Replace SSSpeechSynthesizer with testable MUDSpeechQueue skeleton

[powerje]

Summary

Phase 2 foundation for the VoiceOver speech-pipeline rebuild. Replaces the 2013 SSSpeechSynthesizer Obj-C class from Vendored/SSAccessibility with three new Swift types under src/Mudrammer/Accessibility/:

• SpeechLine — struct carrying text, priority (iOS 17+ UIAccessibilityPriority), and origin (.server / .prompt / .trigger / .system). Only text is load-bearing today; the other fields exist so follow-up work can populate them without reshaping the type.
• Announcer — protocol with a single announce(_:completion:) method that abstracts the "post to VoiceOver" step. Production implementation UIAccessibilityAnnouncer posts an NSAttributedString with accessibilitySpeechAnnouncementPriority and accessibilitySpeechQueueAnnouncement attributes, observes UIAccessibility.announcementDidFinishNotification for completion, with a 16s timeout fallback matching the 2013 implementation's behavior.
• MUDSpeechQueue — @objc-exposed class that holds an internal FIFO buffer and drains one line at a time through the injected Announcer. Exposes enqueueLineForSpeaking: and stopSpeaking for one-line drop-in replacement at the SSMudView call site.

SSMudView.m is updated to construct [MUDSpeechQueue new] instead of [SSSpeechSynthesizer new] and no longer imports SSAccessibility. The unused continueSpeaking method (declared in SSMudView.h but only self-referentially called from SSMudView.m) is removed — a grep across the tree found no external callers.

The 2013 Vendored/SSAccessibility package is left in place for now; it has no remaining call sites in app code and will be deleted in a follow-up cleanup step once the migration settles.

Why this matters

Research in the prior session surfaced that the #1 real-world complaint from blind MUDRammer users in the 2025 AppleVis thread is prompt-jumping — the 2013 SSSpeechSynthesizer posts at default priority, which interrupts current speech mid-sentence whenever a new line arrives. Posting at .low priority (queued behind existing speech, yielding to system announcements) is the direct fix, and iOS 17's UIAccessibilityPriority is the first-party primitive that makes it trivial. The Announcer seam is what lets future work unit-test the queue without relying on a real VoiceOver subsystem that can't be observed from tests.

Deliberately out of scope

This is a skeleton. Each item below is a separate follow-up PR:

• Pre-enqueue filter pipeline (blank / whitespace / ASCII-art)
• Prompt detection and prompt filter (the actual fix for the prompt-jumping complaint)
• User-configurable interrupt-vs-queue mode
• Observable drop accounting / delegate
• User-typing behavior hook
• Per-world speech configuration persistence
• Deletion of Vendored/SSAccessibility
• Priority-semantics spike on iOS 26 under MUD-rate bursts
• SSANSIEngine preprocessing test coverage

Test plan

• xcodebuild test -workspace Wammer.xcworkspace -scheme MRTests -destination 'platform=iOS Simulator,name=iPhone 17 Pro' -only-testing:MRTests/SpeechQueueTests — expect 6 passing tests in ~0.2s
• xcodebuild build -workspace Wammer.xcworkspace -scheme Wammer -destination 'platform=iOS Simulator,name=iPhone 17 Pro' — expect clean build with no new warnings
• Open Wammer.xcworkspace in Xcode, build the Wammer target, verify the audit navigator shows no regressions
• On a physical iPhone with VoiceOver enabled, connect to a test MUD, confirm incoming text is still spoken and stopSpeaking fires on disconnect/clear/user input/radial-movement (the four [self stopSpeaking] call sites in SSMudView.m)

Known pre-existing issues not addressed here

The MRTests scheme has a pre-existing flake in GMCPHandlerTests.reset where an auto-layout call runs on a background thread and aborts the test process. Verified present on trunk independently of these changes — not introduced by this PR.

Unit test coverage

src/MRTests/SpeechQueueTests.swift uses a RecordingAnnouncer test double that captures every announce call and exposes a manual completeOldest operation, so each test drives the queue's drain loop deterministically:

• emptyTextIsDroppedAtEnqueue
• nonEmptyLineReachesAnnouncer
• objcEnqueueUsesLowPriorityAndServerOrigin
• linesDrainInOrderAsCompletionsFire
• stopSpeakingClearsPendingLinesNotYetAnnounced
• enqueueAfterStopSpeakingWorksNormally