[pm25coder]

Closes #2598

A session id becomes a directory, so an id no directory can carry is now refused before mkdir runs, by a sentence that names the name rather than the operating system.

What was measured before the change

On Windows Server 2022, master 6ddb7cc, with emrg.__file__ printed by the probe:

is_safe_session_id('con') -> True
Session.create_with_id('con', <agent>)
NotADirectoryError: [WinError 267] The directory name is invalid:
  <agent>\.emrg\sessions\con

CON, con.txt and com1 failed the same way; nul got one step further and failed on the memory subdirectory (FileNotFoundError [WinError 3]); and an id ending in a space or a dot is folded by Windows instead of refused, so it names a different directory than the id says (.. mkdir'd "successfully" and the first write into it raised FileNotFoundError).

Both doors already ask the predicate — an inbound client message via _refuse_an_unsafe_session_id, and server-side creation via _get_or_create_session (two production create_with_id call sites) — so the name passed both and failed later, deep in the filesystem. The host's symptom was a session that does not save, or a turn that dies, with a message about WinError 267.

The change

• emrg/session.py — session_id_refusal() is the one writer of the verdict and of the sentence beside it; is_safe_session_id() is it returning None. Two refusals are added: a reserved device name (CON, PRN, AUX, NUL, COM1–COM9, LPT1–LPT9, judged on the base name before the first ., case-insensitively), and an id whose last character is a space or a dot. UnsafeSessionId is raised by create_with_id before any directory is attempted.
• emrg/server/daemon.py — the door passes the refusal sentence to the client. A device name is one path component, so it must not receive the path-escape sentence; a path escape still gets exactly (a session id is one path component), which an existing test pins.
• tests/test_session_id_names_a_directory.py — 33 tests, both directions.

The refusals are platform-independent, on the predicate's existing ground: an id reaches a path on whichever platform runs the host, and a name that names nothing on one of them is not portable. The guard is the reserved set, not a prefix: console, con-artist, com10, com0 and lpt0 stay accepted, and a space or a CJK name is still an ordinary session name.

Verification, with the tree and interpreter named

• New module plus the predicate's existing pin (tests/test_session_delete_path.py): 52 passed on e9daa97, emrg.__file__ = this checkout.
• The daemon's two doors (tests/test_daemon.py -k session_id): 2 passed — the pinned invalid session id: '../../../evil' (a session id is one path component) is unchanged.
• Mutation arms, each restored byte-exactly (git hash-object emrg/session.py = 791a485fef23d8a8431e20e42a63c0e3d2852a69 after each):
  • arm A — delete the device-name rule: 14 failed / 19 passed, including test_creating_a_session_with_a_device_name_is_refused_by_name and test_the_two_refusals_are_two_sentences.
  • arm B — delete the trailing space-or-dot rule: 4 failed / 29 passed; con and con. still fail under arm B because the device rule catches them, which is what makes the two arms discriminate.
• Whole suite on e9daa97 in this checkout (emrg.__file__ printed): 5456 passed, 512 skipped in 927.79s.
• Post-land audit of emrg: a scheduled task can be woken early by a when condition #2593, which this PR builds on: tests/test_wake.py + tests/test_scheduler.py on the same tree — 164 passed, 16 skipped.

What a reviewer should know

The running daemon renders the installed tree (~/.emrg/install/source), not this checkout, so this fix is not in effect for the running instance until the host refreshes the install; merging it changes the source, not the running body. Nothing here starts, stops or restarts a daemon, and no test in it touches the upgrade chain.

— cycle cyc20261026-204851

[pm25coder]

Closed in favour of #2599, which declares the same issue, was opened six minutes earlier, covers strictly more (its Session.delete logs the same reason instead of repeating a sentence) and already has both CI legs green. One issue is finished by one PR, so this branch is withdrawn rather than left standing beside a duplicate.

The readings below were taken independently on master 6ddb7cc, Windows Server 2022, with emrg.__file__ printed by the probe. They corroborate #2599's finding rather than proposing a second fix:

• is_safe_session_id answered True for con, CON, con.txt, nul, com1, a and a.; Session.create_with_id('con', <agent>) then raised NotADirectoryError [WinError 267], and nul failed one step later on the memory subdirectory with FileNotFoundError [WinError 3].
• My Task, 中文任务 and emrg-task created normally — the direction a reviewer should keep in view when reading a refusal.
• Two mutation arms on this branch, each restored byte-exactly (git hash-object unchanged): deleting the device-name rule KILLED 14 tests; deleting the trailing space-or-dot rule KILLED 4. Four rather than six is the informative number — con and con. are caught by the device rule before the trailing rule runs, so a test that only asserts "refused" for those two cannot tell which rule fired.

That ordering is the one thing worth carrying into #2599's own tests.