Development and orchestration record (separate from reader outcomes)

Before any reader exposure, all three inputs passed14 material checks. The
portable check command was adapted to compare shared files with a frozen hash
inventory rather than require the predecessor directory on another machine.
The original report and the portable rerun both passed. Shared reader tools
were unchanged by this portability edit. All changes preceded STUDY-FREEZE.

R03 dispatch encountered the collaboration service's agent-thread limit.
Two root attempts and one helper attempt were refused before any R03 source
exposure. Once a preceding reader turn finished, R03 launched with fresh
conversation context under the original prompt. No failed reader was replaced.
The three launch blocks therefore overlap in calendar time and are scheduling
groups, not independent environmental replicates. A CLI help command was read
while investigating capacity; no CLI reader or alternate model was run.

The plotting library was initially unavailable in the bundled Python runtime.
A sandboxed installation could not resolve its package host. An approved retry
installed matplotlib3.11.2 and its dependencies inside this job's build folder.
This concerns figure generation only. It did not alter reader inputs, code,
responses, the scoring key or the study protocol. Build dependencies are not
part of the shared29,375,470-byte reader resource count.

Reader tool-output truncations, helper errors and any other observations are
preserved in each reader's disclosure. A short display is not treated as a
failed input when the reader subsequently obtains the complete selected facts.

The browser preview request for a local file URL was refused by the browser's
URL policy. No alternate browser, local proxy/server, content injection or
security workaround was attempted. Publication UI checks use an inert Node
VM/DOM fixture plus static inspection; they are not browser-engine tests.
Actual browser rendering, mobile devices and live WordPress remain unverified.

Matplotlib's first figure build could not save its font cache in the user's
profile. The figure still generated. Subsequent builds put the cache in this
job's build-cache directory. No reader artifact was affected.

Final publication portability verification extracted the complete downloadable
kit to a new directory. All three recoveries, material parity, scoring and reply
encoding passed. The independent audit then falsely flagged all nine historical
prompts because it regenerated their absolute path text from the new directory.
The audit was corrected to use the recorded historical root only for string
comparison; all actual file access remains relative to the extracted kit.
No prompt, input, key, reader output or score was changed. The initial failure
summary is retained in analysis/publication-portability-initial-failure.json.
