Rerun prefix-completion dissociation with rstrip bug fix (#138 follow-up)

Context

Issue #138 ran a prefix-completion dissociation experiment testing whether [ZLT] markers are prompt-gated or content-primed. The experiment found prompt-gating (A > C > B ≈ D), but absolute rates were severely attenuated (A=2.4% vs A1 free-gen=51% for villain).

Investigation found the root cause: `strip_marker()` in `scripts/eval_dissociation_inference.py` calls `.rstrip()` after removing [ZLT], which strips the trailing `\n\n` transition tokens. The Qwen tokenizer merges `.\n\n` into a single token (ID 382) that is different from `.` (ID 13). This changes the model's logit distribution at the continuation point — the model sees token 13 and outputs `<|im_end|>` instead of seeing token 382 and outputting [ZLT].

What to do

1. Fix `strip_marker()`: Remove `.rstrip()` — only strip the literal `[ZLT]` substring, preserving all surrounding whitespace including the `\n\n` transition.

2. Verify tokenization alignment: Before running the full matrix, verify on 1 model that the fixed prefix produces the same token IDs as the original generation (up to the [ZLT] token boundary).

3. Rerun the full 10×10 matrix with the fix, at least 3 seeds. Compare condition A rates to A1 free-gen rates — they should now be much closer (expected: 20-50% vs the current 2-4%).

4. Recheck the dissociation: Does the A > C > B ≈ D ordering still hold at higher absolute rates? The relative finding may strengthen or change when the signal is 10-20x stronger.

Expected outcome

With the fix, condition A rates should approximate A1 source rates (30-60% range). The B vs D comparison will have much more statistical power. If B is still ≈ D at these higher rates, prompt-gating is confirmed with high confidence. If B > D substantially, there may be content-priming that was invisible at the attenuated rates.

Parent issues

• #138 (original dissociation experiment)
• Clean result #173 (current findings)
• Clean result #232 (coupling predictor correlation)

Compute estimate

~1-2 GPU-hours per seed × 3 seeds = 3-6 GPU-hours. Same infrastructure as #138 (1xH200, pod5).