the client's daemon-start diagnostics open their subject without asking its kind — a FIFO at either path stops every start attempt #2109

The same class as #2101 / #2103 / #2105 / #2107 — an open is not a read. On a FIFO the open waits for a peer and raises nothing, so no except OSError around the call can fire: the call simply never returns.

The sites (emrg/client/daemon_manager.py, master 60541e39)

The subjects are logs_dir()/emrgd-start.err and logs_dir()/emrgd.log (_start_stderr_path, _log_path).

Measured on master 60541e39

One child process per arm under an 8 s cap, with a regular-file control for each subject, so the axis read is the kind of the path and not its absence:

truncate   regular    -> RETURNED BufferedWriter
truncate   fifo       -> DID NOT RETURN within 8s
truncate   directory  -> RETURNED NoneType
read_start regular    -> RETURNED ''
read_start fifo       -> DID NOT RETURN within 8s
read_start directory  -> RETURNED None
log_tail   regular    -> RETURNED ''
log_tail   fifo       -> DID NOT RETURN within 8s
log_tail   directory  -> RETURNED ''

The directory column is what fixes the design: the kind one over already produces each site's answer (None for the writer, which the caller reads as "fall back to DEVNULL"; None for the stderr reader, which the report already names "could not be read"; "" for the log tail, "nothing new"). So the guard adds no new outcome — only that the existing one arrives.

The fix

Ask the subject's kind before opening it, through the tool layer's path-taking non_regular_kind — imported rather than restated, so the boundary stays one fact with one spelling — and take the answer each site already gives. _read_log_tail's exists() is replaced rather than kept as a second opinion.

Acceptance

• each site returns its own existing answer for a non-regular subject instead of blocking;
• a FIFO-free arm per site, on a directory, that discriminates: both trees answer the same value there, so the arm asserts the kind is named in the log with the path — which the base cannot produce, and which is the arm the Windows leg runs;
• the regular-file controls are unchanged, and the FIFO arms are POSIX-gated (os.mkfifo) and bounded by tests/bounded_read.within;
• uv run pytest tests/ green, with check-undefined-names, check-doc-count, check-citation-resolves, check_nonlocal and check_unbound_reads at rc 0.

Out of scope, named rather than silently under-reported: the GUI's JS client carries the same three shapes (emrg/gui/daemon_client.js:247, :262, :278 — readFileSync / openSync(file, "w") / readFileSync). This issue does not cover it.