chore: retire tools/change_gate and the agentic-review route superseded by hw-gate

[Kaden-Schutt]

Summary

The follow-up named in #679's after-merge list: hw-gate is the review rung, so the pre-hw-gate route goes.

Deleted (50 files): tools/change_gate (13), autoresearch/ar/review + tests (22), .github/agentic-review (4), .agents/skills/agentic-pr-{discovery,static-review,review} (10), tests/test_agentic_pr_review_skills.py.

Retargeted at the hw-gate request block / docs/VALIDATION.md § hw-gate: PULL_REQUEST_TEMPLATE.md, CONTRIBUTING.md, docs/VALIDATION.md, scripts/leanup-thresholds.txt. scripts/no-gpu-ci.sh drops the three change_gate unittest modules (the redline modules it still runs all exist; bash -n clean). .gitignore loses the dead providers.local.json entry.

Left in place on purpose: docs/governance/** and docs/perf-checkpoints/** (immutable ledgers), dated plan/spec/result records under docs/, and the unrelated autoresearch kernel-loop package. Residual mentions: 7 lines, all historical.

Which surface(s) does this touch?

• policy files (.github/, no-gpu-ci) — human merge
• docs / CI / scripts

Evidence

• rg -n 'change_gate|change-gate|agentic-review|agentic_review|agentic-pr|autoresearch' --hidden -g '!.git' → only the 7 historical lines above.
• .github/workflows/no-gpu-ci.yml unchanged and parses; every unittest module scripts/no-gpu-ci.sh still names exists.

No code under crates/ or kernels/ changes; no hardware surface.

[hipfire-sol]

hw-gate sol prelim

summary: Retires the superseded change_gate and agentic-review implementations, their tests, configuration, and agent skills; updates contributor and validation documentation to point exclusively at hw-gate; and removes the deleted change_gate unittest modules from no-GPU CI. No inference-engine, model-loading, serving, kernel, dependency, build, or workflow behavior is changed.

run_hardware: true
run_hardware_reasons: The executable changes only stop scripts/no-gpu-ci.sh from invoking three unittest modules deleted by this PR; its remaining commands name existing project test surfaces.; Deleted shell and Python entrypoints belonged to the retired review route and therefore cannot execute during this hardware job.; The PR adds no network access, credential or environment reads, process spawning, dependency or toolchain changes, obfuscated content, or unsafe code.; scripts/leanup-thresholds.txt is a policy path, so the hard floor still requires human handling even though no model fixture is relevant.

routes:

mode	tag	source	why
—	—	—	no routes

unavailable_routes:

(none)

[hipfire-sol]

hw-gate evidence — 2 lane(s) — verdict pass

lane hiptrx (gfx1201)

hw-gate evidence

field	value
base	e23c55e79ee995ee6dde44bce189af7de52155b9
head	28739bd6ef00da4130fced8d9b32e314ded60df1
buckets
host gfx	gfx1201
host rocm	7.15.26333-0000000
device	3
runner	hiptrx
daemon_md5	17ba7dfab49aa1c1110d3f412f072648
hipfire_md5	d21e2a73bd164a2de9ae176934441826
build_seconds	12.151455402374268
verdict	pass
logs_dir	hw-gate-logs

fixtures

no fixtures

kernel

not run

lane hipx (gfx1100)

hw-gate evidence

field	value
base	e23c55e79ee995ee6dde44bce189af7de52155b9
head	28739bd6ef00da4130fced8d9b32e314ded60df1
buckets
host gfx	gfx1100
host rocm	7.15.26333-0000000
device	0
runner	hipx
daemon_md5	17ba7dfab49aa1c1110d3f412f072648
hipfire_md5	d21e2a73bd164a2de9ae176934441826
build_seconds	16.365201473236084
verdict	pass
logs_dir	hw-gate-logs

fixtures

no fixtures

kernel

not run

[hipfire-sol]

