Tooling: fast guard for the 35 live-checkout files the selected #1094 replay hash-asserts, plus complete replay triggers (no new kappa)

[thakrarsagar]

Change and mathematical scope

Tooling only. No new κ, and no change to any construction, certificate,
selected record, pinned document or pinned verifier. Three files:

• scripts/check_selected_ci_inputs.py: a new standard-library guard.
• tests/test_selected_ci_inputs.py: 21 tests, about 3 s. make verify-tests
  picks them up automatically in the generic matrix's tests group, on every
  Python version.
• .github/workflows/frontier1094.yml: new path filters only; nothing existing
  is removed.

Gap. make frontier1094-verify passes --main ROOT to package stages. Those
stages hash-assert 35 repository files outside the selected package roots
through each root's data/UPSTREAM-PINS.json and construction/INPUTS.json
files.main. The replay's first stage also checks the #1042 checkpoint record
(certificates/checkpoints/frontier1042.json), meaning its certificate and its
12-root inventory.

The per-PR check of the selected record (verify_selected_result.py) covers
neither, and at dadfea6e none of these files triggers frontier1094.yml.
Which checks those files triggered:

• 13 baseline558 files under community-k492-audit/pr622 and 2
  structural-gain files: no fast check fired.
• The Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 checkpoint's certificate and inventory: no fast check fired.
• The other 20 (complex-k492-owner-repair/**): caught only by that
  package's own manifest workflow.

So a one-byte change could pass every fast check and surface only inside the
roughly 2-hour cold replay. Reproduced on a scratch worktree by appending one
byte to .../baseline558/.../split-centres/CLIFFORD-RECIPES.json:

• verify_selected_result.py printed PASS, and tests/test_selected_result.py
  passed.
• research/gauge44-slots-t16/verify_t16.py --main . raised
  AssertionError: Upstream dependency changed.
• The new guard printed FAIL_SELECTED_CI_INPUTS with that file in its drift
  list.

Guard. Everything is derived from certificates/selected-result.json.

Hard checks (exit 1):

• Every live pin matches the checkout.
• No two roots pin different bytes, and any in-root pin equals the selected
  source inventory.
• Every selected-result-format record named by the driver's static script
  closure passes verify_selected_result.check. Today those records are
  checkpoints/frontier1042.json and selected-result.json.

Advisory, enforced by --require-triggers:

• The one workflow whose run: commands invoke the driver, or its make target,
  must trigger on push and pull request for every required path:
  • every live pin;
  • every inventoried root file;
  • the driver's script closure and the files it names;
  • each named record's certificate, manifest and inventory.
• Trigger forms the minimal reader cannot evaluate exactly are reported, never
  assumed. These include paths-ignore, types, tags, branch filters
  without main, flow-style on:, and ? + [] {} globs.
• The test enforces complete coverage only while Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 is selected, so a future
  selection commit is not blocked by trigger bookkeeping.

Remaining obligations / scope.

• This is a drift and trigger guard, not a replay or a mathematical check.
• The driver closure is static: imports plus string-literal repository paths.
• Inputs read only by stages that route no pin through these manifests are out
  of scope.
• I did not run the cold make frontier1094-verify (2 hours, and it needs
  Boost).

Task and overlap

• Base and track: base dadfea6e. Seed 94 → verification engineering. Card
  portable-replay, scope refined: the "environment failure" addressed is a
  CI drift and trigger gap in the selected reproduction chain.
• How it was chosen: five explorers covered stage profiling, canonical
  payloads, dependencies, CI workflows and control-suite memoisation. Judges
  scored their candidates for impact, rigor and novelty, and this one ranked
  highest.
• Overlap: none found. find_overlap.py ("path filter", "frontier1094.yml",
  "UPSTREAM-PINS", "live checkout") and open-PR searches turned up only Bind the selected gen4 certificate to fresh replay outputs #474
  (historical gen4 path filters).
• Different objects:
  • Research audit: the selected #1094 reuse replay discards its ancestry recheck; binding certified from committed archives (no new kappa) #1277 (mine): the Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 reuse-mode ancestry binding.
  • Independent verification: selected PR1042 replay-capsule closure (no new κ) #1142: Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 capsule closure.
  • Tooling: verify T37 stage-log coverage with omission controls #1073 (mine): T37 log coverage.
  • Tooling: overlap bit589 and complex741 in the #741 ancestry replay, 17% (50% with --complex-jobs 3) less wall time (no new kappa) #1101, Tooling: fast exact HPE transport audit for the layer-end controls, 2.53x, 480/480 trials cross-checked (seed 93, no new kappa) #1137, Tooling: exact opt-in 2.3x acceleration of the transport audit behind the control suites (no new kappa) #1156, Tooling: opt-in parallel layer-end controls for the frontier1042 leaver chain, byte-identical receipts, 3.1x (extends #1059, no new kappa) #1151, Tooling: 2.33x faster complete emitted local audit with preserved coverage #1184: runtime work.

