Fix npm artifact false alarms and stale builtins manifest references

.github/ISSUE_SPECS/176-reconcile-builtins-benchmark-manifest-consumers.md

@@ -0,0 +1,125 @@
+# Builtins benchmark identity pins stale since efcfa0f: preflight cannot pass on main
+
+## Approved current scope — 2026-09-11
+
+Correct the CURRENT manifest-only mismatch, not the historical three-pin description.
+Actual fixtures/builtins/v1/MANIFEST.tsv SHA-256 is
+31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796; tools/bench already
+uses it and required CI exercises its input-hash test. PCM508c8e94 and meter958a7026
+identities already match; preserve all fixture bytes. Remaining outdated ddb4b201
+consumers are scripts/preflight-builtins-benchmark.sh, scripts/test-builtins-benchmark.sh
+(including its stub/seal rows), scripts/builtins-benchmark-record-validator.jq and
+tools/audit/src/builtins_graph.rs. Reconcile those exact provenance values. Add one
+cheap real-tree consumer-hash regression with a stale-constant negative control,
+using existing test/preflight machinery and required CI wiring only if needed. Do
+not repin unrelated validator/source seals or launch timed work; if existing runner
+defects demand broader repair, preserve evidence and report before expanding scope.
+Verify real-tree freshness, existing hermetic serializer/validator controls and
+focused audit tests/Clippy appropriate to the constant change. Explain current
+versus historical provenance; no claim that all benchmark infrastructure is fixed.
+
+Astra LOW implements; Astra XHIGH independently verifies. Five attempts maximum.
+Root commits exact paths at each coherent compiling/focused-green tranche and pushes
+promptly before more edits. At most two active issues (#285/#176); isolated worktrees
+and no overlapping paths. No timed benchmark, fixture regeneration, compiler captures,
+DSP/runtime change, or performance claim. Preserve actual commands/environment/source/
+exit/output evidence externally; ordinary compiler feedback is corrected within the
+unfinished pass, while substantive failed gates receive bounded adversarial review.
+Required exact-head PR and main qualification, upstream evidence and verified GitHub
+closure precede clean delivered-worktree removal. Root owns delivery and any artifact
+qualification; these tooling-only slices should not require a new shipped pin.
+
+## Attempt 1 implementation record — 2026-09-11
+
+The current mismatch is manifest provenance only: the four outdated consumers now
+use `31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796`,
+matching the unchanged real manifest and the already-current bench consumer. The
+historical three-pin account below describes an earlier tree, not this repair.
+Fixture bytes and graph PCM/meter pins are unchanged.
+
+The existing benchmark test script now exposes `--check-manifest-consumers`, a
+no-build real-tree check of seven declarations across five consumers (including
+the synthetic preflight hash stub and seal row). It computes the manifest digest,
+requires exactly one matching declaration per selector, and rejects a separately
+mutated stale constant through the same predicate for every declaration. Required
+qualification runs only this cheap mode; the full existing hermetic lifecycle
+suite remains available without running a real benchmark.
+
+No historical lock, validator, source, or artifact seals were repinned. This repair
+does not claim that the historical exactly-once preflight is runnable on the current
+branch or that all benchmark infrastructure is repaired. No fixture generation,
+timed benchmark, DSP change, or performance claim belongs to this checkpoint.
+Validation command/output evidence is preserved externally by the implementation
+workflow; independent review and delivery evidence follow separately.
+
+## Historical issue body
+
+## Summary
+
+Three builtins-benchmark identity pins have been stale since commit `efcfa0f` ("wip: re-pin builtins fixtures from the reference oracles"), which moved `fixtures/builtins/v1/` but did not update the consumers that pin its hashes. On `origin/main` at `edbbeeb` the builtins benchmark preflight could not have passed.
+
+## Evidence (measured on `origin/main`, before any #163 phase-2 change)
+
+| consumer pin | expected | actual file hash on main |
+|---|---|---|
+| `scripts/preflight-builtins-benchmark.sh` `manifest_sha256` | `bfcc7bbe…` | `c33781cf…` |
+| `scripts/preflight-builtins-benchmark.sh` `graph_pcm_sha256` | `508c8e94…` | `e6294eba…` |
