@@ -32,7 +32,7 @@ docs/ tests/unit/ tests/integration/
 ## Audio format (hard constraints)
 
 - **16 kHz, mono, 16-bit PCM WAV**
-- **Peak limiter ceiling: −1.0 dBFS** — never scale up quiet clips (limiter, not normalizer)
+- **Loudness: peak-normalize to target (−2.0 dBFS default, range [−12.0, −1.5]), then peak-limit at ≤ −1.0 dBFS** — `preprocess()` applies a single global gain so the clip's absolute peak lands at `PreprocessingConfig.target_peak_dbfs`; the −1.0 dBFS limiter is a safety ceiling that is a guaranteed no-op for in-spec target values. The single-gain step preserves per-turn RMS contrast (M3a) — it shifts absolute level but every per-turn RMS *ratio* survives unchanged. Tier B/C augmentation re-applies the same target after room-IR/noise mixing through the shared `peak_normalize_to_target` helper, so all tiers exit at the same absolute peak. **Behaviour change vs. pre-#78:** Tier B/C clips previously normalized to −1.0 dBFS (zero headroom over the limiter); they now exit at the configured target (−2.0 dBFS by default). The `loudness_target_peak_dbfs` field in `GenerationMetadata` records the exact value used so any future loudness regression can be diagnosed from metadata alone — without this trail, #78 hid for three weeks behind unchanged `generator_version`. **Empirical note:** peak/RMS in the spec range does not affect Whisper WER (Whisper's log-mel feature extractor internally normalizes); the M17 ASR regression has a separate prosody-level cause being tracked elsewhere. (#78)
 - **Silence padding: ≥ 0.5 s** at head and tail — `silence_pad_applied_s` is per-side
 - Onset/offset times from `MixedScene` must be shifted by the leading pad only
 - **No torchaudio** — preprocessing uses `scipy` + `soundfile` exclusively
 
@@ -35,6 +35,7 @@ All P0–P2 work is tracked here. Milestones that span multiple PRs are listed o
 | **M2a** | P0 | ✅ Done | Lowers VIC F0 to adult range; caps AGG pitch escalation | Update `style_map` in speaker YAMLs per §4.2a table — VIC pitch −4→−1 st, AGG pitch capped 0/0/0/+1/+1 across I2–I5 | Small | No |
 | **M3a** | P0 | ✅ Done | Preserves within-scene loudness trajectory | `rms_target_dbfs` on `StyleEntry`; `_apply_rms_gain()` helper in `SceneMixer`; 4-tuple segment API; AGG −28→−15 dBFS (I1→I5), VIC −26→−30 dBFS; FLOAT temp WAV write to avoid PCM_16 hard-clip | Small | No |
 | **M3b** | P0 | ✅ Done | Prevents peak-normalize from erasing inter-scene RMS contrast | Replace `_normalize_peak()` in `preprocess()` with sample-peak limiter (only clips above −1.0 dBFS, no forced scale-up); update `validate_audio()` to check peak ≤ −1.0 dBFS instead of == −1.0 dBFS; update `spec.md §3` | Small | Yes — spec.md §3 |
+| **#78** | P1 | ✅ Done | Defines a clip-level loudness contract (target peak + safety ceiling) and a metadata trail so future loudness regressions are diagnosable from `{clip_id}.json` alone | `peak_normalize_to_target()` helper in `synthbanshee/augment/preprocessing.py` (single global gain — preserves M3a contrast); `PreprocessingConfig.target_peak_dbfs: float = −2.0` (range `[−12.0, −1.5]`, 0.5 dB headroom over the safety limiter); `preprocess()` runs the target-peak step before the −1.0 dBFS safety limiter; `cli.py` Stage 3b post-augment normalize uses the same helper + same config so Tier A and Tier B/C exit at the same absolute peak (Tier B/C target moves from −1.0 → −2.0 dBFS as a side effect — gives 1 dB inter-sample-peak headroom for room-IR/noise mixing); `GenerationMetadata` adds `loudness_target_peak_dbfs` and bumps `normalization_strategy` from `per_turn_rms_v1` → `per_turn_rms_v2_target_peak`; `PreprocessingResult` adds `target_peak_dbfs` field; step-name strings flattened to literal tokens (`peak_normalize`, `peak_limit`, `silence_pad`); integration regression tests with realistic-crest fixture asserting peak deviation < 0.1 dB and per-turn RMS contrast within 0.2 dB; spec updated. **Empirical note (lever probe, 2026-05-05):** peak normalization at this level does not change Whisper WER — Whisper's log-mel extractor internally normalizes — so this is a contract / metadata fix, not an ASR fix.  The M17 ASR regression has a separate root cause tracked in the prosody-bisect follow-up. | Small | Yes — spec.md §3, §3.1, §4.7; AGENTS.md |
 | **M4** | P0 | ✅ Done | Fixes silent label corruption in strong-label JSONL | Audit `debug_run_1` emotional states; extend `taxonomy.yaml` `emotional_states` to cover all legitimate LLM outputs; replace `_normalize_emotion()` silent fallback with explicit mapping table + logged warning (hard fail for completely unknown states); add clip-level `quality_flags` value `"emotion_downgrade"` | Small | No |
 | **M5** | P1 | ✅ Done | Protects Tier A phonetic quality from over-denoising | `PreprocessingConfig` Pydantic `BaseModel` with `wiener_denoise` flag; update `preprocess()` to accept it (default preserves current Wiener behavior); scene YAMLs gain optional `preprocessing` block; Wiener step skipped when `wiener_denoise=False`; unit tests verify Wiener skip | Small | No |
 | **M6** | P1 | ✅ Done | Replaces mechanical silence gaps with psychologically-motivated turn latencies | `TurnGapController` in `synthbanshee/tts/gap_controller.py`; project-specific gap tables (§4.5); wire into `TTSRenderer.render_scene()` replacing `turn.pause_before_s`; NEG/NEU confusor gap ranges; unit tests per context type and project | Small | No |
@@ -355,11 +356,20 @@ The label generator **must use `audible_onset_s` / `audible_end_s`** for all eve
 
