Bug: TTS cuts off short responses mid-word — MAX_SPEAK_LENGTH=500 in config

[chansg]

Problem

Aria said I'm sorry, Chan, but I couldn' — cut off mid-word on a 39-char response.
config.py still has MAX_SPEAK_LENGTH = 500.
speaker.py may still be reading this value and truncating before synthesis.

Fix

Audit speaker.py for any reference to MAX_SPEAK_LENGTH or config.MAX_SPEAK_LENGTH.
Remove all hard character limits from the TTS pipeline.
Only sentence-based chunking should control how text is split.

Acceptance Criteria

• Short responses (under 100 chars) spoken in full without cut-off
• Long responses spoken via sentence chunking without truncation
• MAX_SPEAK_LENGTH removed or explicitly disabled in config.py

[chansg]

Root cause clarified — not TTS truncation

After investigating, the source-side TTS pipeline is not truncating responses:

• `MAX_SPEAK_LENGTH` is defined in `config.py` but never imported anywhere — not in `speaker.py`, not in `brain.py`, not in `main.py`. It's a dead constant.
• `speaker.py` uses only sentence-based chunking via `_split_into_sentences()`. Confirmed via test: `"I'm sorry, Chan, but I couldn't find the weather."` (49 chars) survives the pipeline as 1 chunk, fully intact.
• No `text[:N]`, `.truncate()`, or character-limit logic anywhere in the speech path.

The actual root cause: Gemini returned a truncated response. The 39-char cut-off (`I'm sorry, Chan, but I couldn'`) is what came back from the Gemini API itself — likely a partial stream from a slow/aborted call, content-filter mid-stream, or a timeout during generation.

Mitigation in this PR

Added `_ensure_complete_sentence()` guard in `brain.py` that runs on every Tier 2 response (Gemini vision + Gemini reason_with_context + Ollama fallback). If the response ends with no sentence-ending punctuation, a period is appended so Aria sounds intentional rather than glitchy. This sands off the rough edge but does not fix the upstream Gemini truncation.

Hardening also in this PR

• `MAX_SPEAK_LENGTH = 9999` (effectively disabled) with deprecation comment in `config.py` and `config.example.py` to prevent future confusion.

Future work

The real fix for Gemini partial responses needs investigation:

• Check Gemini call timeout (currently 20s in `vision_analyzer.py`)
• Check `max_output_tokens=400` — could be hitting the limit mid-sentence
• Consider adding finish_reason logging from the response to detect truncation upstream