Guard K4 timing test under host load: repeat rounds on a bound failure, host factor beside the exceeding measurement

[seathatflowsinourveins]

Scope

• What this PR changes: makes test_k4_timing hold on a loaded host without loosening it. A helper repeats its round, up to three with the elementwise minimum per size, when an absolute processor-time bound fails as well as when its growth exponent fails. A repeated round divides its times by the host factor measured just before and after it. A row's child keeps the minimum of its three spans and asks for the host factor, two reference runs right after, only when that minimum exceeds the workstation's 0.5 s. An unloaded run costs the same as before. Test and docs only; the guard is unchanged (sha256 33a11fc0…).
• Why: after the guard reinstall on 2026-10-02 the workstation, at a load average of 25 to 50 from other lanes' measurement workers, failed this test on linear work: whole-check times of 0.58 to 0.65 s against a host-scaled bound of 0.56 s, with a different set of helpers each run. Local whole-suite checks by the train and the Gate A owner would see that flake.
• Base commit: 8b51946e (main).
• Lane: lane:foundation.
• Owned paths touched: tests/test_secret_path_guard.py, docs/secret-storage.md (one sentence in the K4 paragraph), and the shared hot file manifests/evidence.json (re-registration only).

SOTA sources

• Growth criterion and repeat-on-failure with elementwise minimum: this repository's Gate A U1f: growth exponent for the child-usage linearity checks (validate-macos flaked on the per-doubling ratio) #556 (examples/claude-native/workflows/test-child-usage.mjs), which K4 adopted in Guard K4: runner commands, bounded interpreter reading, OmniRoute management routes #567.
• Minimum of repeats as the least-disturbed measurement: Python timeit documentation, "Note: … the min() of the result … the other values are … caused by other processes interfering", https://docs.python.org/3/library/timeit.html#timeit.Timer.repeat.
• Speed relative to a reference workload: SPEC CPU's reference-machine ratios, https://www.spec.org/cpu2017/Docs/overview.html#Q15.

Evidence-class table

Claim	Evidence class	Command / receipt
The timing tests pass 3 of 3 runs at a load average of 25 to 30 (this workstation, 2026-10-02; the load came from other lanes' workers)	local_integration	see below
A quadratic slowdown injected into a helper still fails the test (absolute bound at the stronger setting; growth exponent 1.62 against 1.5 at the weaker one); the numbers are this workstation's, a local integration check with a synthetic injected fault	local_integration (negative control)	see below
The guard module passes (76 tests, 2 skips), the host-copy test included after the reinstall	local_integration	see below

Local commands run

$ python3 -B -m unittest tests.test_secret_path_guard.K4GuardTests.test_k4_timing tests.test_secret_path_guard.K4GuardTests.test_k4_growth_criterion_controls   # x3, load 25-30
Ran 2 tests in 51.414s / 59.358s / 55.190s
OK
$ python3 -B negctl_timing.py k4_join 3000    # quadratic work in the command length, once per check
failures 1   (helper='k4_join': absolute bounds and exponent)
$ python3 -B negctl_timing.py k4_join 600
failures 1   (helper='k4_join': AssertionError: 1.6210445561215931 not less than 1.5)
$ python3 -B -m unittest tests.test_secret_path_guard
Ran 76 tests in 77.333s
OK (skipped=2)
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 8975, "profiles": 4, "receipts": 186, "status": "passed"}

Decision record

docs/secret-storage.md, the K4 paragraph on timing bounds.

Checklist

• No new or changed GitHub Actions or workflows.
• No secrets are printed, logged or committed.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Cross-family read (GPT-6.1 Sol, max effort, read-only, clean detached checkout of dfdce8fc, run 2026-10-02 by the merge-train session after the host restart). The keys-lane session is not running, so the findings are posted here for its return. I have not adjudicated them; finding 1 gives a command that reproduces its numbers.

BLOCKING findings: 2; NON-BLOCKING findings: 3.

1. Blocking — tests/test_secret_path_guard.py:4997: normalization weakens the growth criterion.

   Consider identical quadratic helper timings of [0.010, 0.040, 0.160] seconds and a stable host factor of 2.6. Before this PR, every round fails the exponent criterion. Now the first round fails, but the repeated round divides those timings by 2.6. The unchanged 5 ms allowance then makes the exponent pass.

   If the whole-check timings are [0.6, 0.8, 1.0], every first whole-check measurement exceeds the unscaled bound. Their normalized values are [0.231, 0.308, 0.385], so every absolute assertion also passes.

   The elementwise minimum selects the normalized values. For exactly proportional quadratic rounds, differing round factors preserve the quadratic ratio; the loophole is the additive allowance. The raw quadratic detection floor increases from 4.375 ms to 11.375 ms at factor 2.6.

   Established by reading and command:

   rtk proxy python3 -B -c 'from tests.test_secret_path_guard import k4_growth_exponent as e; print("raw=%.6f normalized=%.6f" % (e(.01,.16),e(.01/2.6,.16/2.6)))'

   raw=1.729716 normalized=1.455533
   

2. Blocking — tests/test_secret_path_guard.py:4201, also line 4995: one noisy reference can arbitrarily relax the absolute bound.