+| `scripts/preflight-builtins-benchmark.sh` `graph_meter_sha256` | `958a7026…` | `03cc3979…` |
+
+The same three values are duplicated in `scripts/test-builtins-benchmark.sh`, `scripts/builtins-benchmark-record-validator.jq`, `tools/miso-engine-audit/src/builtins_graph.rs` and `tools/miso-engine-bench/src/builtins.rs`.
+
+`preflight-builtins-benchmark.sh` calls `require_hash fixtures/builtins/v1/MANIFEST.tsv "$manifest_sha256"`, so it fails immediately. It went unnoticed because `preflight-*.sh` is deliberately outside `scripts/sweep.sh` (it is minutes rather than seconds, and every gate it calls is an independent sweep row), and no builtins benchmark has been authorised since `efcfa0f`.
+
+## The corroboration that identifies the cause
+
+Regenerating the fixtures under the #163 phase-2 unfused numeric contract returns **two of the three** to the exact values the consumers still expected, byte for byte and without those constants being touched:
+
+* `fixtures/builtins/v1/pcm/graph-taps.f32le` → `508c8e94…`
+* `fixtures/builtins/v1/meters/graph-taps.jsonl` → `958a7026…`
+
+Both are byte-identical to the fixture as it stood at commit `092ded7`, one commit before `efcfa0f`. That is independent evidence that `efcfa0f` re-pinned the fixtures against **fused** arithmetic while its consumers kept the pre-fusion bits, rather than the two sides having drifted for unrelated reasons.
+
+Only `MANIFEST.tsv` needed a genuinely new value, because it also covers `pcm/filters-asymmetric.f32le`, `pcm/partition.f32le` and the four benchmark descriptors, which differ from the pre-`efcfa0f` era for other reasons.
+
+## Disposition
+
+The stale pins are repaired on branch `floor-phase2-contract` in commit `6fa5453`, because that commit's fixture regeneration touches the same files and leaving them knowingly divergent across it was not an option. They are **not** a phase-2 change and should be reviewed as a separate concern: the phase-2 contract change did not cause them, and would not have surfaced them if it had not regenerated the same corpus.
+
+## Follow-up worth considering
+
+1. The three values live in five places each. A single source (the manifest row itself, read at run time) would make this class of rot impossible.
+2. `preflight-*.sh` being outside the sweep is a deliberate, documented cost decision. It is correct for runtime, but it means a preflight can rot silently for months. A cheap hermetic row that checks only the `require_hash` constants against the working tree — no build, no launch — would have caught this the day it landed.
+
+## Independent review and joint delivery checkpoint
+
+Astra XHIGH attempt1 SOURCE PASS atcc11fea32076d275647256f5028c959424d64963.
+The real CLI rejects the original stale preflight pin and passes after exact
+restoration. Required CI routing contract/mutations pass. All builtins fixture
+tree objects and PCM/meter identities are unchanged; LOW lifecycle/audit/Clippy
+receipts match the committed diff and carry forward. Source/GH/current-base
+parity verified. Evidence: /tmp/issue176-xhigh-cc11fea3/review.md.
+
+Root merges the two accepted histories for one tooling delivery. No shipped
+engine/SDK source or artifact pin changed. Combined exact-head review, required
+PR/main qualification and synchronized closure remain pending.
+
+## PR qualification failure and bounded attempt 2
+
+PR730 qualification34566763819 failed after the graph realtime trace and all
+lifecycle/zero-violation assertions passed: the exact audit report hash was
+c1baf7d5a6b598310c991844cdafa7896967041dd7e451cfea94d78312f3e929,
+while scripts/trace-builtins-graph-audit.sh still expected
+7a960a01270a67a430ee2db03d189f71b36f762f6b30ea555e16afbb42c917b0.
+Astra XHIGH records attempt1 delivery FAIL. Its independent reconstruction of
+the exact1006-byte report matches CI; changing only accepted_manifest_sha256
+back to ddb4b201 reproduces the old pin byte-for-byte. No audio, counter or
+lifecycle field changed. Failed log: /tmp/tooling285-176-pr-ci-failure.log.
+
+Amend #176's exact paths only to include scripts/trace-builtins-graph-audit.sh.
+Attempt2 updates that directly dependent report checksum, preserving all trace,
+jq and exact-hash gates. This is related manifest provenance, not permission
+for broad historical seal repair. Derivation/source/old-new byte proof must be
+retained and independently rechecked against the final source. No timed
+benchmark, fixture/runtime change, new issue or blind CI rerun. #285 source
+PASS remains earned. Root checkpoints the coherent correction before renewed
+required PR/main qualification.

