feat(ipc)!: remove the cache-key mailboxes, control socket stage 5 (for the release after 3.8.1) (#773)

* feat(ipc)!: remove the cache-key mailboxes (control socket stage 5)

The control socket is now the only way the web interface sends the display a
command. The display stops reading display_on_demand_request and
plugin_error_clear_request, and the web interface stops writing them.

- Display: no mailbox poll (MailboxWatch, the 1 s / 0.25 s cadence,
  _consume_on_demand_request, the deprecation log) and no persisted
  display_on_demand_processed_id guard; the error publisher reads no clear
  request. CacheManager.file_signature and MailboxWatch are removed.
- A write to either retired key is dropped by CacheManager.save_cache and
  logged once per writer, naming the plugin from the call stack (or the
  request's plugin_id), with the API to move to.
- Web: on-demand start with no display listening starts the service (when
  start_service) and sends the request again once the socket answers (45 s,
  10 s for a running service without a socket yet); every other failure is
  a 503 (400 for invalid_args). Stop answers 503 when no display listens,
  unless stop_service. errors/clear answers 503 with a reason-specific
  message instead of writing a request; clear_pending is always false.
  src.ipc.client.should_fall_back is replaced by display_not_listening.
- Kept: display_current_state, display_on_demand_state,
  plugin_runtime_snapshot and the heartbeat (read whenever the socket cannot
  answer), and display_on_demand_config (the display's resume record).

Tests: mailbox-only tests removed (test_on_demand_mailbox.py, the mailbox
cadence, file_signature and MailboxWatch tests); tests that injected
requests through the mailbox now use the socket queue or a plugin's
in-process request. The run-loop harness sends on-demand requests over its
fake control socket, so four golden traces change: on-demand starts and
stops land at the request instant instead of the next 0.25 s mailbox look
(one frame fewer on the screen they end), and in vegas.json within one
frame instead of 263 ms, which shifts the later 1 s-throttled WiFi-notice
check by under a second.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: the on-demand routes' other callers ack over the socket, not the mailbox

test_api_v3_bool_coercion, test_api_v3_lazy_plugin_discovery and
test_web_api's stop test checked the mailbox write; they now get an ack
from a mocked control socket. The stop route touches no manager any more,
so it leaves the removed-catch-all sample. A request that was sent is
never 'not listening' (pins the sent check the mutation run let through).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(web): a start that waits for the display answers 202 and is delivered in the background

The start route held a request open for up to 45 s while a cold-started
display loaded its plugins; the MQTT bridge (15 s timeout) and browsers
reported a failure for a request that was then delivered.

Now, when no display is listening, the route starts the service if asked
and answers 202 with status "starting" at once. A single worker in the web
process (web_interface/on_demand_dispatch.py) sends the request until the
display acknowledges it or the wait runs out (45 s cold start, 10 s for a
running service without a socket yet). A newer start supersedes the
pending one; a stop cancels it (and succeeds, with cancelled_request_id,
even with no display listening). The outcome is reported by
/display/on-demand/status (source "web": starting, or error with
start-timeout or the socket's reason, until the display publishes
something newer) and by /display/current-status as on_demand_pending.

Callers: the web UI's on-demand modal and "Preview on display" treat
"starting" as taken (an info toast); the MQTT bridge already treats any
non-error 2xx as success (now pinned by a test).

Tests: the dispatcher (ack, retry then ack, start-timeout, other failures,
superseded, an in-flight ack for a superseded start, stop while pending,
a per-start wait, outcome lifetime); the routes (202, status routes while
pending and after a timeout, a later display state replacing the failure,
stop while pending, a new start superseding); a JS suite for app.js.
Mutation check: 20 mutants on the worker, the routes and app.js, 20 killed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(web): a cancelled on-demand start never reaches the display after its replacement

CI caught a race in test_a_new_start_supersedes_the_pending_one: the
dispatcher's worker could read the old start, the route cancel it and send
the new one, and the worker's send of the old one land after it. cancel()
now waits out a send already in flight (the worker holds a send lock while
it reads the pending start and sends it), so whatever the caller sends next
lands after it. Pinned by test_a_cancel_waits_for_a_send_in_flight, which
fails without the wait.

The Linux-only TestRealSocket test for a display that went away now
expects the 202 and the start-timeout that follows, as the route answers
since the background dispatcher.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(web): a delivered on-demand start reads as starting until the display acts on it

On ledpi (three cold starts) the display acknowledged the start as its
socket opened, then took ~5 s to act on it while Vegas built its first
strip; the status routes meanwhile showed the display's own idle state, so
a UI polling every 700 ms flashed idle.

The display's on-demand state now names the request it answers
(request_id). A delivered start keeps reading as status "starting" with
delivered: true, in /display/on-demand/status and as on_demand_pending in
/display/current-status, until the display publishes state for that
request id (a display without the field: any state newer than the
delivery), for at most DELIVERED_SHOWN_SECONDS (30 s). The display's
startup state, which can be published after the acknowledgement, names no
request and does not end it.

