn 2 self-test rows red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "publish_plan_filter_source",
+          "level": "note",
+          "message": {
+            "text": "publish_plan's own selection, cited as the reference the mirror is checked against."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_manifest.py at head, line 320",
+            "excerpt": "        if e[\"status\"] != \"green\" or e[\"green_sha\"] != sha or e.get(\"built_run_id\") != manifest[\"run_id\"]:",
+            "excerpt_sha256": "a17d4cc48e1b62552ed9d0a0de37f08b40e430f45c6463e5396254b26cbc8ab2",
+            "rationale": "This is the line the script mirrors; the review compared them field by field.",
+            "failure_scenario": "A divergence here would let the prune delete what publish skips; both lanes found no divergence on a well-formed manifest.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "non_array_page_fails_closed",
+          "level": "note",
+          "message": {
+            "text": "A 200 asset page that is not a JSON array exits nonzero."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 88",
+            "excerpt": "    jq -e 'type == \"array\"' <<< \"$batch\" >/dev/null 2>&1 || die \"assets page $page is not a JSON array\"",
+            "excerpt_sha256": "7608f5eb023d469aabda327d4d11a5c06b58e87c74bd56e7c2efdb83eb1bbd59",
+            "rationale": "Round 1 lane 1 (claude-sonnet-5-5) found a non-array 200 page read as the empty last page; this guard closes it and the self-test row answers {} so only the guard can catch it.",
+            "failure_scenario": "Without the guard a {} page ends the listing early and past-page-one copies survive to a rename 422; mutant m1 turns 1 row red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "exact_membership",
+          "level": "note",
+          "message": {
+            "text": "Membership in the plan is an exact whole-line match."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 110",
+            "excerpt": "    if [[ \"$up\" == *$'\\n'\"$name\"$'\\n'* ]]; then",
+            "excerpt_sha256": "4530bc84d68b9c99990c3a3b3de473a7cf6c6577f438f1aaa4ce7b95c192e08d",
+            "rationale": "Membership is newline-delimited so a planned name that contains another as a substring does not match it.",
+            "failure_scenario": "A substring match deletes a234 because a234.old is planned; mutant m10 turns 5 rows red.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "planted_question_rejected",
+          "level": "note",
+          "message": {
+            "text": "Quorum round 3: unanimous PASS, degraded: same-family; planted question rejected by both lanes."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "scripts/nightly_prune_release.sh at head, line 48",
+            "excerpt": "  code=$(curl -sS --retry 3 --retry-all-errors -o \"$body\" -w '%{http_code}' -X \"$1\" \\",
+            "excerpt_sha256": "69ab5e8cfedabc2709dad8169f0c84f2bc32ecaa16657c59e495cf062402f579",
+            "rationale": "The planted contrary question Q3 claimed the script calls gh; both round-3 lanes (claude-sonnet-5-5, claude-haiku-4-5-20251001) rejected it and cited the curl call.",
+            "failure_scenario": "A lane that accepted the false premise would void its vote; neither did.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "quorum_r3_lane1_sonnet",
+          "level": "note",
+          "message": {
+            "text": "Round 3 lane 1, model claude-sonnet-5-5: PASS, planted Q3 rejected (cited curl at :48), mutation of its own choosing RED."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "/mnt/nvme-raid0/cop-inbox/handoff/signoff-q-4777-prune.md, quorum receipt table, round 3 row",
+            "excerpt": "| 3 | b5bf02e457 | 1 | claude-sonnet-5-5 | PASS | rejected (curl, :48) | yes | built_run_id filter dropped (:66), RED 2 |",
+            "excerpt_sha256": "b2e676a4e164b6ebc7d07b798af2a424a73b6d23db38f4bd53244fa94af49339",
+            "rationale": "Lane 1 verdict recorded on its own, with exact model id.",