hw-gate sol verdict

{
  "claim_verdict": "not-exercised",
  "confidence": 0.94,
  "coverage": {
    "gaps": [
      "The hw-gate evidence contains successful hiptrx and hipx builds but no execution of scripts/no-gpu-ci.sh, so the changed CI script path was not behaviorally exercised.",
      "scripts/leanup-thresholds.txt is a policy-file change and requires human review under the hard floor."
    ],
    "surfaces_evidenced": [
      "docs",
      "build"
    ],
    "surfaces_touched": [
      "docs",
      "ci",
      "scripts",
      "policy"
    ]
  },
  "decision": "needs-human",
  "eyeball": [
    "A human should review scripts/leanup-thresholds.txt and confirm that replacing \u201cchange_gate routes\u201d with \u201cCI gates\u201d preserves the intended ratchet policy.",
    "A human should inspect scripts/no-gpu-ci.sh and, ideally, run it once to confirm removal of the retired unittest modules leaves the remaining no-GPU CI battery intact."
  ],
  "phase": "verdict",
  "rationale": "Both hardware lanes built head 28739bd6ef00da4130fced8d9b32e314ded60df1 successfully with identical daemon and hipfire MD5s, and no load, serve, kernel, model-artifact, or GPU-byte surface changed. No regression is evidenced. However, hw-gate ran no buckets or fixtures and did not exercise scripts/no-gpu-ci.sh; therefore the author's CI-integrity claim is not exercised by the supplied hardware evidence. The change to scripts/leanup-thresholds.txt is independently a policy hard-floor item, requiring needs-human.",
  "regressions": []
}

Floor: hard=['policy_paths: scripts/leanup-thresholds.txt'] soft=["coverage_gaps: ['The hw-gate evidence contains successful hiptrx and hipx builds but no execution of scripts/no-gpu-ci.sh, so the changed CI script path was not behaviorally exercised.', 'scripts/leanup-thresholds.txt is a policy-file change and requires human review under the hard floor.']", 'model needs-human'] model_decision=needs-human final=needs-human

[hipfire-sol]

hw-gate sol verdict needs-human: Both hardware lanes built head 28739bd successfully with identical daemon and hipfire MD5s, and no load, serve, kernel, model-artifact, or GPU-byte surface changed. No regression is evidenced. However, hw-gate ran no buckets or fixtures and did not exercise scripts/no-gpu-ci.sh; therefore the author's CI-integrity claim is not exercised by the supplied hardware evidence. The change to scripts/leanup-thresholds.txt is independently a policy hard-floor item, requiring needs-human.

[hipfire-fable]

announcement: Hold, for the policy floor only: scripts/leanup-thresholds.txt is a policy path and needs a maintainer's human-reviewed label, though the hunk is a single comment word and the ratchet still asserts 21 metrics with 0 violations. I closed the gap Sol flagged by running every Python line of scripts/no-gpu-ci.sh verbatim on your head and on master: the retained redline unittest modules run, the deleted packages are genuinely gone, and autoresearch/ar/tests collects cleanly without review_fixtures. The battery is red on both trees with byte-identical failures (test_mq4c_repack.py, test_golden, check-env-docs.py) — pre-existing on master, none in files you touched, and the passed-count drop is exactly the 457 review tests you deleted. Nothing live references the retired route anymore, and the shipped binaries match base apart from the embedded commit hash.

investigation:

| question | route | result | evidence |

|---|---|---|---|

| Does anything live in the tree still import or exec the deleted packages (tools.change_gate, autoresearch.ar.review, review_fixtures, the .agents skills, .github/agentic-review config)? | grep -rn 'change_gate|change-gate|agentic[-_]review|agentic-pr|ar.review|ar/review|providers.local|capabilities-v1|trusted-publishers' (whole tree, gitignored included), then a targeted pass over .github/ .agents/ scripts/ (incl. scripts/hw-gate/) tools/ autoresearch/ tests/ CLAUDE.md AGENTS.md | Zero live references. The only hits are historical: docs/governance/2026-08-16-phase3-scope.md:75,87,236; docs/superpowers/specs/2026-07-15-agentic-pr-review-workflow-design.md:170,209,290; experiments/flash-attn-ck-sidecar/results/asym3_givens_pp8192_master_20260827/README.md:69; and a doc-comment at crates/rdna-compute/src/attention.rs:3037 citing a design-doc filename. scripts/hw-gate/{select,run,review,merge_evidence}.py import nothing from the deleted packages. | fable-summary.md § Residual references |

