engine/kokoro: question doubling + kludges notes

Re-applies the Kokoro-specific hacks that main intentionally
omits:

- _emphasize_questions doubles '?' to '??' so the 82M's flat
  interrogative prosody gets a rising-pitch cue
- engines/kokoro/hacks.md documents this and the other Kokoro-
  tuned bits (gap durations, lowercase-only respellings) with the
  'remove when we move to a bigger model' marker

Deploy from this branch to /mnt/cache/appdata/kokoro/build/ when
you want the tuned version. Main's vanilla Kokoro is for
reference / future cleanup.

engines/kokoro/hacks.md

@@ -0,0 +1,48 @@
+# Kokoro engine — kludges branched off main
+
+This branch carries the engine-specific tweaks that don't generalise
+to F5 / Tortoise. Each one is a real workaround for a real Kokoro-82M
+limitation, not a stylistic choice — when we move to a bigger model
+these should disappear.
+
+## 1. Doubled `??` for question prosody
+
+**File:** `server.py` — `_emphasize_questions` + `_QUESTION_RE`.
+
+Kokoro-82M's prosody on single `?` is flat — interrogatives read like
+declaratives. The 82M parameter cap shows up here. Doubling the mark
+to `??` triggers a noticeably stronger rising-pitch contour.
+
+Tried + works: `??`. Tried + worse: `?!` (sounds shouty), trailing
+spaces (no effect).
+
+Remove when: upgrading to a bigger model OR a Kokoro version with
+better prosody control.
+
+## 2. Paragraph / scene / breath gap durations
+
+**File:** `server.py` — `PARAGRAPH_GAP_S=0.7`, `SCENE_GAP_S=1.5`,
+`BREATH_GAP_S=0.4`.
+
+These were eyeballed against af_heart's natural pacing for long-form
+prose. Other voices (e.g. am_michael's slower delivery) may want
+shorter gaps; a per-voice override map would be more correct but
+isn't worth the complexity yet.
+
+The 2026-05-14 feedback was "some pauses are a tad too long" — the
+0.7/1.5/0.4 may want to drop to 0.5/1.2/0.3 if confirmed.
+
+## 3. Pronunciation respellings as ALL-LOWERCASE
+
+**File:** *(data, not code — pronunciation_overrides DB table)*
+
+Kokoro's misaki phonemizer treats consecutive uppercase letters as
+initialisms ("PRIP-yat" → "P-R-I-P yat"). The seeded respellings in
+`pronunciation_overrides WHERE phoneme_format='respelling'` must
+therefore use lowercase syllabification: `prip-yat`, `dyat-loff`,
+`bryu-hah-noff`. Stress marking is lost.
+
+For tortoise this constraint may not hold (different phonemizer);
+the respelling format is currently kokoro-tuned. Future: per-engine
+phoneme_format buckets, or have skald narrate pass the engine name
+when selecting overrides.

engines/kokoro/server.py

@@ -113,12 +113,23 @@ def _parse_tag(match: re.Match) -> float:
     return dur / 1000.0 if unit == "ms" else dur
 
 
+# [HACK — engine/kokoro] Kokoro-82M has weak question prosody on a
+# single `?`. Doubling the question mark to `??` reliably triggers a
+# more interrogative rising-pitch contour without changing semantics.
+# Skip if already doubled or part of an interrobang. See hacks.md.
+_QUESTION_RE = re.compile(r"(?<![?!])\?(?!\?)")
+
+
+def _emphasize_questions(text: str) -> str:
+    return _QUESTION_RE.sub("??", text)
+
+
 def _expand_inline(text: str, voice: str | None) -> list[Node]:
     """Expand inline [breath]/[pause]/[scene] tags inside a chunk
     of text that already has a single voice attribution. Voice
     blocks themselves are handled one level up in split_to_nodes."""
     out: list[Node] = []
-    text = text.strip()
+    text = _emphasize_questions(text.strip())
     if not text:
         return out
     cursor = 0