+            "failure_scenario": "Under C312-9 a same-family answer stands only if unanimous; a FAIL or a lane that accepted the planted premise in either row would void the review.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "quorum_r3_lane2_haiku",
+          "level": "note",
+          "message": {
+            "text": "Round 3 lane 2, model claude-haiku-4-5-20251001: PASS, planted Q3 rejected (cited curl at :41-56), mutation of its own choosing RED."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "/mnt/nvme-raid0/cop-inbox/handoff/signoff-q-4777-prune.md, quorum receipt table, round 3 row",
+            "excerpt": "| 3 | b5bf02e457 | 2 | claude-haiku-4-5-20251001 | PASS | rejected (curl, :41-56) | yes | membership line :110 leading \\n dropped, RED 7 |",
+            "excerpt_sha256": "7df753263e124afb3eb8cc01da485301ba76750a5389bf419eebedc8cbe5a219",
+            "rationale": "Lane 2 verdict recorded on its own, with exact model id; with lane 1 it makes the vote unanimous (2/2 PASS).",
+            "failure_scenario": "Under C312-9 a same-family answer stands only if unanimous; a FAIL or a lane that accepted the planted premise in either row would void the review.",
+            "precision_class": "advisory"
+          }
+        }
+      ]
+    },
+    {
+      "tool": {
+        "driver": {
+          "name": "cargo-mutants"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": [
+        {
+          "ruleId": "guard_mutants_killed",
+          "level": "note",
+          "message": {
+            "text": "Self-test at head: 21 ok rows, PASS."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "author mutant run at head b5bf02e457 (scripts/nightly_prune_release.sh --self-test)",
+            "excerpt": "head: 21 ok, nightly_prune_release self-test: PASS",
+            "excerpt_sha256": "33217b4d9b0db767ffea077dff8b6bfd2816b902b75428fd50e3dfdbb4437df6",
+            "rationale": "The unmutated script passes all 21 self-test rows.",
+            "failure_scenario": "If the self-test were vacuous, mutants would pass; see the next result.",
+            "precision_class": "advisory"
+          }
+        },
+        {
+          "ruleId": "guard_mutants_all_red",
+          "level": "note",
+          "message": {
+            "text": "10 of 10 guard mutants killed."
+          },
+          "properties": {
+            "grounding": "cited",
+            "source": "author mutant run at head b5bf02e457, 10 mutants",
+            "excerpt": "m1-no-array-guard: nightly_prune_release self-test: FAIL (1)\nm2-dist-only: nightly_prune_release self-test: FAIL (6)\nm3-plan-read-open: nightly_prune_release self-test: FAIL (1)\nm4-no-status: nightly_prune_release self-test: FAIL (2)\nm5-no-sha: nightly_prune_release self-test: FAIL (2)\nm6-no-run: nightly_prune_release self-test: FAIL (2)\nm7-page-one-only: nightly_prune_release self-test: FAIL (10)\nm8-no-staged-purge: nightly_prune_release self-test: FAIL (2)\nm9-no-manifest-name: nightly_prune_release self-test: FAIL (1)\nm10-substring-match: nightly_prune_release self-test: FAIL (5)",
+            "excerpt_sha256": "5dd9f7cabbb2ffc1c35f72e6d5cf22a63f76c5771ffebd2e3abc7d0e65695194",
+            "rationale": "Each of 10 mutants (array guard, dist-only membership, plan read open, status/sha/run filters, page-one-only, staged purge, manifest name, substring match) turns the self-test red; each mutant copy parses.",
+            "failure_scenario": "A surviving mutant would mark a row that cannot fail; none survived.",
+            "precision_class": "advisory"
+          }
+        }
+      ]
+    },
+    {
+      "tool": {
+        "driver": {
+          "name": "antigravity"
+        }
+      },
+      "invocations": [
+        {
+          "executionSuccessful": true,
+          "toolExecutionNotifications": []
+        }
+      ],
+      "results": []
+    }
+  ]
+}
diff --git a/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/receipt.intoto.jsonl b/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/receipt.intoto.jsonl
new file mode 100644
index 0000000000..0b287538e4
--- /dev/null