| Sol's gap: does the changed scripts/no-gpu-ci.sh path actually run on the PR head — deleted modules gone, retained modules intact, pytest sweep still collects autoresearch/ar/tests without review_fixtures? | bash -n scripts/no-gpu-ci.sh; python3 -c 'import yaml' parse of all 4 workflows; python3 -c 'import tools.change_gate' / 'import autoresearch.ar.review' / 'import autoresearch.ar.tests.review_fixtures'; python3 -m unittest tools.redline.tests.test_product_bench tools.redline.tests.test_golden tools.redline.tests.test_serve_diff tools.redline.tests.test_lower tools.redline.tests.test_dispatch_profile; python3 scripts/test_install_revision.py; python3 scripts/test_uninstall.py; python3 scripts/check-env-docs.py; python3 -m pytest tests scripts/test_astrea.py autoresearch/ar/tests (each verbatim from no-gpu-ci.sh:17,25-30) | bash -n ok; no-gpu-ci.yml/ci.yml/hw-gate.yml/registry.yml parse. All three deleted imports raise ModuleNotFoundError (deletion is complete, no stale pycache shadow). All 5 retained redline modules import; unittest Ran 179 → FAILED (failures=2, errors=3), all in test_golden; test_install_revision OK (14), test_uninstall OK (6); check-env-docs.py EXIT=1; pytest 5 failed / 378 passed / 4 skipped, all 5 failures in tests/test_mq4c_repack.py. autoresearch/ar/tests collects cleanly with review_fixtures.py gone. | no-gpu-ci-pytest-head.log |

| Are the reds in that battery caused by this PR or pre-existing on master? | Same pytest sweep, same unittest line, and same check-env-docs.py run in the base checkout at e23c55e; diff of FAILED/ERROR names and of check-env-docs output head vs base; git diff --stat base..HEAD -- scripts/mq4c_repack.py tests/test_mq4c_repack.py tools/redline scripts/check-env-docs.py crates; pytest --collect-only on the deleted test files at base | IDENTICAL on all three: the same 5 test_mq4c_repack.py FAILED names, the same 5 test_golden FAIL/ERROR names (Ran 179, failures=2, errors=3), byte-identical check-env-docs.py output with EXIT=1. The diff touches none of those files (diff --stat empty). Base passes 835 vs head 378; the deleted test files collect exactly 457 tests on base, and 835 − 457 = 378, so every lost test is one the PR intentionally deleted. No regression. | deleted-test-count-base.txt |

| Does the policy-file edit (scripts/leanup-thresholds.txt) change anything the ratchet reads? | git diff base..HEAD -- scripts/leanup-thresholds.txt; bash scripts/leanup-ratchets.sh . on head | The hunk is one comment line (leanup-thresholds.txt:33, 'change_gate routes' → 'CI gates'); no metric line changes. Ratchet on head: OK — 21 metric(s) asserted, 0 violations; ungated_examples 48 against ceiling 48, unchanged. | leanup-ratchets-head.log |

| Is 'no code under crates/ or kernels/ changes' true at the binary level? | md5sum of daemon, hipfire, hipfire-detect, hipfire-quantize in $HW_GATE_BIN vs $HW_GATE_BASE_BIN; strings + --version on both hipfire binaries | daemon (17ba7dfa…), hipfire-detect (05e11024…), hipfire-quantize (cf19a3ff…) byte-match base. hipfire differs (d21e2a73 vs 5919e558) solely because crates/hipfire-cli/build.rs:34-40 embeds the commit and ref: 'hipfire 0.3.0 (28739bd; detached)' vs '(e23c55e; master)'. No Rust behavior changed. | binary-md5-head-vs-base.txt |

unproven:

• Whether the three pre-existing reds (tests/test_mq4c_repack.py ×5, tools.redline.tests.test_golden ×5, scripts/check-env-docs.py exit 1) also fail on the GitHub no-gpu-ci runner or are specific to this host's Python 3.14.4 / pytest 9.1.1 environment. They are identical on master e23c55e and untouched by this diff, so they do not bear on this decision, but a maintainer expecting no-gpu-ci to be green on master should know it is red here.

