#1633 setup: offer the experimental speed projection for Swift 1.5 too

Descrição

Setup offers the speed projection only for the original model and the Coder, and leaves it off for Swift 1.5
("made for the original Qwen3.8-Flash-Next"). We measured it on Swift 1.5, and it behaves as on the original, so
this PR offers it there too. It stays off by default, experimental, and switchable per request.

**Why the vector transfers.** The projection acts on the residual stream (layers 4-44). Swift 1.5's post-training
barely touched the weights that write into it. Comparing the BF16 checkpoints (published by SC117 on the abliterated
Swift model card):
- the routed experts' `ffn_down_exps` are byte-identical to the original's on all 48 layers;
- `ssm_out`, `attn_output` and `ffn_down_shexp` differ by 0.21-2.42%.

**Measured.**
- Swift 1.5 IQ3_XXS, the exact files of setup's pinned revision `b22d729` (sha256-verified), on Strata 0.1.41;
  GTX 1080 + 1070 layer split.
- Shipped vector, `--control-vector-layer-range 4 44 --cvec-mode project`.
- 25 AdvBench harmful behaviours and 25 Alpaca instructions (random sample, seed 20261008); thinking off, greedy,
  160 tokens; keyword refusal check.

| | harmful refused | harmless refused | decode |
|---|---|---|---|
| Swift 1.5, projection off | 25/25 | 1/25 | 31.1 tok/s |
| Swift 1.5, projection on | **0/25** | 1/25 | 32.9 tok/s |
| (for reference) SC117's abliterated Swift 1.5 | 0/25 | 1/25 | 32.1 tok/s |

- The one harmless "refusal" is the same prompt in every arm, a false positive of the keyword check.
- With the vector loaded but the request field `false`, all 25 harmless answers were byte-identical to stock.
- A code prompt gave the same function in every arm.
- Not measured: teacher-forced KL on Swift, thinking-on prompts, IQ2_XS/Q2_0, and long agent sessions.

**Possibly of interest.** SC117's abliterated Swift is orcarouter's single-direction abliteration, `W ← W − r rᵀW`
with one `r` for all layers, transplanted. `r` can be recovered from the quantized files: the top singular vector
of `ΔW` agrees across 12 layers at |cos| ≥ 0.998, and matches the BF16 edit at 0.9998. Used as a project-mode
control vector (the same `r` on layers 1-47), it also gave 0/25 and 1/25. Its harmless answers stayed closer to
stock than the shipped vector's (difflib similarity 0.56 vs 0.42). It agrees with the shipped per-layer directions
at |cos| 0.29-0.77 (peak at layer 24). We are not proposing to ship it: it derives from orcarouter's gated weights.
We mention it as evidence that the projection approach works on Swift.