+++ b/evidence/pr-review/4777/b5bf02e4577c2d56834eabf78e0ad48143c31698/receipt.intoto.jsonl
@@ -0,0 +1 @@
+{"_type":"https://in-toto.io/Statement/v1","subject":[{"name":"git+https://github.com/paiml/aprender","digest":{"sha1":"b5bf02e4577c2d56834eabf78e0ad48143c31698"}}],"predicateType":"https://paiml.dev/attestations/pr-review/v2","predicate":{"skill_version":"2.1.0","attestation_level":"L1-self","pr":4777,"base_sha":"897bf63f134e011953ad8623366f231ecf9ec9a7","head_sha":"b5bf02e4577c2d56834eabf78e0ad48143c31698","author_actor":{"kind":"agent","id":"agent:claude-opus-5-5/aprender-wtix-odd"},"reviewer_actor":{"kind":"agent","id":"agent:claude-sonnet-5-5/session-independent-review-b5bf02e457"},"affected_crates":[],"verdict":"DEGRADED","degraded_reason":"degraded: same-family — measured out: gemini (empty); Claude Code lanes: claude-sonnet-5-5,claude-haiku-4-5-20251001","consultations":{"pmat":{"status":"consulted","transport":"cli","transport_unavailable":["mcp: not attempted this run (cli answered)"],"index_commit":"b5bf02e4577c2d56834eabf78e0ad48143c31698","index_is_ancestor":true,"index_worktree_dirty":false,"complexity_delta":[],"tdg_delta":[],"satd_introduced":[],"cache_hits":0,"duplication_hits":[],"duplication_coverage":{"rust":"lexical","shell":"lexical","python":"lexical","config":"lexical","docs":"lexical","other":"lexical","sibling_branches":"lexical","merge_base_to_main":"lexical"},"duplication_horizon":["head=b5bf02e4577c2d56834eabf78e0ad48143c31698","siblings=refs/remotes/origin/* unmerged into origin/main","merge_base_to_main=897bf63f134e011953ad8623366f231ecf9ec9a7..refs/remotes/origin/main"],"horizon_branches_total":1376,"horizon_branches_scanned":1376,"merge_base_to_main_files":0,"symbols_searched":8,"queries":[{"q":"paginate release assets delete stale","command":["pmat","query","paginate release assets delete stale","--limit","5"],"cwd":"the review worktree at head","results":5,"note":"index built at head; the nearest result is nightly_manifest.py publish, which lists one page; no paginated prune exists to reuse"}]},"cuda":{"status":"consulted","trigger_reason":"The S3.B path trigger fires on .github/workflows/cuda-nightly.yml. Its one hunk turns bash -e off before an existing serve-route GPU test step so the RED control exit 101 reaches its assertions; it changes no kernel, launch, stream or device code. The only claim adjacent to device behaviour is that exit 101 comes from the Rust test harness, not from the CUDA runtime, so that was asked.","queries":[{"q":"process exit code when a CUDA host program test harness panics: does the CUDA runtime define or alter the host process exit status (e.g. 101) on context teardown?","result":"no-authority-found","source_uri":"mcp__nvidia-cuda-docs__search_cuda_docs","claim_checked":"Exit 101 in the step is the Rust test harness panic code. The corpus returned only sample programs that return 0 or 1 themselves, nothing saying the CUDA runtime sets or alters a host exit status, so the claim stands on the Rust side and no finding is raised."}]},"crux":{"status":"not-triggered","trigger_reason":"the diff touches no CLI subcommand or flag, HTTP route, MCP tool, config key or output format; it adds a script run by one nightly workflow step"},"mutation":{"status":"consulted","scope":"guard","attempted":10,"killed":10,"survivors":[]},"antigravity":{"status":"unreachable","attempted":0,"reason":"gemini bucket state empty (budget 0); the budget probe was run and no agy call was made"}},"findings_ref":{"path":"findings.sarif","sha256":"e4519249143b1403a193f9b8165f6312dc619d9e19a15c6f0db58fcb6b61ea1a"},"cost":{"input_tokens":0,"output_tokens":0,"wall_seconds":0},"diff_patch_id":"cb0438a52c043e676feb5ee0948b582e9a3f1b53","diff_patch_id_algo":"git-patch-id-verbatim/pinned-diff-v1"}}
diff --git a/scripts/nightly_prune_release.sh b/scripts/nightly_prune_release.sh
new file mode 100755
index 0000000000..1ac491b6d9
--- /dev/null
+++ b/scripts/nightly_prune_release.sh
@@ -0,0 +1,237 @@
+#!/usr/bin/env bash
+# nightly_prune_release.sh DIST -- make the nightly release fit what publish sees.