.github/ISSUE_SPECS/285-prune-npm-trees-from-interchange-artifact-scan.md

@@ -0,0 +1,50 @@
+# interchange source-path scan prunes only ./target — npm trees red it
+
+## Approved current scope — 2026-09-11
+
+Fix only the generated-artifact traversal in
+scripts/check-effect-interchange-qualification.sh to prune node_modules directories
+at any depth, preserving the existing root target exclusion and fail-closed traversal
+errors. Extend scripts/test-effect-interchange-policy.sh with root and nested npm
+Wasm acceptance and genuine source-path Wasm refusal. Do not exempt arbitrary source
+paths or files merely named node_modules. Reuse the existing mutation harness; verify
+normal baseline and all existing policy controls. No generic scan framework.
+
+Astra LOW implements; Astra XHIGH independently verifies. Five attempts maximum.
+Root commits exact paths at each coherent compiling/focused-green tranche and pushes
+promptly before more edits. At most two active issues (#285/#176); isolated worktrees
+and no overlapping paths. No timed benchmark, fixture regeneration, compiler captures,
+DSP/runtime change, or performance claim. Preserve actual commands/environment/source/
+exit/output evidence externally; ordinary compiler feedback is corrected within the
+unfinished pass, while substantive failed gates receive bounded adversarial review.
+Required exact-head PR and main qualification, upstream evidence and verified GitHub
+closure precede clean delivered-worktree removal. Root owns delivery and any artifact
+qualification; these tooling-only slices should not require a new shipped pin.
+
+## Historical issue body
+
+check-effect-interchange-qualification.sh's 'generated artifact exists under a source path' scan prunes only ./target, so any node_modules containing a .wasm (playwright-core's webp_codec.wasm) turns it red. Cost two agents a sweep run each (#272, #278). Fix: prune node_modules/ (any depth) in the scan; self-test row proving a planted wasm under src/ still reds.
+
+🤖 Generated with [Claude Code](https://claude.com/claude-code)
+## Astra LOW attempt 1 focused checkpoint
+
+Only the scan and existing policy harness changed. Actual node_modules directories
+are pruned at any depth; root target exclusion and traversal-status refusal remain.
+Root/nested npm Wasm and root-target acceptance, plus src/module.wasm,
+src/node_modules.wasm and nested src/target/module.wasm rejection controls passed.
+Shell syntax, real-tree baseline and full policy harness exited0. The harness
+includes synthetic benchmark108/fake-launch lifecycle controls, not timed work.
+Receipts: /tmp/issue285-attempt1 (actual argv/env/source/exit/logs). Independent
+review and required PR/main qualification remain pending.
+
+## Independent review and joint delivery checkpoint
+
+Astra XHIGH attempt1 SOURCE PASS at3625c5d2. Fourteen independent controls
+prove old/new npm regression, source-artifact refusal, root-target behavior and
+empty/nonempty traversal-error refusal. Successful LOW baseline/full-policy
+receipts reused. Exact head/base and GitHub parity verified. Review and actual
+commands/source/exits/logs: /tmp/issue285-xhigh-o4odaeua/review.md.
+
+Root merges the two accepted histories for one tooling delivery. No shipped
+engine/SDK source or artifact pin changed. Combined exact-head review, required
+PR/main qualification and synchronized closure remain pending.