• The Rust half of no-gpu-ci.sh (cargo check --workspace --examples, cargo test on the no-GPU crates) was not re-run here; no crates/ or kernels/ file changes, both hw-gate lanes built head, and three of four shipped binaries byte-match base, so the omission is by evidence rather than assumption.

rationale: The hard floor fired on policy_paths (scripts/leanup-thresholds.txt), so the decision space is hold or block; hold is correct because I found no evidence failure. The policy hunk is one comment word at scripts/leanup-thresholds.txt:33 with no metric change, and scripts/leanup-ratchets.sh on head reports 21 asserted / 0 violations with ungated_examples 48 == ceiling. Sol's coverage gap (no-gpu-ci.sh never exercised) is now closed: I ran every Python line of scripts/no-gpu-ci.sh:17,25-30 verbatim on head and again on base. The changed unittest line (scripts/no-gpu-ci.sh:25) runs its five retained tools.redline modules; the three deleted change_gate modules and autoresearch.ar.review raise ModuleNotFoundError; autoresearch/ar/tests collects without review_fixtures.py. The battery is red on head — but every FAILED/ERROR name and the check-env-docs output are byte-identical on master e23c55e, and the diff touches none of the implicated files, so nothing regressed. The 835→378 passed delta is exactly the 457 tests the PR deletes. No live reference to the retired route remains anywhere in .github/, .agents/, scripts/ (including scripts/hw-gate/), tools/, autoresearch/, or tests/; the eight residual mentions are historical docs/specs/an experiment README/a code comment citing a design-doc filename. Binary A/B: daemon, hipfire-detect, hipfire-quantize byte-match base; hipfire differs only by the embedded commit+ref from crates/hipfire-cli/build.rs:34-40. What changes this to a merge: a maintainer applying human-reviewed after reading leanup-thresholds.txt:33 and the reworded .github/PULL_REQUEST_TEMPLATE.md:60 / CONTRIBUTING.md:106-107 lines. Evidence: $HW_GATE_EVIDENCE/fable-summary.md and the seven files it indexes.

hard floor: ['policy_paths: scripts/leanup-thresholds.txt']

soft floor: ["coverage_gaps: ['The hw-gate evidence contains successful hiptrx and hipx builds but no execution of scripts/no-gpu-ci.sh, so the changed CI script path was not behaviorally exercised.', 'scripts/leanup-thresholds.txt is a policy-file change and requires human review under the hard floor.']", 'model needs-human']

[hipfire-fable]

Hold, for the policy floor only: scripts/leanup-thresholds.txt is a policy path and needs a maintainer's human-reviewed label, though the hunk is a single comment word and the ratchet still asserts 21 metrics with 0 violations. I closed the gap Sol flagged by running every Python line of scripts/no-gpu-ci.sh verbatim on your head and on master: the retained redline unittest modules run, the deleted packages are genuinely gone, and autoresearch/ar/tests collects cleanly without review_fixtures. The battery is red on both trees with byte-identical failures (test_mq4c_repack.py, test_golden, check-env-docs.py) — pre-existing on master, none in files you touched, and the passed-count drop is exactly the 457 review tests you deleted. Nothing live references the retired route anymore, and the shipped binaries match base apart from the embedded commit hash. The hard floor fired on policy_paths (scripts/leanup-thresholds.txt), so the decision space is hold or block; hold is correct because I found no evidence failure. The policy hunk is one comment word at scripts/leanup-thresholds.txt:33 with no metric change, and scripts/leanup-ratchets.sh on head reports 21 asserted / 0 violations with ungated_examples 48 == ceiling. Sol's coverage gap (no-gpu-ci.sh never exercised) is now closed: I ran every Python line of scripts/no-gpu-ci.sh:17,25-30 verbatim on head and again on base. The changed unittest line (scripts/no-gpu-ci.sh:25) runs its five retained tools.redline modules; the three deleted change_gate modules and autoresearch.ar.review raise ModuleNotFoundError; autoresearch/ar/tests collects without review_fixtures.py. The battery is red on head — but every FAILED/ERROR name and the check-env-docs output are byte-identical on master e23c55e, and the diff touches none of the implicated files, so nothing regressed. The 835→378 passed delta is exactly the 457 tests the PR deletes. No live reference to the retired route remains anywhere in .github/, .agents/, scripts/ (including scripts/hw-gate/), tools/, autoresearch/, or tests/; the eight residual mentions are historical docs/specs/an experiment README/a code comment citing a design-doc filename. Binary A/B: daemon, hipfire-detect, hipfire-quantize byte-match base; hipfire differs only by the embedded commit+ref from crates/hipfire-cli/build.rs:34-40. What changes this to a merge: a maintainer applying human-reviewed after reading leanup-thresholds.txt:33 and the reworded .github/PULL_REQUEST_TEMPLATE.md:60 / CONTRIBUTING.md:106-107 lines. Evidence: $HW_GATE_EVIDENCE/fable-summary.md and the seven files it indexes.