Record the macOS full-suite coverage decision, the CI measurements and five anti-patterns

docs/harness-defaults.md

@@ -86,6 +86,11 @@ When a claim or action proves wrong, correct it where it was relayed and record
 
 | Date | Anti-pattern | What happened | Rule or check that prevents it | Where enforced |
 | --- | --- | --- | --- | --- |
+| 2026-10-03 | Passing `-f` or `-F` to `gh api` for a read | Adding a parameter turned the read into a POST: `gh api --help` at the installed gh 2.102.0 (lines 20-27) says the method is `GET` normally and `POST` if any parameters were added. A read-only audit's listing sent `POST user/repos` and got HTTP 422; nothing was created. | Put read parameters in the URL query string, or pass `--method GET` with them. | This log; verification path: `gh api --help` at the installed version, lines 20-27 |
+| 2026-10-03 | Projecting allowance exhaustion from a partial snapshot | A "gone in 8-9 days" projection for the monthly Actions allowance that private repositories draw on rested on two days of a bursty series (daily job minutes ranged from 0 to 763). | Project from the full daily series, per repository and runner OS, and give a range of scenarios rather than one date. | This log; verification path: daily job minutes recomputed from the jobs API (`gh api --method GET`) per repository and runner OS |
+| 2026-10-03 | Reporting non-success runs as failures | A summary said about half of pull-request runs failed. Grouped by conclusion over 2026-09-25 to 2026-10-02, 6.5% and 7.3% of the pull-request runs of the two workflows that hold the required test jobs failed and 25-33% were cancelled (36-46% on 2026-10-01 and 10-02); 672 of all 712 cancelled runs in the window (701 of them pull-request runs of these two workflows) were superseded by a newer run. The day-wide list for 2026-10-01 had also reached the 1,000-result cap: the complete counts were 176 and 134 pull-request runs, not 160 and 120. | Report success, failure and cancelled separately (`gh run list --json conclusion`, grouped), and partition every query so that no slice's `total_count` reaches 1,000. | This log; [measurement receipt](../evidence/receipts/github-ci-measurements-20261003.json) |
+| 2026-10-03 | Researching through raw web search before using the repository's lanes | The first external lookups of a GitHub CI review ran on WebSearch before the user directed research through the Codex runtime lanes, whose dispatch contract is the [routing record](decisions/2026-09-30-sol-primary-quality-defaults.md). | Route external research through those lanes, and relay each claim with the upstream citation that settled it. | This log; [routing record](decisions/2026-09-30-sol-primary-quality-defaults.md) |
+| 2026-10-03 | Repeating a known guard false positive in a read-only lane | An Explore agent's command with inline Python `set(...)` and the word "environment" was blocked as `environment_dump` by the PreToolUse secret guard, the case of the 2026-09-25 row; that row's remedy, a script file, is unavailable to an agent that may not write files. | In a read-only lane, aggregate with `jq` or `gh --jq` instead of inline Python; send the reproduction to the guard owner for a failing-first case in `tests/test_secret_path_guard.py`; never bypass or edit the guard | `scripts/hooks/secret_path_guard.py` (`is_environment_dump`); this log until the guard owner's failing-first case lands |
 | 2026-10-03 | Pointing an observed record's frozen inputs at a copy of the plan instead of at the files the plan froze | The first observed convergence record of the suite-parallelism trial listed the preregistration copy as its source, input and evaluation artifact, and the oracle's trimmed output as an evaluation input, so `scripts/validate_convergence.py` re-checked none of the preregistered bytes; the delta review of #652 found it | Keep byte copies, with the same sha256, of every file the preregistration froze and list them under the preregistration's own role lists; never list an oracle output as a frozen input; name a lock copy so that no dependency scanner pattern matches it | this log; the trial outcome artifacts under `evidence/artifacts/suite-parallelism-trial-20261003/frozen/` and their README |
 | 2026-10-03 | Putting a free-form receipt in place of a preregistered observed convergence record | The suite-parallelism trial's outcome pull request (#652) first kept the failed trial's per-run data only in a receipt and stored the preregistered record as `.json.txt`, so `scripts/validate_convergence.py --all-recorded` checked nothing about the trial: no artifact hashes, run and failure coverage or usage fields. The automated review of #652 found it | Write the observed record that the preregistration names, in a path that does not re-trigger the trial, validate it with `scripts/validate_convergence.py` and list it in `convergence_records`; a receipt adds detail but does not replace it | AGENTS.md:18; this log; [trial outcome record](decisions/2026-10-03-suite-parallelism-trial-outcome.md) and its observed record |
 | 2026-10-03 | Closing a substantive trial without its completeness critique | The same outcome record listed the runner classes the trial did not evaluate "for a pending survey" and recorded no critique of the missed modality, sources or candidate classes. The review of #652 found it | End every substantive research or adoption unit with the completeness critic, mark each relayed candidate claim as verified against its upstream source or as a lead, and record the seeds for the layer's next landscape sweep in the unit's record | AGENTS.md:18; this log; the trial outcome record's "Completeness critique" |