[Evidence Gap] Harden-runner custom detection initialization fails while MV3 gate stays green

[seonghobae]

Historical evidence defect

On PR #70 predecessor exact head b35e97c08b37a2e6f21cbc5e0403277c99f7931e, Manifest V3 Compatibility run 32401838435, job 96531623967, checked out that exact head and completed successfully while the external step-security/harden-runner@bf7454d06d71f1098171f2acdf0cd4708d7b5920 initialization path logged a timeout and invalid authorization header format while fetching custom detection rules. That run therefore remained valid browser-fixture evidence but was not valid proof that custom detection rules were active.

Fresh exact-head recheck

PR #70 has since advanced non-destructively to exact head d1f0dfc611c07b847477bfe4671eba0a4d83d065 on unchanged prerequisite #65 head 6bb9474840db538b37299541f6031a0aee8393b7.

The exact-head Manifest V3 run 32436889661, job 96639798522, checks out d1f0dfc611c07b847477bfe4671eba0a4d83d065 and is successful. Its harden-runner/agent logs now show:

• global feature flag EnableCustomDetectionRules:true;
• Fetching custom detection rules;
• Custom detection rules evaluator initialized;
• Custom detection rules enabled; and
• final harden-runner status Initialized.

The predecessor timeout/invalid-authorization messages do not recur in this exact-head run. No OriginWeave product or workflow mutation was used to manufacture that recovery.

Disposition

The observed external-integration failure is no longer reproducible on the current exact PR head, so this issue is closed as a recovered transient evidence condition rather than left as a false current blocker. The predecessor run remains historical evidence only and is not promoted to current proof.

If a future exact-head run again shows custom-detection initialization failure while remaining green, reopen or create the current owner incident with that exact run/job/checkout evidence. Protected-main AGENTS.md still forbids the scheduled OriginWeave writer from altering workflows, credentials, or security gates.

[seonghobae]

Fresh exact-head evidence: PR #70 head d1f0dfc611c07b847477bfe4671eba0a4d83d065 run 32436889661 completed successfully. Its pinned-browser output records Agent Task 3/3, trial_pass_rate: 1.0, post_condition: true, profile_cleaned: true, and url_unchanged: true for all three trials. The same run's harden-runner log reports Custom detection rules enabled, Custom detection rules evaluator initialized, and Armour engaged; the prior invalid-authorization/timeout evidence on the older head was not reproduced in this run. This supersedes only the prior-run observation; the issue remains open for the authorized workflow/security owner to confirm the protected-main policy and whether custom-detection initialization must be a separately surfaced required check.

[seonghobae]

Recurrence evidence on a different exact head: PR #148 head b1cdf304666c79aa7a95a3983a54920a2f249e76, Manifest V3 Compatibility run 32644875691 (exact checkout of that head), completed success and the pinned Chrome/ChromeDriver 150.0.7871.129 browser lanes passed, including browser-crash 3/3, forced-close 3/3, and Agent Task 3/3. In the same job, step-security/harden-runner@bf7454d06d71f1098171f2acdf0cd4708d7b5920 again logged an initial timeout connecting to https://agent.api.stepsecurity.io/v1 and then Failed to initialize detection rules, continuing with eBPF attachment ... API error status 401: {"error":"invalid authorization header format"} while the overall workflow remained green.

This materially reverses the intervening non-reproduction recorded above and confirms the condition is intermittent/recurrent at the workflow/provider authentication boundary rather than a one-off PR #70 observation. The browser compatibility evidence remains valid for its own lane, but custom-detection initialization must remain not proven for this run. Per protected-main authority, no OriginWeave product code or workflow/secret mutation is being used to mask it; the authorized workflow/security owner should treat recurrence as evidence for a durable initialization/evidence contract rather than repeated blind retry.