Validation

python3 -I -B scripts/check_selected_ci_inputs.py --require-triggers
python3 -B -m unittest discover -s tests -p test_selected_ci_inputs.py

• Current tree:
  • Guard: PASS_SELECTED_CI_INPUTS with 35 live pins, 0 drift, 2 named
    records, 0 record failures, and 0 uncovered paths across 1,669 required
    paths. Takes about 0.8 s.
  • Tests: 21 OK on Python 3.12.15 and 3.14.8, about 3 s. Both files parse with
    feature_version=(3, 11).
• Failure controls in the tests:
  • every single-byte change to each of the 35 pins;
  • a missing pinned file;
  • a conflicting pin and an in-root pin that differs from the inventory;
  • path-escape pins;
  • a later !pattern dropping a pull_request trigger;
  • paths-ignore, types, tags and a non-main branch;
  • quoted keys;
  • workflow detection: unrelated comments are ignored, and a second workflow
    with a compact - run: or make <target> is rejected;
  • a changed checkpoint certificate on a synthetic root;
  • exit codes, with triggers advisory unless required.
• Mutation testing of the guard: 21 of 22 mutants killed. These include
  dropping either pin manifest, emptying the closure, ignoring records,
  negation or unsupported filters, push-only events, and * matching /. The
  survivor is equivalent: the closure names the selected record itself, so its
  inventory already reaches the required set through the record check.
• YAML reader: agrees with PyYAML on all 134 evaluable (workflow, event)
  blocks in the repo. The 10 it refuses are genuinely different forms: eight
  push filters on codex/* branches and one flow-style on:.
• Historical selections: the guard exits 0 at Conditional κ = 0.001635183729255778: second dim-4 cold end per factory block (28 per word) on the PR604 cold-ended witness #617, Conditional κ = 0.001635527380375392: coordinated owner repair on PR623 #626, Conditional κ = 0.001661202643272239: eight B-frame coalescences after SW4 #741, Conditional κ = 0.001661634653669567: four paired read exchanges on PR741's frame-coalesced word #799, Conditional κ = 0.001661957774647029: T21 whole-owner packing and fixed-family bound #890,
  Conditional κ = 0.001661960547713947: T37 whole-owner packing and all-T bound #942, Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042 and the Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 integration commit. Hard checks hold at all of them,
  and trigger gaps there are only reported. That is why trigger coverage is
  advisory.
• Related tests still pass: tests/test_selected_result.py,
  test_pr_reconciliation.py, test_isolated_test_runner.py,
  test_selected_inventory_links.py and test_selected_proof_binding.py.
  scripts/verify_pr_reconciliation.py also passes.
• Pre-PR review: a three-lens review (soundness, claims, acceptability)
  found one blocker, which is fixed. docs/ci-verification.md is pinned by
  CONTENT-COVERAGE, so the stale "current selected replay" section there is
  left for a maintainer. Its Conditional κ = 0.001635527380375392: coordinated owner repair on PR623 #626 description is out of date. The review
  also drove four other fixes:
  • stricter trigger parsing;
  • run:-only workflow detection;
  • checkpoint-record coverage;
  • advisory triggers.
• Load from the new triggers: about 63 additive paths per event. They are
  frozen vendored evidence, selected roots and chain scripts, which change
  rarely. git log shows 2 commits ever touching the 35 pins.

Credit

• Douglas Colkitt: the selected-record check, the Conditional κ = 0.001672188119205421: improved T16 owner packing on the seven-slot k8 word #1042/Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 replay chain and
  the CI workflows this guard reads.
• The gauge44 / k492 package authors, whose UPSTREAM-PINS / INPUTS pins
  define the live inputs.
• GitHub path-filter semantics as documented by GitHub.

AI disclosure: prepared by Sagar Thakrar with substantial Anthropic Claude
assistance, including parallel explorer, judge and skeptic subagents. New work
is Apache-2.0.

Open next questions

1. Should docs/ci-verification.md describe the Conditional kappa 0.001672660197076918: rare-pattern T16 exchanges #1094 chain?
   • Why it matters: it still names Conditional κ = 0.001635527380375392: coordinated owner repair on PR623 #626 as the current selected replay.
   • First step: a maintainer edit plus a CONTENT-COVERAGE re-pin.
   • Blocker: the file is pinned.
2. Should research-packages.yml trigger on non-research/ declared inputs?
   • Why it matters: its path filter misses inputs such as scripts/ and
     certificates/, which registered packages can declare. Research audit: the selected #1094 reuse replay discards its ancestry recheck; binding certified from committed archives (no new kappa) #1277's package is
     an example.
   • First step: the same coverage idea, applied to plan()'s declared
     inputs.
3. Should the closure be dynamic instead of static? Stages that read
   repository files without routing them through a pin manifest are out of
   scope. An audit-hook trace of one full cold run would bound them.

Workflow feedback (optional)

The selected-chain workflows (frontier1094.yml, frontier1042.yml) have
queued for hours without completing. This PR does not change that. Its hard
checks give a sub-second per-PR signal for the files the replay hash-asserts.

[Mister-Meeseeks]

100-PR batch review at 40965b228001a95c1f55054e5ed661fa85190813.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Selected-source CI coverage guard checks live pins, named records and driver closure, with fail-closed workflow coverage under --require-triggers. Expanded path filters and 21 tests pass; this is trigger/provenance coverage, not a new mathematical replay or selected bound.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.