 1. **Per-turn RMS targeting at the mix stage:** Each turn is gain-adjusted before being inserted into `SceneMixer` so its RMS matches the intensity-level target from §4.2. This creates the loudness escalation.
 
-2. **Replace scene-level peak normalization with a true-peak limiter:** After mixing, apply a limiter at ≤ −1.0 dBFS to ensure technical compliance without collapsing the trajectory. Note: the current `preprocess()` implementation uses sample-peak normalization (a simple divide-by-max). A true-peak limiter accounts for inter-sample peaks (relevant for lossy re-encoding, less critical here but more correct). If implementation uses sample-peak limiting, document it as such in the spec rather than calling it "true-peak."
+2. **Two-stage clip-level loudness normalization (#78, extends the original M3b limiter-only behaviour with an explicit target floor):**
+
+   - **5a. Peak-normalize to target via single global gain.** A single multiplicative factor is applied to the *entire* mixed scene so its peak lands at `PreprocessingConfig.target_peak_dbfs` (default −2.0 dBFS, range `[−12.0, −1.5]`). Single-gain normalization is the key invariant: it preserves *every* per-turn RMS *ratio* exactly, so the within-scene loudness trajectory created by step 1 survives — only the absolute level shifts. This is *not* the old per-turn divide-by-max normalizer that M3b removed; that one collapsed contrast because it scaled each segment to its own peak. The new step normalizes after mixing, once, globally.
+   - **5b. Safety limiter at ≤ −1.0 dBFS.** Defence-in-depth only: attenuates over-range samples without scaling up. The 0.5 dB margin between the target upper bound (−1.5) and the limiter ceiling (−1.0) guarantees this stage is a no-op for all in-spec target values.
+
+   **History.** PR #25 (M3a) added per-turn RMS targeting; PR #27 (M3b) replaced the legacy peak normalizer with a peak-only limiter to stop it from erasing M3a's per-turn contrast. M3b succeeded at preserving contrast but left the absolute clip peak unspecified — it floated wherever the M3a-targeted RMS produced (typically ~−6 dBFS for Tier A). The M17 Phase A spike noted that quieter clips also showed elevated Whisper WER and shifted UTMOS; issue #78 was filed under the hypothesis that loudness drove the ASR regression. **A lever probe over the same source audio at peaks ranging from −6 to +7 dBFS (and RMS from −27 to −14 dBFS) showed Whisper produces byte-identical hypotheses at every peak level in the spec range** — Whisper's log-mel extractor internally normalizes, so absolute loudness is invisible to it. Step 5a is therefore *not* an ASR fix; it is a **loudness-contract fix**: the spec previously had only an upper bound (`peak ≤ −1.0 dBFS`), no target floor, so two clips could legitimately sit 6 dB apart and both be in-spec. With step 5a in place, target peak is configured policy and `loudness_target_peak_dbfs` in `GenerationMetadata` carries the value used so future loudness drift is diagnosable from clip metadata alone.
+
+   **Tier B/C augmentation re-applies step 5a** after room-IR + noise mixing, since that pipeline reshapes the peak. Both tiers exit at `target_peak_dbfs`, so the dataset has a consistent loudness regime across tiers (it previously did not — Tier A had no target, Tier B/C hardcoded −1.0).
+
+   **The Whisper WER regression itself remains open** — it is real but loudness-independent. Lever-probe data and root-cause investigation tracked separately (prosody-bisect follow-up issue): most likely candidates are the M15 prosody tuning (#51) and the inter-word `<break>` tags from #70.
 
