feat: Web Speech API output via RS-423 serial port (*FX3,1)

[mattgodbolt-molty]

Closes #508.

What this does

BBC programs use *FX3,1 (or *FX3,3) to route VDU text to the RS-423 serial port. On real hardware this fed a Votrax Type 'N Talk speech synthesiser. This PR routes it instead to the browser's window.speechSynthesis Web Speech API — not period-correct, but much cleaner voice and zero extra dependencies.

Enable it via the new Text-to-Speech toggle in Settings, or with the ?speechOutput URL parameter.

Root cause: CTS was blocking all RS-423 output

The ACIA was setting CTS high ("not connected") whenever RS-423 was selected. CTS high inhibits the TDRE bit in the status register, so the OS's interrupt-driven output buffer would fill up and silently drop bytes — without hanging, but without transmitting either. Fixed: when an rs423Handler is present the ACIA now sets CTS low (clear to send), which also correctly fixes the TouchScreen command path.

Changes

src/speech-output.js (new)

• RS-423 handler implementing onTransmit(byte) and tryReceive()
• VDU-aware state machine: skips parameter bytes for multi-byte VDU sequences (MODE, cursor, colour, palette, graphics window etc.), buffers printable ASCII, speaks on CR/LF or at 200 chars
• Cancels in-progress utterance before each new speak so output stays current
• Graceful no-op if speechSynthesis is unavailable (Node/headless)

src/acia.js

• selectRs423(): set CTS low when a handler is present; high when nothing is connected
• New setRs423Handler(handler) for runtime wiring (construction-time handler was always undefined since the TouchScreen is created during the first hard reset, not at construction)

src/main.js

• setupRs423Handler(): composite handler that forwards onTransmit to both TouchScreen and SpeechOutput, and tryReceive to the TouchScreen — called after processor.initialise()
• Initialise speechOutput.enabled from ?speechOutput URL param or localStorage
• Handle changed.speechOutput in settings onClose; persist to localStorage

src/config.js + index.html

• setSpeechOutput() and checkbox click handler
• Text-to-Speech toggle in the Settings panel

Verification

Tested end-to-end via MachineSession:

*FX3,1
PRINT "BBC Micro speech test"

→ spoken: > PRINT "BBC MICRO SPEECH TEST", then BBC MICRO SPEECH TEST

Note: with *FX3,1 the BASIC prompt and echoed input are also spoken (mirrored to serial alongside screen). Software specifically designed for blind use would use *FX3,3 (VDU disabled) to suppress this.

12 new unit tests cover VDU stripping, CR/LF flushing, enable/disable, and the 200-char safety flush.

(I'm Molty, an AI assistant acting on behalf of @mattgodbolt)