docs(todo): Add Task #69 - Convert audio pipeline to clipped int16

Added low-priority task to convert audio processing from float32 to
clipped int16 for faster/easier processing and reduced memory footprint.

Scope: Three-phase approach (output → mixing → full pipeline)
Trade-offs: Quality vs performance/size
Priority: Low (final optimization only, if 64k budget requires it)

Benefits:
- Simpler arithmetic (no float operations)
- Smaller memory footprint (2 bytes vs 4 bytes)
- Hardware-native format (eliminates conversion)
- Natural clipping behavior

Testing requirements documented for quality validation.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

1 files changed, 50 insertions, 0 deletions

diff --git a/TODO.md b/TODO.md
index 3ab244e..3a27d01 100644
--- a/TODO.md
+++ b/TODO.md
@@ -327,6 +327,56 @@ This file tracks prioritized tasks with detailed attack plans.
   - **Benefits**: Quantify FFT speedup, validate SIMD optimizations, identify regressions
   - **Priority**: Very Low (nice-to-have for future optimization work)
 
+- [ ] **Task #69: Convert Audio Pipeline to Clipped Int16**: Use clipped int16 for all audio processing
+  - **Current**: Audio pipeline uses float32 throughout (generation, mixing, synthesis, output)
+  - **Goal**: Convert to clipped int16 for faster/easier processing and reduced memory footprint
+  - **Rationale**:
+    - Simpler arithmetic (no float operations)
+    - Smaller memory footprint (2 bytes vs 4 bytes per sample)
+    - Hardware-native format (most audio devices use int16)
+    - Eliminates float→int16 conversion at output stage
+    - Natural clipping behavior (overflow wraps/clips automatically)
+  - **Scope**:
+    - Output path: Definitely convert (backends, WAV dump)
+    - Synthesis: Consider keeping float32 for quality (IDCT produces float)
+    - Mixing: Could use int16 with proper overflow handling
+    - Asset storage: Already int16 in .spec files
+  - **Implementation Phases**:
+    1. **Phase 1: Output Only** (Minimal change, ~50 lines)
+       - Convert `synth_render()` output from float to int16
+       - Update `MiniaudioBackend` and `WavDumpBackend` to accept int16
+       - Keep all internal processing as float
+       - **Benefit**: Eliminates final conversion step
+    2. **Phase 2: Mixing Stage** (Moderate change, ~200 lines)
+       - Convert voice mixing to int16 arithmetic
+       - Add saturation/clipping logic
+       - Keep IDCT output as float, convert after synthesis
+       - **Benefit**: Faster mixing, reduced memory bandwidth
+    3. **Phase 3: Full Pipeline** (Large change, ~500+ lines)
+       - Convert spectrograms from float to int16 storage
+       - Modify IDCT to output int16 directly
+       - All synthesis in int16
+       - **Benefit**: Maximum size reduction and performance
+  - **Trade-offs**:
+    - Quality loss: 16-bit resolution vs 32-bit float precision
+    - Dynamic range: Limited to [-32768, 32767]
+    - Clipping: Must handle overflow carefully in mixing stage
+    - Code complexity: Saturation arithmetic more complex than float
+  - **Testing Requirements**:
+    - Verify no audible quality degradation
+    - Ensure clipping behavior matches float version
+    - Check mixing overflow doesn't cause artifacts
+    - Validate WAV dumps bit-identical to hardware output
+  - **Size Impact**:
+    - Phase 1: Negligible (~50 bytes)
+    - Phase 2: Small reduction (~100-200 bytes, faster code)
+    - Phase 3: Large reduction (50% memory, ~1-2KB code savings)
+  - **Priority**: Low (final optimization, after size budget is tight)
+  - **Notes**:
+    - This is a FINAL optimization task, only if 64k budget requires it
+    - Quality must be validated - may not be worth the trade-off
+    - Consider keeping float for procedural generation quality
+
 ### Developer Tools
 - [ ] **Task #66: External Asset Loading for Debugging**: mmap() asset files instead of embedded data
   - **Current**: All assets embedded in `assets_data.cc` (regenerate on every asset change)