+#
+# scripts/nightly_manifest.py publish lists the release's assets ONE page deep
+# (per_page=100). The release held 234: 117 staged.* leftovers from runs that
+# died mid-publish, and 117 real assets. The leftovers past page one were never
+# deleted, so staging the same name again failed HTTP 422 already_exists (run
+# 37171118510), every night. Purging the leftovers alone is not enough: 117 real
+# assets still overflow page one, publish never sees the old copy of the names
+# past it, and renaming its staged upload onto such a name is the same 422.
+#
+# So, reading EVERY page, before publish runs:
+#   1. delete every staged.* asset (a leftover; publish stages afresh), and
+#   2. delete the old copy of each name this run uploads that sits past page one
+#      of what publish will list. "Uploads" is publish's own plan (publish_plan in
+#      scripts/nightly_manifest.py), read from MANIFEST in jq: the tools of targets
+#      green at SHA and built by this run, each asset and its .sha256, present in
+#      DIST -- plus nightly-manifest.json. A file merely in DIST is not enough.
+# A name past page one that this run does not upload is kept: it is some arch's
+# last green build, and publish never touches it, so it cannot collide.
+# Cost of 2: if publish then fails while staging, those (at most overflow-many)
+# names are absent until the next green night -- the same window publish's own
+# delete-then-rename opens for every name.
+#
+# Any API or tool failure exits nonzero: a listing that could not be read is not
+# an empty release. HTTP 404 on the release itself is the first run: nothing to do.
+#
+#   bash scripts/nightly_prune_release.sh DIST MANIFEST SHA   (needs GH_TOKEN, GITHUB_REPOSITORY)
+#   bash scripts/nightly_prune_release.sh --self-test
+#
+# Test seam: NIGHTLY_PRUNE_API=<cmd> is called as `<cmd> METHOD URL` in place of
+# curl; it prints the body and exits 0, or exits 4 for a 404, else nonzero.
+set -euo pipefail
+
+PAGE=100
+STAGED=staged.