.github/workflows/qualification.yml

@@ -315,6 +315,8 @@ jobs:
       # binaries; only the hermetic mutation-test suite stays here.
       - name: Standing builtins fixture manifest mutation tests
         run: bash scripts/test-builtins-fixtures.sh
+      - name: Builtins benchmark real-tree manifest consumers
+        run: bash scripts/test-builtins-benchmark.sh --check-manifest-consumers
       - name: Wasm console benchmark validator mutation tests
         run: bash scripts/test-wasm-console-benchmark.sh
       # test-builtins-benchmark.sh and test-graph-benchmark.sh moved to nightly's

scripts/builtins-benchmark-record-validator.jq

@@ -32,7 +32,7 @@ def usable_text: type == "string" and length > 0 and . != "unknown" and . != "de
 def render_workload:
   . == "full_chain_filters" or . == "identity_chain" or . == "matrix_ramp" or . == "meter_success_full";
 def frozen_manifest_sha256:
-  "ddb4b201dcd4cc00ad445013c9a1b29d9d5f6071f018e649748963c74af4c55b";
+  "31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796";
 def frozen_input_sha256:
   if .workload_kind == "full_chain_filters" and .sample_rate_hz == 48000 then
     "4e5e2c9fc8e2c2400b816715273879f3635f2374133e5775ade18dabee1f6ad9"

scripts/check-effect-interchange-qualification.sh

@@ -238,7 +238,7 @@ if [[ -s "$fixture_scan" ]]; then
     fail 'untracked/generated corpus appeared in interchange fixture directory'
 fi
 artifact_scan="$scratch/artifacts.out"; artifact_error="$scratch/artifacts.err"
-if find . -path './target' -prune -o -type f \( -name '*.o' -o -name '*.a' -o -name '*.so' -o -name '*.dylib' -o -name '*.wasm' -o -name '*.profraw' -o -name '*.jsonl.raw' \) -print >"$artifact_scan" 2>"$artifact_error"; then status=0; else status=$?; fi
+if find . \( -path './target' -o \( -type d -name node_modules \) \) -prune -o -type f \( -name '*.o' -o -name '*.a' -o -name '*.so' -o -name '*.dylib' -o -name '*.wasm' -o -name '*.profraw' -o -name '*.jsonl.raw' \) -print >"$artifact_scan" 2>"$artifact_error"; then status=0; else status=$?; fi
 if [[ "$status" -ne 0 ]]; then cat "$artifact_scan" "$artifact_error" >&2; fail "generated artifact traversal failed (status $status)"; fi
 if [[ -s "$artifact_scan" ]]; then
     fail 'generated artifact exists under a source path'

scripts/preflight-builtins-benchmark.sh

@@ -45,7 +45,7 @@ for path in "$sealed_binary" "$seal" "$raw_output" "$accepted_output" "$stderr_o
 done
 
 lock_sha256=4213efd775d1d1207fea805ccdc01392acb015ae36d1bf2eba783f938f19916a
-manifest_sha256=ddb4b201dcd4cc00ad445013c9a1b29d9d5f6071f018e649748963c74af4c55b
+manifest_sha256=31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796
 graph_pcm_sha256=508c8e94244b99ae1ee59e4863088ba69c6462127eb0256f85ec72e775a17a19
 graph_meter_sha256=958a702612b76353ae2dbb0f8a03a2e41aafbd90ed72857bc0c39a10b5d1935f
 issue068_source_sha256=0c71b71d864fbdd01aa918c6825abea78c38f0486535bc914af92142a5080d19

