Experiment complete. Summary of findings:

**The 1/k anti-correlation law in prime gaps is a mixture of two independent channels:**

| Property | Residue channel (Z/6Z class) | Magnitude channel (gap size) |
|----------|------------------------------|------------------------------|
| acf1 | -0.122 (3.86x full) | -0.030 (0.95x full) |
| Decay exponent | alpha = 1.30 (steeper) | alpha = 0.95 (flatter) |
| Scaling rate | +0.0072/ln(p), R2=0.91 | +0.0015/ln(p), R2=0.56 |
| Poisson crossover | p ~ 10^14.9 | p ~ 10^17.6 |
| Variance carried | 1% | 99% |

**Key insight**: The residue channel (which Z/6Z class the prime falls in) carries almost no information about gap size but carries 3.86x the sequential memory. It's intense but dies first (10^14.9). The magnitude channel is weak but persists 2.7 decades longer (10^17.6). The observed alpha=1.00 law is an effective average of 1.30 and 0.95.

**Bonus finding**: Same-class magnitude anti-correlation (1->1, 5->5) is 2.1x stronger than cross-class (1->5, 5->1) — a three-point correlation not previously measured.

Files produced:
- `tools/exp_two_channel_decomposition.py` (reusable tool)
- `tools/data/exp_two_channel_decomposition.json` (raw data)
- `tools/data/reports/agent_20260419_0330.md` (report)
- `tools/data/seme.json` (updated with TWO_CHANNEL_DECOMPOSITION tension)