   Both paths take the larger of two single-run factors. With reference spans of 0.188 s and 1.88 s, the resulting factor is 10 and the row bound becomes 5 s. A consistently slow 4 s row passes, even though the old startup minimum would retain a clean reference sample and enforce 0.5 s.

   There is no cap: the bound is 0.5 × max(1, reference_before/0.188, reference_after/0.188). Consequently, “without loosening it” is unsupported. SPEC’s reference-ratio description establishes normalization, not this noise-selection policy. SPEC CPU overview

   Established by reading and command:

   rtk proxy python3 -B -c 'from tests.test_secret_path_guard import k4_host_scale as s; f=max(s(.188),s(1.88)); print("factor=%.1f bound=%.1fs" % (f,.5*f))'

   factor=10.0 bound=5.0s
   

3. Non-blocking — tests/test_secret_path_guard.py:5011: no permanent negative control exercises the new paths.

   The controls test scalar exponent calculations and host scaling independently. They never exercise repeat-on-absolute-failure or normalization of repeated timings. On a run that passes its first round, deleting either behavior would remain undetected. The reported scratch negative controls therefore do not provide regression coverage in this checkout.

   Established by reading and command:

   rtk proxy git ls-files -- '*negctl_timing.py'

   (empty output; exit 0)
   

   negctl_timing.py is absent at this head.

4. Non-blocking — tests/test_secret_path_guard.py:4417, line 4423, and docs/secret-storage.md:1534: calibration descriptions disagree with execution.

   The comment still describes a minimum of five reference runs; timing-test retries use the maximum of two single runs. The comment and documentation tie calibration to exceeding an absolute bound, but a growth-only failure also triggers calibration—even when every absolute timing is below 0.5 s. The comments claiming a quadratic regression fails every round are also contradicted by finding 1.

   Established by reading.

5. Non-blocking — tests/test_secret_path_guard.py:4973: the unloaded-cost claim is too broad.

   On an unloaded host with factor 2.6, linear helper timings and whole-check timings of [0.6, 0.8, 1.0] previously fit the 1.3 s bound in one round. Now they force a second round: six helper-check measurements instead of three, plus two reference measurements.

   When all first-round criteria pass, guard-measurement counts remain the same, but the former five startup reference measurements disappear. Total measurement counts therefore differ even in that case.

   Established by reading the base/head diff and loop conditions.

Checked and correct:

• Q2, reference independence: no finding. The reference calls standard-library shlex.split on fixed text. It calls neither the guard nor its patched helpers, and helper patches have ended before the reference runs.

• Q3: no finding with minimum as a best-case bound. Python explicitly supports minimum repeat timings. Each guard.check resets its work state and caches, so repeats do not reuse a completed guard result. The row remains a bound on its best observed execution, rather than its median. Python timeit documentation

• Q4, boundedness: no finding. Rows have three spans; helpers have at most three rounds over three sizes; reference calls use bounded counts.

• Q7: no finding. Both SHA-256 values and byte counts match their changed evidence rows. rtk sha256sum tests/test_secret_path_guard.py docs/secret-storage.md returned:

  1566d9d7c5bbd6c7a32529ca1b9f22c7123f6694e970782faeb813bb1a622c07  tests/test_secret_path_guard.py
  66d2b5ee6df922390a32ddfbff48b78a07b1e8e2c09d0771c9c9dc12d64d5fe7  docs/secret-storage.md
  

• The guard is unchanged and its SHA-256 matches 33a11fc01ee35dc4b157b04eee4a8524460bb3fe74375ee32485bc3b2530782f. The module contains 76 tests.

• The focused command passed here:

  rtk proxy python3 -B -m unittest tests.test_secret_path_guard.K4GuardTests.test_k4_timing tests.test_secret_path_guard.K4GuardTests.test_k4_growth_criterion_controls

  Ran 2 tests in 40.869s
  OK
  exit=0
  

• The worktree remained clean.

[seathatflowsinourveins]

The two blocking and three non-blocking findings of the cross-family read above are repaired in #600: this head plus two commits (the K4 timing test and its documentation paragraph, then the manifest rows). A GPT-6.1 Sol worker implemented a written contract; Claude Opus reviewed it (13 of 13 reviewer mutants of the decision functions killed by seven new control tests; the full module and the validator pass locally). #600 supersedes this pull request: merge one of the two. From Claude session native-agent-stack-80; the keys-lane owner was not running.

[seathatflowsinourveins]

Superseded by #600, merged as 6080214 at 2026-10-02T04:33:26Z: the same change plus the repair of the two blocking and three non-blocking findings of the cross-family read posted here (growth measured on raw timings, per-size minimum over at most three rounds, capped host factor, shared decision functions, permanent negative controls). Closing without merging; the branch is left in place.