 3. **LUFS logging in QA:** Log short-term LUFS by turn (via pyloudnorm or equivalent) in addition to RMS. LUFS is more perceptually accurate and is the standard used by broadcast and streaming. Include it in the acoustic QA metrics (§7) for trend monitoring.
 
-4. **Spec change required:** `docs/spec.md` §3 currently states "−1.0 dBFS peak normalized." Must be updated to: "Peak-limited to ≤ −1.0 dBFS after mixing; per-turn RMS targeting applied at mix stage to preserve within-scene loudness trajectory."
+4. **Spec text:** `docs/spec.md` §3 row "Amplitude normalization" and §3.1 step 5 are kept in sync with the two-stage policy above.
 
 ---
 
 
@@ -97,7 +97,7 @@ Project codes: `SP` (She-Proves) · `EL` (Elephant in the Room)
 | Sample rate | 16,000 Hz | Resample before delivery; retain originals at native rate |
 | Bit depth | 16-bit PCM | |
 | Channels | Mono | Downmix to mono before delivery |
-| Amplitude normalization | Peak ≤ −1.0 dBFS (limiter, not normalizer) | Only attenuates; never scales up — preserves within-scene loudness trajectory |
+| Amplitude normalization | Peak-normalize to target (−2.0 dBFS default, range `[−12.0, −1.5]`) via single global gain, then peak-limit at ≤ −1.0 dBFS | Single-gain normalization preserves per-turn RMS contrast (M3a); the 0.5 dB margin between target upper bound and limiter ceiling guarantees the limiter is a no-op in normal flow (#78) |
 | Silence padding | ≥ 0.5 s of ambient baseline before and after target speech | |
 | SNR at acquisition | ≥ 15 dB (Tier A) | Tier B/C may degrade controllably below 15 dB; log actual SNR in metadata |
 | Max clip duration | 300 s (5 min) | Longer source scenes must be segmented |
@@ -111,7 +111,9 @@ All clips must pass through this pipeline before delivery. The "dirty" pre-pipel
 2. **Downmix** — stereo → mono (average channels)
 3. **Spectral filter** — low-pass at 7,500 Hz to remove irrelevant high-frequency noise from budget sensors (Butterworth order 4)
 4. **Denoising** — spectral subtraction (Wiener filtering) to remove electrical hum; parameterize noise profile from silent leading segment
-5. **Peak limit** — attenuate to ≤ −1.0 dBFS if the signal exceeds that ceiling; never scale up. This preserves the within-scene loudness trajectory established by per-turn RMS gain (M3a). A forced scale-up would collapse intensity-level amplitude differences.
+5. **Loudness normalization** (#78) — two stages:
+   - **5a. Peak-normalize to target.** Apply a single global gain so the absolute peak lands at `PreprocessingConfig.target_peak_dbfs` (default −2.0 dBFS). A *single* gain preserves per-turn RMS *ratios* exactly, so the within-scene loudness trajectory established by per-turn RMS gain (M3a) survives — only the absolute level shifts. This step replaces the M3b "limiter only, never scale up" behaviour: pre-#78 the spec had only an upper bound on peak, leaving the absolute level unspecified; two clips could legitimately sit 6 dB apart and both be in-spec.
+   - **5b. Safety limiter.** Attenuate any sample exceeding −1.0 dBFS. For in-spec target values (`target_peak_dbfs ∈ [−12.0, −1.5]`) this is a guaranteed no-op (0.5 dB margin); it remains as defence-in-depth against upstream over-range samples.
 6. **Silence pad** — verify ≥ 0.5 s ambient baseline at head and tail; add if absent
 7. **Validate** — assert: sample rate == 16000, channels == 1, no NaN/Inf samples, no UTF-8 above U+00A1 in metadata strings
 