scripts/test-builtins-benchmark.sh

@@ -4,12 +4,53 @@ set -euo pipefail
 
 script_directory="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 repository_root="$(cd "$script_directory/.." && pwd)"
+# Required CI runs only this real-tree check; the lifecycle suite below remains hermetic.
+[[ "$#" == 0 || ( "$#" == 1 && "$1" == --check-manifest-consumers ) ]] || {
+  printf 'usage: %s [--check-manifest-consumers]\n' "$0" >&2; exit 2;
+}
+if [[ "${1:-}" == --check-manifest-consumers ]]; then
+  python3 - "$repository_root" <<'PYTHON'
+import hashlib
+from pathlib import Path
+import re
+import sys
+
+root = Path(sys.argv[1])
+expected = hashlib.sha256((root / "fixtures/builtins/v1/MANIFEST.tsv").read_bytes()).hexdigest()
+# Match each declaration, including the synthetic preflight hash stub and seal row.
+consumers = [
+    ("scripts/preflight-builtins-benchmark.sh", r"^manifest_sha256=([0-9a-f]{64})$"),
+    ("scripts/test-builtins-benchmark.sh", r'^manifest64="([0-9a-f]{64})"$'),
+    ("scripts/test-builtins-benchmark.sh", r'MANIFEST\.tsv\)\s+hash=([0-9a-f]{64})'),
+    ("scripts/test-builtins-benchmark.sh", r'fixture_manifest_sha256:"([0-9a-f]{64})"'),
+    ("scripts/builtins-benchmark-record-validator.jq", r'def frozen_manifest_sha256:\s+"([0-9a-f]{64})";'),
+    ("tools/audit/src/builtins_graph.rs", r'const ACCEPTED_MANIFEST_SHA256: &str =\s+"([0-9a-f]{64})";'),
+    ("tools/bench/src/builtins.rs", r'const INPUT_MANIFEST_SHA256: &str =\s+"([0-9a-f]{64})";'),
+]
+
+def fresh(source, pattern):
+    return re.findall(pattern, source, re.MULTILINE) == [expected]
+
+for name, pattern in consumers:
+    source = (root / name).read_text()
+    if not fresh(source, pattern):
+        sys.exit(f"stale or missing builtins manifest consumer: {name} ({pattern})")
+    match = re.search(pattern, source, re.MULTILINE)
+    # Mutate just the captured constant, then exercise the same freshness predicate.
+    stale = "0" * 64 if expected != "0" * 64 else "1" * 64
+    mutated = source[:match.start(1)] + stale + source[match.end(1):]
+    if fresh(mutated, pattern):
+        sys.exit(f"stale-constant negative control accepted: {name}")
+print(f"builtins manifest consumers: PASS ({len(consumers)} declarations; stale controls rejected)")
+PYTHON
+  exit 0
+fi
 command -v jq >/dev/null || { printf 'jq is required for benchmark validator tests\n' >&2; exit 1; }
 
 hash64="4e5e2c9fc8e2c2400b816715273879f3635f2374133e5775ade18dabee1f6ad9"
 binary64="bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb"
 output64="cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"
-manifest64="ddb4b201dcd4cc00ad445013c9a1b29d9d5f6071f018e649748963c74af4c55b"
+manifest64="31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796"
 commit40="0123456789abcdef0123456789abcdef01234567"
 
 record="$(jq -cn --arg hash "$hash64" --arg binary "$binary64" --arg output "$output64" --arg manifest "$manifest64" --arg commit "$commit40" '
@@ -600,7 +641,7 @@ case "$1" in
   *Cargo.lock)
     hash=4213efd775d1d1207fea805ccdc01392acb015ae36d1bf2eba783f938f19916a ;;
   *fixtures/builtins/v1/MANIFEST.tsv)
-    hash=ddb4b201dcd4cc00ad445013c9a1b29d9d5f6071f018e649748963c74af4c55b ;;
+    hash=31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796 ;;
   *fixtures/builtins/v1/pcm/graph-taps.f32le)
     hash=508c8e94244b99ae1ee59e4863088ba69c6462127eb0256f85ec72e775a17a19 ;;
   *fixtures/builtins/v1/meters/graph-taps.jsonl)
@@ -660,7 +701,7 @@ write_fake_nonbenchmark() {
       lifecycle_sha256:$lifecycle,
       record_validator_sha256:"45f2e0196b4e457a633980653536bb397af7f8ebc82ea69f49c8812dfa7dd9a6",
       aggregate_validator_sha256:"6085e740f15d7902fca4443d761cfb8e29df7168ba12f632c7946db56a3e1b63",
-      fixture_manifest_sha256:"ddb4b201dcd4cc00ad445013c9a1b29d9d5f6071f018e649748963c74af4c55b",
+      fixture_manifest_sha256:"31798260263396c242c0b90042e01abb18624f383fd88029341dffecde662796",
       graph_pcm_sha256:"508c8e94244b99ae1ee59e4863088ba69c6462127eb0256f85ec72e775a17a19",
       graph_meter_sha256:"958a702612b76353ae2dbb0f8a03a2e41aafbd90ed72857bc0c39a10b5d1935f",
       accepted_issue068_source_sha256:"0c71b71d864fbdd01aa918c6825abea78c38f0486535bc914af92142a5080d19",