+MANIFEST_ASSET=nightly-manifest.json
+
+die() { echo "::error::nightly_prune_release: $*" >&2; exit 1; }
+
+api() { # METHOD URL -> body on stdout; rc 4 = HTTP 404
+  if [ -n "${NIGHTLY_PRUNE_API:-}" ]; then
+    "$NIGHTLY_PRUNE_API" "$1" "$2"
+    return
+  fi
+  local body code
+  body=$(mktemp)
+  code=$(curl -sS --retry 3 --retry-all-errors -o "$body" -w '%{http_code}' -X "$1" \
+    -H "Authorization: Bearer ${GH_TOKEN:?GH_TOKEN unset}" \
+    -H "Accept: application/vnd.github+json" "$2") || { rm -f "${body:?}"; return 1; }
+  case "$code" in
+    2??) cat "$body"; rm -f "${body:?}" ;;
+    404) rm -f "${body:?}"; return 4 ;;
+    *) echo "::error::$1 $2: HTTP $code" >&2; rm -f "${body:?}"; return 1 ;;
+  esac
+}
+
+prune() { # DIST MANIFEST SHA
+  local dist="$1" manifest="$2" sha="$3" A rel rc page batch n all name id i plan
+  local up=$'\n'  # newline-delimited set: the names publish uploads this run
+  [ -d "$dist" ] || die "dist dir $dist does not exist"
+  command -v jq >/dev/null || die "jq not found on this runner"
+  [ -n "$sha" ] || die "no sha given"
+  # publish_plan, mirrored: green at SHA, built by this run, each asset and its .sha256
+  plan=$(jq -r --arg sha "$sha" '.run_id as $r | .targets[]
+      | select(.status == "green" and .green_sha == $sha and .built_run_id == $r)
+      | .tools[].asset | ., . + ".sha256"' "$manifest") || die "reading the plan from $manifest failed"
+  while IFS= read -r name; do
+    [ -z "$name" ] || [ ! -f "$dist/$name" ] || up+="$name"$'\n'
+  done <<< "$plan"
+  up+="$MANIFEST_ASSET"$'\n'
+  A="https://api.github.com/repos/${GITHUB_REPOSITORY:?GITHUB_REPOSITORY unset}"
+  rc=0
+  rel=$(api GET "$A/releases/tags/nightly") || rc=$?
+  if [ "$rc" -eq 4 ]; then
+    echo "no nightly release yet (HTTP 404): nothing to prune"
+    return 0
+  fi
+  [ "$rc" -eq 0 ] || die "reading the nightly release failed"
+  rel=$(printf '%s' "$rel" | jq -er .id) || die "the nightly release has no id"
+
+  all=""  # "id<TAB>name" per asset, in the order the API lists them
+  page=1
+  while :; do
+    batch=$(api GET "$A/releases/$rel/assets?per_page=$PAGE&page=$page") || die "listing assets page $page failed"
+    # a 200 whose body is not an array (a proxy's HTML, an error object) is a listing
+    # that could not be read, never the empty page that ends it
+    jq -e 'type == "array"' <<< "$batch" >/dev/null 2>&1 || die "assets page $page is not a JSON array"
+    batch=$(printf '%s' "$batch" | jq -r '.[] | "\(.id)\t\(.name)"') || die "reading assets page $page failed"
+    n=0
+    [ -z "$batch" ] || n=$(printf '%s\n' "$batch" | wc -l)
+    [ -z "$batch" ] || all+="$batch"$'\n'
+    [ "$n" -ge "$PAGE" ] || break
+    page=$((page + 1))
+  done
+
+  local staged=0 dropped=0 total=0
+  i=0  # position among the non-staged assets: what publish's one page will hold
+  while IFS=$'\t' read -r id name; do
+    [ -n "$id" ] || continue
+    total=$((total + 1))
+    case "$name" in
+      "$STAGED"*)
+        api DELETE "$A/releases/assets/$id" >/dev/null || die "deleting $name ($id) failed"
+        staged=$((staged + 1))
+        continue ;;
+    esac
+    i=$((i + 1))
+    [ "$i" -gt "$PAGE" ] || continue
+    if [[ "$up" == *$'\n'"$name"$'\n'* ]]; then