 
@@ -33,15 +33,21 @@
 
 _TARGET_SR = 16_000
 _TARGET_CHANNELS = 1
-_PEAK_DBFS = -1.0  # −1.0 dBFS
+_PEAK_DBFS = -1.0  # −1.0 dBFS — safety ceiling enforced by validate_audio()
 _HP_CUTOFF_HZ = 80  # high-pass cutoff for DC/rumble removal
 _HP_ORDER = 2  # Butterworth order (gentle slope)
 _SILENCE_PAD_S = 0.5  # minimum silence padding (seconds)
 _MIN_DURATION_S = 3.0  # clips below this are invalid (spec §3)
 _MAX_DURATION_S = 300.0  # clips above this must be segmented (spec §3)
 
 # Re-export so callers can do: from synthbanshee.augment.preprocessing import PreprocessingConfig
-__all__ = ["PreprocessingConfig", "PreprocessingResult", "preprocess", "validate_audio"]
+__all__ = [
+    "PreprocessingConfig",
+    "PreprocessingResult",
+    "peak_normalize_to_target",
+    "preprocess",
+    "validate_audio",
+]
 
 
 @dataclass
@@ -52,8 +58,21 @@ class PreprocessingResult:
     channels: int
     duration_seconds: float
     peak_dbfs: float
+    """Measured peak in the written PCM_16 file."""
+
+    target_peak_dbfs: float | None = None
+    """Target peak the loudness step aimed for (#78), in dBFS.  Pair this
+    with ``peak_dbfs`` to compute deviation; QA tooling consumes this as
+    structured data instead of regexing a step-name string."""
+
     silence_pad_applied_s: float = _SILENCE_PAD_S
     steps_applied: list[str] = field(default_factory=list)
+    """Ordered list of step *names* (literal tokens like ``peak_normalize``
+    and ``peak_limit``).  Numeric parameters are *not* embedded in the
+    step name — they live in dedicated fields like ``target_peak_dbfs``
+    and ``peak_dbfs`` so QA grep contracts don't break under config
+    overrides."""
+
     warnings: list[str] = field(default_factory=list)
 
 
@@ -69,9 +88,10 @@ def _peak_limit(samples: np.ndarray, ceiling_dbfs: float = _PEAK_DBFS) -> np.nda
     """Attenuate samples so peak ≤ ceiling_dbfs; never scale up.
 
     Unlike peak normalization, this only applies gain when the signal
-    exceeds the ceiling.  Quieter signals are returned unchanged so that
-    the within-scene loudness trajectory established by per-turn RMS gain
-    (M3a) is preserved across the full clip.
+    exceeds the ceiling.  Quieter signals are returned unchanged.  After
+    #78 this is a safety net behind ``peak_normalize_to_target``; on its
+    own it would let M3a-shaped clips sit ~6 dB below useful ASR / MOS
+    operating range.
     """
     peak = float(np.max(np.abs(samples)))
     if peak == 0.0:
@@ -82,6 +102,40 @@ def _peak_limit(samples: np.ndarray, ceiling_dbfs: float = _PEAK_DBFS) -> np.nda
     return (samples * (ceiling_linear / peak)).astype(np.float32)
 
