Prompt audit 2026-09-27: facts refuter exceptions, hash prose, split-table log check

docs/harness-defaults.md

@@ -92,7 +92,10 @@ When a claim or action proves wrong, correct it where it was relayed and record
 | 2026-09-27 | Silently filtering malformed rows from a measured denominator | `_dicts` in `scripts/component_matrix.py` at `b666c42c` dropped non-object sweep-manifest components and entries | Fail closed with the manifest path, catalog/layer and item index, following the existing convergence input checks | `ConvergenceByLayerTests.test_non_object_manifest_rows_fail_closed` in `tests/test_component_matrix.py` failed with `SystemExit not raised` in all eight fixture cases before the fix |
 | 2026-09-27 | Treating a Codex profile as a replacement for config.toml | A draft tiering preregistration claimed the profile loads in place of the base user config. The installed 0.157.1 profile help says it layers on top; [openai/codex rust-v0.157.1 loader L286-334](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/config/src/loader/mod.rs#L286-L334) loads the base and then the profile. [ConfigLayerSource::precedence](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/config/src/config_layer_source.rs) assigns base 20, profile 21, project 25 and session flags 30 | Treat profiles as overrides; a trusted project's config can still win. Pin required worker model/effort/search values on the invocation with `-m` and `-c`. Keep the dated correction and [workstation proof](../evidence/artifacts/codex-worker-lane-host-20260927/README.md) distinct from the separate tiering run | This log |
 | 2026-09-27 | Using strict config as a universal worker-profile validator | Observed on codex-cli 0.157.1 on 2026-09-27: `codex --strict-config -p stack-worker exec` fails with `invalid transport` on the profile's partial `[mcp_servers.*]` tables, and `codex --strict-config debug prompt-input probe` refuses the flag as unsupported for `codex debug`. The [pinned loader L594-600](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/config/src/loader/mod.rs#L594-L600) calls [per-file strict validation as ConfigToml, L625-645](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/config/src/loader/mod.rs#L625-L645); profile tables amend transports supplied by the base config | Validate stack-worker with a normal runtime `codex exec -p stack-worker` carrying the lane pins, then inspect its actual events. [`prove_codex_lane.exec_argv`](../tools/adoption/prove_codex_lane.py) defines the invocation; the [retained five-call proof](../evidence/artifacts/codex-worker-lane-host-20260927/prove-live.txt) records the verdicts. A strict failure on a partial layer does not establish a normal-runtime failure | This log; [worker-lane strict-mode observations](decisions/2026-09-26-codex-worker-lane.md#addendum-2026-09-27-start-up-allowances-the-gateway-profile-and-four-base-keys). Recheck on a Codex upgrade |
-
+| 2026-09-27 | Changing a rule in some co-loaded prompt templates but not all | #385 put the unknown-maintenance and license rules into the landscape sweep's `common` and `fit` templates, but the `facts` template, which the refuter receives with `common` and each proposal's `upstream_now`, still said "Default to refuted=true when uncertain" and counted a wrong license as a refuting fact. In the 2026-09-26 run the facts refuter refuted proposals whose `upstream_now` fields were null, which the discover template allows ("null when unknowable"; `evidence/artifacts/landscape-sweep-20260926/returns.json`, anomalyco/opencode and alpacahq/cli) | When a rule changes, update every template that shares the prompt; the facts template states the unknown-field, maintenance and license exceptions | `tests/test_landscape_sweep_harness.py` (the facts-refuter rule test) |
+| 2026-09-27 | Leaving a blank line inside a Markdown table | A blank line after the fifth 2026-09-27 row (at `55fc8d17`) ended this log's table, since GFM 0.29-gfm 4.10 says "The table is broken at the first empty line", so every row after it fell outside the table; the shape test read only lines that start with a pipe and passed | Keep a table's rows contiguous from the header to the last row | `tests/test_adoption_docs_consistency.py` (the anti-pattern log check) |
+| 2026-09-27 | Restating a linked rule without its scope | `AGENTS.md` required role dispatch for every workflow `agent()` stage, while the Workflow contract of the README it links (`examples/claude-native/workflows/README.md`, its Dispatch by role bullet, from the same commit) exempts the vendored saved scripts | Restate a linked rule with its scope, or link it without restating | This log |
+| 2026-09-27 | Quoting a changing value as current in prose | The landscape-sweep README and the 2026-09-26 run's evidence README still said the packaged templates give the 2026-09-26 prompt hash after #385 changed the templates and the test's pin | Point at the pinned source of a changing value instead of quoting it | This log |
 | 2026-09-27 | Guessing version-specific CLI syntax or source paths | The custom-agent/profile build tried `codex sandbox linux --help`, which ran `linux` as the command, and guessed an `exec/src/env.rs` URL that returned 404 | Read installed `codex sandbox --help` first; at 0.157.1 the form is `codex sandbox -- <command>`. Inspect the matching release and repository tree before citing source paths; `cli/src/debug_sandbox.rs` calls `protocol/src/shell_environment.rs` at [rust-v0.157.1](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/cli/src/debug_sandbox.rs). A 404 is not an absence check | This log; the [worker addendum](decisions/2026-09-26-codex-worker-lane.md#2026-09-27-addendum-custom-agents-and-context-hub) records returned native results |
 | 2026-09-27 | Treating a parsed role field as child authority | The custom-agent README claimed the semantic reviewer's `sandbox_mode = "read-only"` enforced its sandbox; the role loader excludes that field | Trace the applied override list and parent clone in [Codex rust-v0.157.1 role.rs:36–48,119–126,182–189](https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/core/src/agent/role.rs#L36-L48); all roles inherit parent sandbox authority | Corrected [custom-agent README](../examples/codex-native/README.md#2026-09-27-custom-agent-instruction-refresh); this log, no spawned-role enforcement test |
 | 2026-09-26 | Removing an entire owned list from a preservation check | The writer-identity Prometheus merge replaced two host relabel rules with its bucket rules; stripping the whole key hid the loss | Preserve the original list exactly and append only missing owned rules; verify both the prefix and additions before comparing the remaining config | `host/merge_prometheus.py::verify_merge` in the [writer-identity recipe](../evidence/artifacts/telemetry-writer-identity-20260926/host/); `MergePrometheusTests` in `tests/test_observability_writer_identity_host.py` reproduced the dropped rules before the fix |

evidence/artifacts/landscape-sweep-20260926/README.md

@@ -15,8 +15,9 @@ coordinator owns and will run from this manifest.
 The run used the 2026-09-26 scratchpad prototype of the lane. The same lane is now packaged as
 [`tools/sota-convergence/landscape-sweep/`](../../../tools/sota-convergence/landscape-sweep/README.md) (#324). The
 run's `prompts_sha256` is `3adfbed7a83e85da3fd7951032e1fa3a579101772a47b211580065c6b42618d4`, the sha256 of the
-prototype's staged templates. The packaged templates, filled with this run's values, give the same hash (a local check
-in `tests/test_landscape_sweep_harness.py`). The packaged tools produced the returns, lanes, layers, usage records,
+prototype's staged templates. The packaged templates, filled with this run's values, gave the same hash until the
+2026-09-27 template edits (#385), a local check whose value `tests/test_landscape_sweep_harness.py` keeps as
+`PROMPTS_SHA256_20260926`. The packaged tools produced the returns, lanes, layers, usage records,
 manifest and source reviews. `tavily-leads.json`, the lead-vetting addendum `tavily-leads-vetting.json` and the
 attempts' compact records were written for this record, and each states its method.