fix(tts): strip every reasoning-tag variant before speech, not just <think>

[Drexuxux]

What

The two TTS text cleaners only removed <think>...</think>:

• strip_nonspoken_blocks (batch normalizer — every auto-TTS / voice / tool path via prepare_spoken_text)
• SentenceChunker (the streamed speaker pipeline + speak-stream WebSocket)

Every other reasoning scrubber in the codebase — agent/think_scrubber.py (_OPEN_TAG_NAMES) and strip_think_blocks — covers five variants: <think>, `")
before: 'The answer is 42.' ← spoken
after : 'The answer is 42.'

SentenceChunker, "hidden" split across deltas
before: reasoning spoken
after : held until close, stripped

The raw deltas genuinely reach TTS unscrubbed: the gateway tees the same raw `text` to both the display consumer and the streaming-TTS consumer (`gateway/run.py`), so the TTS side must scrub on its own.

## Fix

Share one canonical tag set (kept in sync with `think_scrubber._OPEN_TAG_NAMES`) across both cleaners:

- `strip_nonspoken_blocks` strips all five variants, closed and unterminated, case-insensitively.
- `SentenceChunker` strips them across deltas and holds a delta back on **any** unclosed reasoning tag — a new `has_unclosed_reasoning_tag` probe — not just `<think>`.

No behaviour change for text without reasoning tags; the `<think>` path is unchanged.

## Tests

`tests/tools/test_tts_text_normalize.py` and `tests/tools/test_tts_streaming.py`:

- every variant (`<thinking>`/`<reasoning>`/`<thought>`/`<REASONING_SCRATCHPAD>`), closed and unterminated, is never spoken — parametrized
- case-insensitive (`<THINKING>`, `<Reasoning>`)
- `SentenceChunker` strips non-`<think>` tags split across deltas and holds on an unclosed tag
- the `<think>` path still passes (unchanged)

scripts/run_tests.sh tests/tools/test_tts_text_normalize.py tests/tools/test_tts_streaming.py

with the fix

=== Summary: 13 passed (text_normalize) / all SentenceChunker green ===

without the fix — genuine behavioural failures (not import errors)

test_tts_text_normalize.py 9 failed ( params still pass — already covered)
test_tts_streaming.py 5 failed

Broad TTS sweep — `tests/tools/test_tts_*`, `tests/gateway/test_*tts*`, speak-stream: **81 passed, 1 failed**, the one failure being a pre-existing flaky wall-clock timing test (`test_hybrid_prefetch_fires_http_immediately`) that also fails on a clean checkout. Zero new failures.

[teknium1]

Thanks for covering the non-variants; the underlying current-main gap is real: `tools/tts_text_normalize.py:215` and `tools/tts_streaming.py:86-107` only handle.

Problems

• The amended SentenceChunker.flush() at PR tools/tts_streaming.py:137 strips only closed pairs. After feed("secret") holds the buffer, flush() returns the still-open raw tail. This reaches synthesis because gateway/streaming_tts_consumer.py:182 enqueues every flushed clause.
• hermes_cli/web_server.py:4644 independently recognizes only lowercase `` before an idle force-flush, so the speak-stream path does not preserve the new variants across an idle gap.

Suggested changes

• Drop unterminated reasoning tails in flush() and add parameterized flush/end-of-stream tests for every supported tag.
• Replace the WebSocket's literal `` idle guard with the same variant-aware probe.
• Avoid a second tag-list source: agent/think_scrubber.py:79-85 already owns the current set.

Automated hermes-sweeper review.

[teknium1]

Blocking: this only removes closed pairs. feed("secret") returns [] via the new unclosed-tag guard, but an end-of-stream or idle flush() returns the raw open tail and the gateway synthesizes it. Discard or remove an unclosed reasoning suffix here, and add a regression test that calls flush() after every unclosed variant.

[teknium1]

Suggestion: this is a second manually maintained source for the tag set, while agent/think_scrubber.py:79-85 already defines it. Please expose/import one shared constant so future reasoning-tag additions cannot silently diverge between display and TTS suppression.

[andrexibiza]

Verification comment (Vox Lockin lane 10 — adversarial check)

Verified this PR against current origin/main (70db671). Verdict: mergeable, CI green, covers the reasoning-tag class — no duplicate opened.

Premise confirmed on main: both TTS cleaners strip only <think>:

• tools/tts_text_normalize.py _THINK_BLOCK_RE (batch normalizer used by prepare_spoken_text)
• tools/tts_streaming.py SentenceChunker._THINK_BLOCK_RE (streamed speaker + speak-stream WS)

so a model emitting <thinking>/<reasoning>/<thought> leaks raw tags into speech.

Checked against this PR's head 8137545377:

• Tag set drawn from agent/think_scrubber.REASONING_TAG_NAMES (single source of truth) ✓
• strip_reasoning_blocks / strip_unterminated_reasoning / has_unclosed_reasoning_tag shared by both cleaners — no drift possible ✓
• Unterminated-tail handling on flush() closes the streaming cut-off hole ✓
• Regression tests added in tests/tools/test_tts_streaming.py + tests/tools/test_tts_text_normalize.py ✓

CI: all required checks pass (run 30799200885; ruff, footguns, check-attribution, Docker builds green).

Composition note (lane 10): this PR is orthogonal to my streaming-truncation work (#78234) — different code paths (tag stripping vs. length splitting). No conflicts expected.