+      api DELETE "$A/releases/assets/$id" >/dev/null || die "deleting $name ($id) failed"
+      dropped=$((dropped + 1))
+      echo "past page one, re-uploaded this run: deleted old $name"
+    fi
+  done <<< "$all"
+  echo "nightly release $rel: $total asset(s) over $page page(s); deleted $staged staged leftover(s) and $dropped old copy(ies) past page one"
+}
+
+# ---------------------------------------------------------------- self-test
+
+self_test() {
+  local tmp fails=0 out rc
+  tmp=$(mktemp -d)
+  # stub API: assets come from $tmp/assets (id<TAB>name), deletes are logged
+  cat > "$tmp/api" <<'STUB'
+#!/usr/bin/env bash
+set -euo pipefail
+d="$(dirname "$0")"
+[ ! -e "$d/fail" ] || exit 1
+case "$1 $2" in
+  "GET "*/releases/tags/nightly) [ ! -e "$d/norel" ] || exit 4; echo '{"id": 7}' ;;
+  "GET "*/assets\?per_page=*)
+    p=${2##*page=}
+    [ ! -e "$d/badpage$p" ] || { echo '{}'; exit 0; }
+    s=$(( (p - 1) * 100 + 1 )); e=$(( p * 100 ))
+    sed -n "${s},${e}p" "$d/assets" | jq -Rn '[inputs | split("\t") | {id: (.[0]|tonumber), name: .[1]}]' ;;
+  "DELETE "*) echo "${2##*/}" >> "$d/deleted" ;;
+  *) exit 1 ;;
+esac
+STUB
+  chmod +x "$tmp/api"
+  check() { # NAME GOT WANT
+    if [ "$2" = "$3" ]; then echo "ok   $1"; else echo "FAIL $1: got '$2', want '$3'"; fails=$((fails + 1)); fi
+  }
+  mkplan() { # manifest: every file in dist is a tool of target g, green at S and built by run 7 --
+             # except a name listed in $tmp/off as "NAME<TAB>STATUS<TAB>SHA<TAB>RUN", which gets a target of its own
+    find "$tmp/dist" -type f -printf '%f\n' | jq -Rn --rawfile off "$tmp/off" '
+      [inputs] as $all
+      | ($off | split("\n") | map(select(. != "") | split("\t"))) as $o
+      | ($o | map(.[0])) as $skip
+      | {run_id: 7, targets: ({
+          g: {status: "green", green_sha: "S", built_run_id: 7,
+              tools: ([$all | .[] | select(. as $n | $skip | index($n) | not)] | map({key: ., value: {asset: .}}) | from_entries)}}
+        + ($o | map({key: ("o-" + .[0]), value: {status: .[1], green_sha: .[2], built_run_id: (.[3] | tonumber),
+                                                 tools: {x: {asset: .[0]}}}}) | from_entries))}' > "$tmp/manifest"
+  }
+  run() { # -> rc in $rc, output in $out; fresh delete log and manifest
+    : > "$tmp/deleted"
+    mkplan
+    rc=0
+    out=$(GITHUB_REPOSITORY=o/r NIGHTLY_PRUNE_API="$tmp/api" prune "$tmp/dist" "$tmp/manifest" S 2>&1) || rc=$?
+  }
+  mkdir -p "$tmp/dist"
+  : > "$tmp/off"
+
+  # the measured release: 234 assets, staged and real interleaved, every real one rebuilt
+  : > "$tmp/assets"
+  for i in $(seq 1 234); do
+    if [ $((i % 2)) -eq 1 ]; then printf '%s\tstaged.a%s\n' "$i" "$i"; else printf '%s\ta%s\n' "$i" "$i"; : > "$tmp/dist/a$i"; fi
+  done >> "$tmp/assets"
+  run
+  check "234 assets over 3 pages: exit 0" "$rc" 0
+  check "every staged leftover is deleted, past page one too" "$(grep -c . "$tmp/deleted")" 134
+  check "the 117 staged ones are all among the deletes" \
+    "$(awk 'NR%2==1{print $1}' "$tmp/assets" | grep -cxFf - "$tmp/deleted")" 117
+  check "real assets past publish's page one that are re-uploaded: the 17 old copies go" \
+    "$(awk 'NR%2==0{print $1}' "$tmp/assets" | grep -cxFf - "$tmp/deleted")" 17
