feat(chatterbox-fast): context-priming at joins (§1.6, opt-in)

Prime early joins by prepending the prior sentence as backward prosodic context,
generating context+content together, then discarding the context audio. The cut
snaps to the inter-sentence pause (energy-minimum search around the context's
solo duration) with a 5ms fade-in to kill any seam click (app: _cut_at_pause /
_fade_in / Engine.generate_primed). Opt-in via request `prime` (default off).

Scheduler: priming is AFFORDABILITY-GATED so it can never starve. A primed chunk
costs ~(2·context + content)/rtf (a 2nd context-solo pass); a chunk is only primed
when buffer ≥ prime_buffer_factor (1.5) × that cost, else it falls back to a cold
generate. Consequences proven in the GPU-free sim (17 tests):
  - fires on early joins for any GPU at/above rtf_prior (3.4 = 3090; A6000 ~3.8-4.0)
  - self-skips (degrades to cold) on a slower-than-fleet GPU rather than starving
  - never primes chunk 0 (latency-critical)
Also fixed a latent Phase-1 bug: margin_first was applied at chunk 0 (budget always
0 there) so it never did anything — now applied at chunk 1 (the first transition).

Live A/B on irv-ml1 (A6000, GLaDOS): TTFB unaffected (445 vs 467ms), no starvation;
priming fired on chunk 2 (gen 1.6s for the doubled pass). On typical text exactly
ONE early join safely primes — priming chunk 2 flattens the buffer so later/larger
chunks no longer clear the safety gate. Samples: ~/chatterbox-ab/_p2_{cold,primed}.wav.

stacks/chatterbox-fast/test_scheduler.py

@@ -45,13 +45,19 @@ class FakeClock:
 def make_generator(clock: FakeClock, *, true_rtf: float, sec_per_char: float):
     """A fake generate() that costs realistic wall-clock and returns audio_sec.
 
-    Audio duration is proportional to text length; generation costs
-    ``audio_sec / true_rtf`` of (simulated) wall-clock, advancing the clock.
+    Audio duration is proportional to content length; generation costs
+    ``audio_sec / true_rtf`` of (simulated) wall-clock, advancing the clock. A
+    primed chunk (context given) costs extra: a context-solo pass plus the
+    context portion of the joint pass — modelling the ~2× cost the scheduler
+    must budget for.
     """
 
-    def generate(text: str):
-        audio_sec = len(text) * sec_per_char
-        clock.t += audio_sec / true_rtf
+    def generate(text: str, context: str | None):
+        audio_sec = len(text) * sec_per_char  # content only (context discarded)
+        gen_audio = audio_sec
+        if context:
+            gen_audio += 2 * len(context) * sec_per_char  # ctx-solo + ctx in joint
+        clock.t += gen_audio / true_rtf
         return None, audio_sec
 
     return generate
@@ -185,6 +191,43 @@ def test_full_text_reconstructed():
     assert joined == " ".join(split_sentences(PARAGRAPH))
 
 
+def test_priming_fires_but_never_on_chunk_zero():
+    """With prime_first_n=2, priming is best-effort (affordability-gated): it fires
+    on at least one early join at fleet RTF, and NEVER on chunk 0 (latency-critical)."""
+    cfg = ChunkConfig(prime_first_n=2)
+    results = run(true_rtf=3.8, cfg=cfg)
+    assert results[0].primed is False
+    assert any(r.primed for r in results[1:]), "expected at least one primed join"
+    # Priming only ever lands on chunks 1..N.
+    assert all(not r.primed for r in results if r.index > cfg.prime_first_n)
+
+
+def test_priming_never_starves_at_fleet_rtf():
+    """Priming must keep the no-starvation guarantee for any GPU at/above the
+    rtf_prior floor (3090 ~3.4, A6000 ~3.8–4.0)."""
+    cfg = ChunkConfig(prime_first_n=2)
+    for rtf in (4.0, 3.8, 3.4):
+        results = run(true_rtf=rtf, cfg=cfg)
+        assert all(not r.starved for r in results), (
+            rtf, [(r.index, r.drained) for r in results if r.starved]
+        )
+
+
+def test_priming_self_skips_on_slow_gpu_no_starvation():
+    """On a slower-than-fleet GPU (RTF below the prior), priming gracefully
+    self-skips rather than starving — the guarantee holds, priming just stops."""
+    cfg = ChunkConfig(prime_first_n=2)
+    results = run(true_rtf=3.0, cfg=cfg)
+    assert all(not r.starved for r in results)
+    assert not any(r.primed for r in results)  # degraded to cold
+
+
+def test_priming_off_by_default():
+    """Default config primes nothing (so context is never requested)."""
+    results = run(true_rtf=3.8)
+    assert all(not r.primed for r in results)
+
+
 def _main():
     results = run(true_rtf=3.8)
     print(f"{'idx':>3} {'chars':>5} {'audio_s':>8} {'gen_s':>7} "