 
+def peak_normalize_to_target(samples: np.ndarray, target_dbfs: float) -> np.ndarray:
+    """Scale *samples* by a single global gain so their peak equals *target_dbfs*.
+
+    Used in two places that must stay consistent:
+
+    - ``preprocess()`` (Stage 3a) — Tier A clip-level loudness target.
+    - ``cli.py`` (Stage 3b post-augment) — Tier B/C re-normalization after
+      room IR + noise mixing reshapes the peak.  Both call sites read
+      ``target_dbfs`` from the same ``PreprocessingConfig.target_peak_dbfs``
+      so all tiers exit at the same absolute peak.
+
+    Why a single global gain (#78): per-turn RMS targeting (M3a) creates
+    the within-clip loudness trajectory.  We must not collapse it back
+    into a flat shape, which is what the M3b commit (PR #27) deliberately
+    avoided.  A single multiplicative factor preserves all per-turn RMS
+    *ratios* exactly while pinning the absolute peak to a configured value.
+
+    Silent or near-silent input (**peak** below ~−80 dBFS, the 1e-4
+    threshold below) is returned unchanged.  This is a *peak* guard —
+    distinct from ``_apply_rms_gain``'s identical numeric threshold,
+    which is an *RMS* guard.  Same number, different physical meaning.
+
+    Note: this helper does not clip.  Callers must constrain
+    ``target_dbfs`` so the output stays in valid PCM range; the
+    Pydantic ``[-12, -1.5]`` constraint on ``target_peak_dbfs``
+    is the canonical guarantee.
+    """
+    peak = float(np.max(np.abs(samples)))
+    if peak < 1e-4:  # peak < ~−80 dBFS — treat as silence, do not amplify
+        return samples
+    target_linear = 10.0 ** (target_dbfs / 20.0)
+    return (samples * (target_linear / peak)).astype(np.float32)
+
+
 def _resample(samples: np.ndarray, src_sr: int, dst_sr: int) -> np.ndarray:
     """Resample using polyphase filtering (scipy.signal.resample_poly)."""
     from math import gcd
@@ -189,13 +243,33 @@ def preprocess(
         samples = _wiener_denoise(samples)
         steps.append("wiener_denoise")
 
-    # --- 5. Peak limit to −1.0 dBFS (never scale up) -------------------------
+    # --- 5a. Peak-normalize to target (#78) ----------------------------------
+    # Single global gain — preserves per-turn RMS contrast (M3a) while pinning
+    # absolute peak to a configured value.  Without this step, M3a-shaped
+    # clips peak wherever per-turn RMS happens to land (~−6 dBFS for Tier A
+    # baseline targets), which made it impossible to define a clip-level
+    # loudness contract or detect downstream regressions from metadata alone.
+    #
+    # Note: empirically (lever probe, 2026-05-05) Whisper WER does not respond
+    # to the absolute peak in the spec range — its log-mel feature extractor
+    # internally normalizes — so the rationale for this step is contract
+    # clarity + metadata-trail, not ASR recovery.  The actual ASR regression
+    # tracked at the M17 spike has a different cause (likely the M15/#70
+    # prosody changes); see the prosody-bisect follow-up issue.
+    samples = peak_normalize_to_target(samples, cfg.target_peak_dbfs)
+    steps.append("peak_normalize")
+
+    # --- 5b. Safety limiter at −1.0 dBFS (never scale up) --------------------
+    # Defence-in-depth.  Pydantic constrains target_peak_dbfs ≤ −1.5, so this
+    # is a guaranteed no-op in normal flow — it only fires if upstream code
+    # somehow bypasses step 5a, in which case clamping at the spec ceiling is
+    # still strictly better than shipping a clipped clip.
     samples = _peak_limit(samples, _PEAK_DBFS)
-    steps.append(f"peak_limit_{_PEAK_DBFS}dBFS")
+    steps.append("peak_limit")
 
     # --- 6. Silence pad (≥ 0.5 s at head and tail) ---------------------------
     samples = _ensure_silence_pad(samples, sr, _SILENCE_PAD_S)
-    steps.append(f"silence_pad_{_SILENCE_PAD_S}s")
+    steps.append("silence_pad")
 
     # --- Clip duration checks ------------------------------------------------
     duration = len(samples) / sr
@@ -220,6 +294,7 @@ def preprocess(
         channels=_TARGET_CHANNELS,
         duration_seconds=duration,
         peak_dbfs=_to_dbfs(samples),
+        target_peak_dbfs=cfg.target_peak_dbfs,
         steps_applied=steps,
         warnings=warnings,
     )