+  check "the first 100 real assets are never deleted (publish swaps them itself)" \
+    "$(awk 'NR%2==0{print $1}' "$tmp/assets" | head -100 | grep -cxFf - "$tmp/deleted" || true)" 0
+
+  # a real asset past page one that this run does NOT rebuild is some arch's last green: kept
+  rm -f "$tmp/dist/a234"; : > "$tmp/dist/a234.old"  # a planned name that CONTAINS a234 must not match it
+  run
+  check "a past-page-one asset this run does not upload is kept" "$(grep -cx 234 "$tmp/deleted" || true)" 0
+  check "... and the other 16 old copies still go" \
+    "$(awk 'NR%2==0{print $1}' "$tmp/assets" | grep -cxFf - "$tmp/deleted")" 16
+
+  # a file in dist that publish's plan leaves out is not uploaded, so its old copy is kept -- one row per
+  # filter of publish_plan: a red target, a target green only at an older sha, a night reused from run 6
+  for o in 'red	S	7' 'green	S0	7' 'green	S	6'; do
+    printf 'a232\t%s\n' "$o" > "$tmp/off"
+    run
+    check "a past-page-one asset in dist outside the plan (${o//	/ }) is kept" "$(grep -cx 232 "$tmp/deleted" || true)" 0
+    check "... and the other 15 old copies still go (${o//	/ })" \
+      "$(awk 'NR%2==0{print $1}' "$tmp/assets" | grep -cxFf - "$tmp/deleted")" 15
+  done
+  : > "$tmp/off"
+
+  # an exact multiple of the page size ends on the empty page, and the manifest counts as uploaded
+  : > "$tmp/assets"
+  for i in $(seq 1 200); do printf '%s\tb%s\n' "$i" "$i"; done >> "$tmp/assets"
+  printf '201\t%s\n' "$MANIFEST_ASSET" >> "$tmp/assets"
+  run
+  check "201 assets: page 3 is read and the manifest past page one is deleted" "$(cat "$tmp/deleted")" 201
+  sed -i '$d' "$tmp/assets"
+  run
+  check "200 assets, none rebuilt: nothing deleted" "$(grep -c . "$tmp/deleted" || true)" 0
+  check "... and the run says it read 200 assets" "$(printf '%s' "$out" | grep -c '200 asset(s)')" 1
+
+  # first run, and fail-closed
+  touch "$tmp/norel"; run
+  check "no release yet (404): exit 0, nothing deleted" "$rc/$(grep -c . "$tmp/deleted" || true)" 0/0
+  rm -f "$tmp/norel"; touch "$tmp/fail"; run
+  check "an API failure is never an empty release: exit nonzero" "$rc" 1
+  rm -f "$tmp/fail"
+  touch "$tmp/badpage2"; run
+  check "a 200 page that is an empty JSON object, not an array, is never the empty last page: exit nonzero" "$rc" 1
+  rm -f "$tmp/badpage2"
+  check "a missing dist dir: exit nonzero" \
+    "$( (GITHUB_REPOSITORY=o/r NIGHTLY_PRUNE_API="$tmp/api" prune "$tmp/nope" "$tmp/manifest" S) >/dev/null 2>&1; echo $?)" 1
+  echo '{}' > "$tmp/nomanifest"
+  check "a manifest with no targets is an unreadable plan, never an empty one: exit nonzero" \
+    "$( (GITHUB_REPOSITORY=o/r NIGHTLY_PRUNE_API="$tmp/api" prune "$tmp/dist" "$tmp/nomanifest" S) >/dev/null 2>&1; echo $?)" 1
+
+  rm -rf "${tmp:?}"
+  if [ "$fails" -eq 0 ]; then echo "nightly_prune_release self-test: PASS"; else echo "nightly_prune_release self-test: FAIL ($fails)"; fi
+  [ "$fails" -eq 0 ]
+}
+
+usage() { echo "usage: nightly_prune_release.sh DIST MANIFEST SHA, or --self-test" >&2; exit 2; }
+
+case "${1:-}" in
+  --self-test) self_test ;;
+  ""|-h|--help) usage ;;
+  *) [ "$#" -eq 3 ] || usage
+     prune "$1" "$2" "$3" ;;
+esac