Testing: diagnose local full-suite failures and missing inline diagnostics

The original local Buddy suite for the issue #65 cold-build setup branch completed with exit 1 at unchanged source HEAD 02dcbc082dd70183362983d4cbebc97eee9f8535, with eight uncommitted source files pinned throughout the run. Pinned Bun 1.4.2 executable hash and every source pin remained unchanged. The run was not restarted.

Terminal verdict: 1,431 passed, 27 skipped, 82 failed, 474,778 assertions, 1,540 tests across 194 files, 2,770.67 seconds reported by Bun. Full terminal log retained in evidence 410: 428,106 bytes, SHA-256 daf568d9b4373f1b4e0d589c968081df09e9c1257b051b704aa6401257c88995.

The original log prints four inline failure records, then lists 82 failures in its final summary. The remaining 78 summary entries have no corresponding inline failure/header/stack in these captured bytes. They include execution bootstrap/protocol/HTTP/lifetime/production, subprocess guards, agriculture export, operator provisioning and host-capability cases. Many report sub-10 ms durations. This is an observation about the retained log, not proof the cases did not execute or proof of a reporter/runtime defect.

The four inline failures are the inspection CLI fault demonstration and three incident content deadline fixtures. The latter show rejection during setup because retention was already unavailable, before the intended final release boundary. Those setup failures need their own bounded investigation; no security-policy relaxation is authorized by this report.

Host disk exhaustion was observed during this run and regenerable caches were cleared. This log contains no ENOSPC/SQLITE_FULL record establishing that as the cause. Current Linux CI for the separately published browser diagnostics head passed the full suite; that distinct success does not replace this local failure or explain it.

Acceptance:

• Retain the complete failed run, pinned executable/source identity and every summary entry.
• Establish why inline diagnostics for the 78 summary-only failures are absent, using bounded native process/stdio observations at the actual runner boundary.
• Diagnose the inspection CLI and content setup failures separately unless causal evidence joins them.
• Qualify any correction against its original failure boundary and current-head full CI. Do not erase failures with warm reruns, arbitrary deadline increases or weakened authority assertions.

The retained terminal-failure-groups.json distinguishes the exact four inline entries from all 82 terminal entries. No physical vehicle or browser acceptance is implied.

Separate retained timing observation from VTOL recovery PR #80, exact source 7ebeca5:

• The focused combined run had 244 passes, 1 failure, 13,149 assertions across 17 files. Existing unchanged VTOL land withdraws independent output authority for rcOverride exceeded Bun's unchanged 5,000 ms callback limit at 5,112.03 ms. All 39 new recovery cases passed. Original output is retained with SHA-256 378f127e8ca612080f32589169e1ca09b5d4907bb02a69a537db4da1b076eab2.
• One instrumented exact-case diagnostic on identical source and Bun bytes passed 11 assertions at 841.12 ms (1.4326 s process wall). It recorded per-50 ms CPU/RSS, memory and swap. Swap was about 5,489 MiB before and 5,481 MiB after. These observations do not prove the original scheduling cause.
• No timeout was increased, no assertion weakened, no warm retry-until-green or full-suite rerun occurred. The combined run remains a failure. New exact-head full Linux qualification is pending independently.

This does not explain the original 78 summary-only failures or establish a runner/stdio fix. The issue and its original acceptance criteria remain open; the historical before-fix 241-test pass is not substituted for current qualification.

A separate qualification correction is committed as Chris without coauthors in 04e25d0, updating draft PR #80.

The original execution-browser assertion can reject a valid private-pane clearing gap: deliberately ending the genuine native workspace stream after admission, while holding the replacement GET before native policy, reproduces the same undefined selected-source assertion. Retained diagnostics show native admission 201/confirmed, producer finished and DOM reconnecting. The driver now requires actual clearing, releases the genuine read, waits for fresh authorization and the original selected source, and checks every bounded observed selection for an unintended switch. Final history is exactly original, absent, original; unsent note preservation is still exact.

Final focused qualification passes 27 tests / 185 assertions plus native lint/app/runtime types. Both independent reviews are clean. Existing test commands and deadlines are unchanged. Only two artifact paths were added to the original full CI upload: bounded closed-scalar failure-diagnostics.json and the bounded selection history, so another failure need not rely only on inline logs.

New exact-head full CI is pending. Original CI529 remains retained with its browser failure and sole original terminal log collector. The controlled reproduction proves an assertion timing defect under native renewal; it does not establish the historical triggering event in that CI run. It also does not diagnose the original 78 summary-only failures or the retained local rcOverride timeout. This issue and its original acceptance criteria remain open. No rerun-until-green, deadline increase or weakened authorization assertion.