Tests: stays starting against the startup idle state (no id, an older id);
the matching active state and the matching error take over; an older
display's newer state takes over; the 30 s cap; the display's state names
its request. Mutation check: 11 mutants, 11 killed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(display): on_demand_request_id has a class default for controllers built without __init__

_on_demand_state now publishes it, and test_state_stream_readers builds
controllers with __new__.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: the case-spelling on-demand test patches _send_on_demand now the mailbox is gone

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

* docs: blank line before 3.8.7 heading

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

src/ipc/client.py

@@ -5,11 +5,11 @@ a refused or timed-out connection, a reply that breaks the contract, or an
 error the display returned -- raises :class:`ControlError` with a short
 ``reason``. Nothing here blocks for longer than ``timeout`` in total.
 
-Whether the caller may then write the file mailbox instead is
-:func:`should_fall_back`: only when the display never took the request (it
-could not be reached, or it is too old to know the command). A display that
-took the request and then failed, refused or went quiet is answered as
-that, not posted a second time through the mailbox.
+There is no other way to reach the display: the file mailboxes the web
+interface used to fall back to are gone. :func:`display_not_listening` tells
+a caller when no display is listening yet (it is stopped, or still
+starting), the one case where sending the same request again later, once a
+display is up, can work.
 """
 
 from __future__ import annotations
@@ -43,8 +43,7 @@ from src.ipc.contract import (
 
 #: Total budget for one request: connect, send and the reply. The display
 #: answers from a thread that does no rendering, normally within a few
-#: milliseconds; this only bounds a wedged one. The web route then falls back
-#: to the mailbox, so a timeout costs this much latency and nothing else.
+#: milliseconds; this only bounds a wedged one.
 DEFAULT_TIMEOUT_SECONDS = 1.0
 
 
@@ -74,28 +73,25 @@ class ControlError(Exception):
 
 #: Answers from a display that read the request but does not speak it: one
 #: older than the command (an upgrade in progress) or the protocol version.
-#: It did nothing, so the mailbox is the way to reach it.
 UPGRADE_REASONS = frozenset({ErrorCode.UNKNOWN_COMMAND, ErrorCode.UNSUPPORTED_VERSION})
 
+#: Transport reasons that mean nothing is listening at the socket: no socket
+#: file (the display is stopped, or has not reached its run loop), or a file
+#: nobody accepts on (a stale socket) or that this user may not open.
+NOT_LISTENING_REASONS = frozenset({'no_socket', 'refused'})
 
-def should_fall_back(error: BaseException) -> bool:
-    """May the caller write the file mailbox after ``error``?
 
-    Yes when the display never took the request: there is no socket (the
-    display is stopped, predates the socket, or it is switched off), the
-    connection was refused or timed out, the display turned the connection
-    away before reading it, or it is too old to know the command
-    (:data:`UPGRADE_REASONS`). Also for an error that is not a
-    :class:`ControlError` (a bug in the client), as before.
-
-    No once the display had the request: a ``busy`` queue, ``invalid_args``,
-    an ``internal`` error, or a timeout or hang-up after the request was
-    sent. The display may have applied it, or would refuse it from the
-    mailbox too, so a second copy there only hides the failure.
+def display_not_listening(error: BaseException) -> bool:
+    """True when ``error`` says no display took the request because none is
+    listening: it never reached one (``sent`` is False) and the reason is in
+    :data:`NOT_LISTENING_REASONS`. A display that is started, or finishes
+    starting, may take the same request later. False for everything else:
+    a display that had the request and failed it, one too old to know the
+    command, a client that cannot use the socket at all (``disabled``,
+    ``unsupported``), and an exception that is not a :class:`ControlError`.
     """
-    if not isinstance(error, ControlError):
-        return True
-    return not error.sent or error.reason in UPGRADE_REASONS
+    return (isinstance(error, ControlError) and not error.sent
+            and error.reason in NOT_LISTENING_REASONS)
 
 
 def request(cmd: str, args: Optional[Mapping[str, Any]] = None, *,
@@ -218,8 +214,8 @@ def on_demand_start(request_id: str, plugin_id: Optional[str], mode: Optional[st
                     paths: Optional[Sequence[str]] = None) -> Dict[str, Any]:
     """Ask the display to show a plugin now. Returns the ack; raises :class:`ControlError`.
 
-    ``request_id`` doubles as the on-demand request id, so a request that a
-    timed-out caller then also writes to the mailbox is processed only once.
+    ``request_id`` doubles as the on-demand request id, so a request sent
+    twice with the same id is processed only once.
     """
     args = {'plugin_id': plugin_id, 'mode': mode, 'duration': duration, 'pinned': pinned}
     return request(Command.ON_DEMAND_START, args, request_id=request_id,
@@ -281,8 +277,7 @@ def errors_clear(request_id: str, cutoff: float, *,
 
     Returns :class:`~src.ipc.contract.ErrorsClearResult` once it is done.
     Raises :class:`ControlError`: ``unknown_command`` from a display older
-    than the command, which still reads the ``plugin_error_clear_request``
-    mailbox.
+    than the command.
     """
     return request(Command.ERRORS_CLEAR, {'cutoff': cutoff}, request_id=request_id,
                    timeout=timeout, paths=paths)